Contents
Vigranumpy exports the functionality of the C++ image processing library VIGRA to Python. It can be invoked by importing the vigra module:
import vigra
Vigranumpy is based on the popular numpy module and uses its ndarray data structure to represent image and volume data. It introduces the C++ wrapper class NumpyArray to allow transparent execution of VIGRA C++ functions on numpy arrays. Thus, vigranumpy is fully interoperable with existing numpy functionality, including various tools for image display such as matplotlib. Since vigranumpy uses boost_python, it is able to support function overloading (which plain Python does not provide), so that calling syntax is largely uniform, regardless of the type and dimension of the input arguments.
Basic calling syntax is similar to C++, with one important difference: Arguments for output images are optional. If no output image is provided, vigranumpy will allocate it as appropriate. In either case, the output image will be returned by the function, for example:
# allocate new result image
>>> smoothImage = vigra.gaussianSmoothing(inputImage, scale)
# reuse and overwrite existing result image
>>> smoothImage = vigra.gaussianSmoothing(inputImage, scale, out=smoothImage)
Unless otherwise noted, all functions expect and create arrays with dtype=numpy.float32.
Another important concern is the interpretation and ordering of the array axes. Numpy does not provide any means to attach semantics to axes, but relies purely on the convention that the most important axis is last, as in array[y, x] or array[z, y, x] (“Corder”). However, there is no way to enforce this convention in a program, since arrays can be transposed outside of the user’s control (e.g. when saving data to a file). Moreover, many imaging libraries (e.g. Image Magick, OpenCV, Qt and the C++ version of VIGRA) use the opposite convention where the xaxis comes first, as in array[x, y] or array[x, y, z]. This makes it very difficult to convert solutions developed in Python into a fast C++ version, because one has to reverse all indices without making mistakes. Matters become even more complicated when multichannel (e.g. color) images are considered – should the color index now be first or last?
To solve these ambiguities in a clean way, vigranumpy introduces the concept of axistags which is realized in class vigra.AxisTags. Every VigraArray (which is a subclass of numpy.ndarray) gets a new property array.axistags that describes axis semantics, and all vigranumpy functions account for and preserve axistag information. Unfortunately, this functionality cannot easily be retrofitted to numpy.ndarray itself. Therefore, we employ the following conversion rules between Python and C++ arrays:
When the Python array has no array.axistags property, it is mapped to the C++ NumpyArray without any change in axis ordering. Since most VIGRA functions can work on arbitrarily transposed arrays, you will get correct results, but execution may be slower because the processor cache is poorly utilized in certain axis orders.
Moreover, this may lead to overload resolution ambiguities. For example, when the array has shape (3, 60, 40), vigranumpy has no way to decide if this is a 2dimensional RGB image or a 3dimensional array that happens to have only 3 slices. Thus, vigranumpy may not always execute the function you actually intended to call.
When the Python array has the array.axistags property, it is transposed into a canonical axis ordering before vigranumpy executes a function, and the results are transposed back into the original ordering. Likewise, functions that change axis ordering (such as array.swapaxes(0,1)) or reduce the number of axes (such as array.max(axis=1)) as well as array arithmetic operations preserve axistags (see section Mathematical Functions and Type Coercion). Thus, you can work in any desired axis order without loosing control. Overload ambiguities can no longer occur because a function cannot be called when the axistags are unsuitable.
Detailed information about the use of axistags is given in section Axistags and the VigraArray Data Structure below. Section Writing Your Own C++ Modules describes how you can take advantage of the axistags mechanism in your own C++ code.
While vigranumpy can directly work on numpy.ndarrays, this would not give us the advantages of axistags as described above. Therefore, vigranumpy introduces its own array class VigraArray which is a subclass of numpy.ndarray, but reimplements many of its methods so that axistags are respected. Arrays with a conforming axistags property are most easily constructed by one of the predefined array factories. A view with axistags can be created from an existing numpy.ndarray by means of the function taggedView() (in contrast, factory functions create copies of the given arrays, not views). We illustrate the ideas by some examples:
>>> width, height, depth = 300, 200, 3
# create a 2dimensional RGB image
>>> rgb = vigra.RGBImage((width, height))
>>> rgb.shape
(300, 200, 3)
>>> rgb.axistags # short output: only axis keys
x y c
>>> print(rgb.axistags) # long output
AxisInfo: 'x' (type: Space)
AxisInfo: 'y' (type: Space)
AxisInfo: 'c' (type: Channels) RGB
# create a 3dimensional scalar volume
>>> volume = vigra.ScalarVolume((width, height, depth))
>>> volume.shape
(300, 200, 3) # same shape as before
>>> volume.axistags
x y z # but different semantic interpretation
>>> print(volume.axistags)
AxisInfo: 'x' (type: Space)
AxisInfo: 'y' (type: Space)
AxisInfo: 'z' (type: Space)
It is also possible to attach additional information to the axistags, in particular the resolution of the axis, and a text comment. The resolution will be correctly adjusted when the image is resized:
>>> rgb.axistags['x'].resolution = 1.2 # in some unit of length
>>> rgb.axistags['y'].resolution = 1.4 # in some unit of length
>>> rgb.axistags['c'].description = 'fluorescence microscopy, DAPI and GFP staining'
>>> print(rgb.axistags)
AxisInfo: 'x' (type: Space, resolution=1.2)
AxisInfo: 'y' (type: Space, resolution=1.4)
AxisInfo: 'c' (type: Channels) fluorescence microscopy, DAPI and GFP staining
# interpolate the image to twice its original size
>>> rgb2 = vigra.sampling.resize(rgb, shape=(2*width1, 2*height1))
>>> print(rgb2.axistags)
AxisInfo: 'x' (type: Space, resolution=0.6)
AxisInfo: 'y' (type: Space, resolution=0.7)
AxisInfo: 'c' (type: Channels) fluorescence microscopy, DAPI and GFP staining
When the array is transposed, the axistags are transposed accordingly. When axes are dropped from the array, the corresponding entries are dropped from the axistags property:
# transpose the volume into reverse axis order
>>> transposed_volume = volume.transpose()
>>> transposed_volume.shape
(3, 200, 300)
>>> transposed_volume.axistags
z y x
# get a view to the first slice (z == 0)
>>> first_slice = volume[..., 0]
>>> first_slice.shape
(300, 200)
>>> first_slice.axistags
x y
# get the maximum of each slice
>>> volume.max(axis=0).max(axis=0)
VigraArray(shape=(3,), axistags=z, dtype=float32, data=
[ 0. 0. 0.])
# likewise, but specify axes by their keys
>>> volume.max(axis='x').max(axis='y')
VigraArray(shape=(3,), axistags=z, dtype=float32, data=
[ 0. 0. 0.])
The initial ordering of the axes is controlled by the argument order that can optionally be passed to the VigraArray constuctor or the factory functions. If order is not explicitly provided, it is determined by the static property VigraArray.defaultOrder (which yields ‘V’ order). The following orders are currently supported:
 ‘C’ order:
 Both strides and axes are arranged in descending order, as in a plain numpy.ndarray. For example, axistags will be ‘y x c’ or ‘z y x c’. array.flags[‘C_CONTIGUOUS’] will be true.
 ‘F’ order:
 Both strides and axes are arranged in ascending order, i.e. opposite to ‘C’ order. For example, axistags will be ‘c x y’ or ‘c x y z’. array.flags[‘F_CONTIGUOUS’] will be true.
 ‘V’ order:
 VIGRAorder is an interleaved memory layout that simulates vectorvalued pixels or voxels: Channels will be the last axis and have the smallest stride, whereas all other axes are arranged in ascending order. For example, axistags will be ‘x y c’ or ‘x y z c’.
 ‘A’ order:
 Defaults to ‘V’ when a new array is created, and means ‘preserve order’ when an existing array is copied.
The meaning of ‘ascending’ or ‘descending’ order is determined by two rules: the primary order is according to axis type (see vigra.AxisType), where Channels < Space < Angle < Time < Frequency < Unknown. The secondary order (between axes of the same type) is lexicographic, such that ‘x’ < ‘y’ < ‘z’. Usage examples:
>>> rgbv = vigra.RGBImage((width, height), order='V')
>>> rgbv.shape
(300, 200, 3)
>>> rgbv.axistags
x y c
>>> rgbc = vigra.RGBImage((width, height), order='C')
>>> rgbc.shape
(200, 300, 3)
>>> rgbc.axistags
y x c
>>> rgbf = vigra.RGBImage((width, height), order='F')
>>> rgbf.shape
(3, 300, 200)
>>> rgbf.axistags
c x y
Functions that reduce the array to a onedimensional shape (flatten(), flat, ravel(), take()) always transpose the array into ‘C’ order before flattening.
Axistags are stored in a listlike class vigra.AxisTags, whose individual entries are of type vigra.AxisInfo. The simplest way to attach axistags to a plain numpy.ndarray (by creating a view of type VigraArray) is via the convenience function vigra.taggedView().
A Fortran array declaration:
real f(20,10)
is compiled such that the elements of the first index are consecutive in memory. In contrast, a C array declaration:
float c[20][10];
places the elements of the last index consecutive in memory. The two possibilities are commonly referred to as “Fortran order” and “C order”, but we will see that these terms are actually ambiguous because their meaning depends on what the array is used for.
When the array represents a matrix, users require the syntax to conform to the mathematical syntax conventions (in order to avoid bugs when typing formulas). In mathematics, m_{ij} means that the first index i refers to rows, and the second index j refers to columns. Thus, Fortran’s f(i, j) and C’s c[i][j] have exactly the same meaning, but the memory layouts of the two matrices differ: Since in Fortran the first index changes quickest, this is referred to as the “column major” format (the elements of a column are consecutive in memory). In contrast, C uses the “row major” format (the elements of a row are consecutive in memory).
In short: When the array is a matrix, the syntax is the same, but the memory layouts differ.
When the array represents an image, users require the memory layout to be the same as in the image files which store the data (in order to ensure fast I/O). All image formats (JPEG, TIFF, PNG, ...) follow the convention of analog television to scan an image left to right, then top to bottom. Consequently, the horizontal pixels are consecutive in memory, and therefore Fortran (where the first axis changes fastest) must use the syntax f(x, y), whereas C (where the last axis changes fastest) must use c[y][x].
In short: When the array is an image, the memory layout is the same, but the syntax differs.
Thus, we have four basic conventions: FM and CM for matrices, FI and CI for images. The meaning of the terms “Fortran order” and “C order” depends on whether we are talking about matrices (where they refer to FM and CM, i.e. a difference in memory layout) or images (where they refer to FI and CI, i.e. a difference in index order).
When we add more dimensions, the confusion increases, because there are no universally accepted memory and indexing conventions. For example, an RGB image can be stored in “interleaved format”:
RGB RGB RGB ...
RGB RGB RGB ...
:
:
where the color values of each pixel are consecutive in memory, or in “banded format”:
R R R ...
R R R ...
:
G G G ...
G G G ...
:
B B B ...
B B B ...
:
where we have a separate scalar image for each color. In Fortran, interleaved and banded images must be indexed as f(color, x, y) and f(x, y, color) respectively, whereas in C we must use c[y][x][color] or c[color][y][x].
From the beginning, VIGRA adopted Fortran conventions, i.e. its default behavior is according to FM and FI (this is possible because VIGRA uses array classes, where the mapping from indices to memory is encapsulated in the appropriate way).
In contrast, numpy adopted C conventions, i.e. its default behavior is CM and CI.
In addition, both packages provide array views which keep the memory layout intact, but change the index order. Thus, VIGRA also supports the CI convention, and numpy also supports FI. Note that changing the index order is only allowed for images. Matrices always use the fixed index order dictated by mathematics where transpose(m) is a welldefined mathematical operation (which just happens to revert the index order). Therefore, the existence of array views does not imply that VIGRA supports CM or numpy supports FM.
However, numpy’s array constructor provides an argument ‘order’ which can take the values ‘C’ (default) and ‘F’ to construct arrays with C or Fortran memory order. By this mechanism, numpy also supports the FM convention (and thus all four basic possibilities).
But here is the catch: When you get a numpy array view, you have no way to tell which convention it adheres to. It simply doesn’t contain this information.
The usual solution to this problem is to enforce a fixed axis order in the entire application, but this workaround breaks down when the application must simultaneously handle arrays with different meaning (e.g. sequences ‘xyt’ vs. volumes ‘xyz’ vs. multispectral images ‘xyc’) or when the application uses modules with conflicting requirements (e.g. numpy’s ‘yx’ vs. PyQt’s ‘xy’).
This is precisely where axistags enter: They attach information to array views that allows the user to figure out which convention applies to a given view. Thus, the primary purpose of axistags is entirely passive  they just keep track of how users manipulate the axis order when creating new array views. The main use case of axistags is therefore to readjust the axis order whenever an algorithm has specific order requirements. Using axistags, one can easily ensure that arrays conform to the desired order at the beginning of every algorithm. Consider the following example: Let ‘image’ be a scalar 2D image with axistags. Then we can ask for the image’s width and height independently of the current axis order:
width = image.width # this works for any axis order!
height = image.height
Now suppose we want to execute a numpy algorithm which expects the [y, x] ordering. We simply transpose the array before calling the algorithm like this:
# adjust the axis order
numpy_image = image.transposeToNumpyOrder()
# execute the algorithm
for y in xrange(height):
for x in xrange(width):
numpy_image[y, x] = ... # note the index order
When we want to execute a VIGRA algorithm which expects the [x, y] ordering, we do:
# adjust the axis order
vigra_image = image.transposeToVigraOrder()
# execute the algorithm
for y in xrange(height):
for x in xrange(width):
vigra_image[x, y] = ... # note the index order
Notice that the order of the loops (the inner loop runs over x) is the same in both cases: To maximize cache locality and therefore speed, the inner loop should operate on consecutive memory cells. Since image memory order derives from file memory order (where the xaxis is consecutive), and array views can never change the memory order, this loop ordering is always preferable, regardless of the index order.
To handle axis meaning in a welldefined way, vigranumpy adopts the following conventions, which are designed such that The Right Thing (TM) should happen automatically, at least most of the time:
When the array represents a matrix, no axistags are allowed because the index order has a fixed semantic meaning and must not be messed around with. In vigranumpy, this requirement is enforced by an assertion:
vigra_precondition( !matrix.axistags(), "matrix must not have axistags");
in the C++ gluecode functions. This applies, for example, to the feature matrices passed to a random forest and to unsupervised learning algorithms. If desired, we can introduce additional axistags for features and samples in the future because this is a common use case.
When arrays represent image data with up to five dimensions, axistags should be used. To sort indices according to the requirements of the next algorithm to be executed, the appropriate convenience function should be called (many more convenience functions are documented in vigra.VigraArray):
numpy_array = array.transposeToNumpyOrder() # gives 'yx', 'zyx' etc.
vigra_array = array.transposeToVigraOrder() # gives 'xy', 'xyz' etc.
ilastik_array = array.view5D() # gives 'txyzc' (inserts singleton axes if necessary)
user_defined = array.withAxes('y', 'x', 't') # specify order explicitly (inserts singleton axes if necessary)
Algorithms with special order requirements can then check for the correct order in an assertion.
The function vigra.taggedView() allows to attach axistags to an array very conveniently. For example, when you know from the context that the axes of a given array are to be interpreted as ‘xyzt’ in that order, you can make this knowledge explicit by calling:
tagged_array = vigra.taggedView(array, 'xyzt')
When you call a vigranumpy function that executes a C++ VIGRA function, image.transposeToVigraOrder() will be invoked automatically on the input arrays, and the original axis order will be restored in the output arrays.
When you call a vigranumpy function that is forwarded to a numpy function (in particular, a ufunc like +, , sqrt etc.), image.transposeToNumpyOrder() will be invoked automatically on the input arrays, and the original axis order will be restored in the output arrays.
When vigranumpy writes arrays to a file, it will always order the axes such that the memory order conforms to the established file conventions (e.g. values along the xaxis are consecutive). In particular, when you use vigra.impex.writeHDF5() to create HDF5 datasets, array.transposeToNumpyOrder() will be called before writing the data (this is a consequence of item 5, because writeHDF5() eventually forwards the actual work to h5py). In addition, the axistags (in numpy order) will be stored in a dataset attribute axistags.
When vigranumpy reads data from a file, it preserves the file’s memory order and attaches the appropriate axistags. In case of images, the axis order follows from the usual file conventions. If you call vigra.impex.readHDF5() to read HDF5, the axistags will be read from the attribute axistags (if present). Upon return, the read functions automatically call array.transposeToVigraOrder(), but this only changes the index order, not the memory layout. This latter convention was adopted to ensure that the default index order is the same in the Python and C++ versions of VIGRA.
When you display an image via image.imshow() or image.show(), the axes are reordered automatically such that the image is displayed upright (i.e. x goes to the right, y goes down). If you want to override this (i.e. want to enforce transposed display), you can remove the axistags by calling image.view(numpy.ndarray).
Item 4. has two important consequences one should be aware of:
Enum to encode the type of an axis described by an AxisInfo object. Possible values:
 AxisType.Channels:
 a channel axis
 AxisType.Space:
 a spatial axis
 AxisType.Edge:
 an edge axis
 AxisType.Angle:
 an axis encoding angles (e.g. polar coordinates)
 AxisType.Time:
 a temporal axis
 AxisType.Frequency:
 an axis in the Fourier domain
 AxisType.UnknownAxisType:
 type not specified
 AxisType.NonChannel:
 any type except Channels
 AxisType.AllAxes:
 any type
Types can be combined by the bitwise ‘or’ operator. For example, Space  Frequency denotes a spatial axis in the Fourier domain.
An object describing a single axis.
Constructor:
Parameters: 


In addition, AxisInfo defines the following factories for the most common cases:
 AxisInfo.c or AxisInfo.c(description='a description'):
 Factory for an axisinfo object describing the ‘c’ (channel) axis.
 AxisInfo.x or AxisInfo.x(resolution=0.0, description=''):
 Factory for an axisinfo object describing the ‘x’ (spatial) axis.
 AxisInfo.y or AxisInfo.y(resolution=0.0, description=''):
 Factory for an axisinfo object describing the ‘y’ (spatial) axis.
 AxisInfo.z or AxisInfo.z(resolution=0.0, description=''):
 Factory for an axisinfo object describing the ‘z’ (spatial) axis.
 AxisInfo.z or AxisInfo.n(resolution=0.0, description=''):
 Factory for an axisinfo object describing the ‘n’ (spatial) axis.
 AxisInfo.e or AxisInfo.e(resolution=0.0, description=''):
 Factory for an axisinfo object describing the ‘e’ (edge) axis.
 AxisInfo.t or AxisInfo.t(resolution=0.0, description=''):
 Factory for an axisinfo object describing the ‘t’ (time) axis.
 AxisInfo.fx or AxisInfo.fx(resolution=0.0, description=''):
 Factory for an axisinfo object describing the ‘x’ axis in the Fourier domain.
 AxisInfo.fy or AxisInfo.fy(resolution=0.0, description=''):
 Factory for an axisinfo object describing the ‘y’ axis in the Fourier domain.
 AxisInfo.fz or AxisInfo.fz(resolution=0.0, description=''):
 Factory for an axisinfo object describing the ‘z’ axis in the Fourier domain.
 AxisInfo.ft or AxisInfo.ft(resolution=0.0, description=''):
 Factory for an axisinfo object describing the ‘t’ axis in the Fourier domain.
axisinfo1.compatible(axisinfo2) yields True when the two axisinfo objects have the same keys and types, or either of the two is ‘unknown’.
(read/write property, type ‘string’) the string description of the axis.
(readonly property, type ‘string’) the key of the axis.
(read/write property, type ‘float’) the resolution of the axis. The resolution will be automatically adjusted whenever the image size changes, e.g. due to a call to resize() or slicing with nonunit step size:
>>> a = vigra.RGBImage((200,100))
>>> a.axistags['x'].resolution = 1.0
>>> a.axistags['y'].resolution = 1.2
>>> print(a.axistags)
AxisInfo: 'x' (type: Space, resolution=1)
AxisInfo: 'y' (type: Space, resolution=1.2)
AxisInfo: 'c' (type: Channels) RGB
>>> b = a[::2, ::4, :]
>>> print(b.axistags)
AxisInfo: 'x' (type: Space, resolution=2)
AxisInfo: 'y' (type: Space, resolution=4.8)
AxisInfo: 'c' (type: Channels) RGB
Object to describe axis properties and axis ordering in a VigraArray.
Constructor:
The parameters ‘i1’...’i5’ are the AxisInfo objects describing the axes. If all are None, an empty AxisTags object is created. Alternatively, ‘i1’ can also be a Python sequence of AxisInfo objects, or an integer (in which case an AxisTags object with that many ‘?’ entries is created).
Most AxisTags methods should not be called directly, but via the corresponding array methods, because this ensures that arrays and their axistags are always kept in sync (rule of thumb: if an axistags function is not documented, you call it on your own risk).
The entries of an axistags object (i.e. the individual axisinfo objects) can be accessed via the index operator, where the argument can either be the axis index or the axis key:
>>> print(array.axistags[0])
AxisInfo: 'x' (type: Space, resolution=1.2)
>>> print(array.axistags['x'])
AxisInfo: 'x' (type: Space, resolution=1.2)
>>> array.axistags['x'].resolution = 2.0
>>> print(array.axistags['x'])
AxisInfo: 'x' (type: Space, resolution=2)
How many axes of the given type(s) are in this axistags object?:
axistags.axisTypeCount(types) > int
The ‘types’ of the query must be single AxisType instances or a combination of them. Examples:
>>> a = vigra.defaultAxistags('txyc')
>>> a.axisTypeCount(vigra.AxisType.Space)
2
>>> a.axisTypeCount(vigra.AxisType.Time)
1
>>> a.axisTypeCount(vigra.AxisType(vigra.AxisType.Space  vigra.AxisType.Time))
3
>>> a.axisTypeCount(vigra.AxisType.NonChannel)
3
(readonly property, type ‘int’) the index of the channel axis, or len(axistags) when no channel axis exists (i.e. axistags.channelIndex is similar to axistags.index('c'), but doesn’t throw an exception when there is no ‘c’ axis.)
Construct a new AxisTags object from the given JSON representation. This is mainly used to reconstruct arrays from HDF5 datasets with a suitable axistags attribute (see readHDF5()).
Get the axis index of a given axis key:
>>> axistags.index('x')
0
In this example, the ‘x’axis corresponds to index 0 (i.e. the first index).
(readonly property, type ‘int’) the index of the innermost nonchannel axis.
Set a description for the channel axis, if one exists:
axistags.setChannelDescription('colors are in Lab color space')
It is similar to:
axistags['c'].description = 'colors are in Lab color space'
except when the axistags contain no channel axis, in which case setChannelDescription() is simply ignored, whereas axistags[‘c’] would cause an exception.
Create a string representation of this axistags in JSON format.
Bases: numpy.ndarray
This class extends numpy.ndarray with the concept of axistags which encode the semantics of the array’s axes. VigraArray overrides all numpy.ndarray methods in order to handle axistags in a sensible way. In particular, operations acting on two arrays simultaneously (e.g. addition) will first transpose the arguments such that their axis ordering matches.
Constructor:
Parameters: 


obj may be one of the following
order can be ‘C’ (C order), ‘F’ (Fortran order), ‘V’ (VIGRA order), ‘A’ (any), or None. This parameter controls the order of strides and axistags (unless axistags are explicit passed into the constructor). See the order definitions for details. If ‘order=None’, the order is determined by VigraArray.defaultOrder.
The AxisTags object of this array.
Get the default axis ordering, currently ‘V’ (VIGRA order).
Equivalent to self.transpose()
array.__getitem__(index) implements the indexing operator array[index]. In addition to the usual numpy.ndarray indexing functionality, this function also updates the axistags of the result array. There are three cases:
 getitem creates a scalar value => no axistags are required
 getitem creates an array view => axistags are transferred from the corresponding axes of the base array
 getitem creates a copy of an array (fancy indexing) => all axistags are ‘?’
If the index contains ‘numpy.newaxis’, a new singleton axis is inserted at the appropriate position, whose axisinfo is set to ‘?’ (unknown). If the index contains ‘vigra.newaxis(axisinfo)’, the singleton axis will get the given axisinfo.
Expand a scalar array (i.e. an array with a single channel) into an RGB array with three identical color channels. This is useful when you want to paste color annotations (e.g. user labels) into the array.
The parameter normalize can be used to normalize the array’s value range to 0..255:
Bind the axis identified by ‘which’ to the given ‘index’. This is similar to:
array[:, index, ...]
but you do not need to know the position of the axis when you use the axis key (according to axistags). For example, to get the green channel of an RGBImage, you write:
>>> rgb = vigra.RGBImage((200, 100))
>>> green = rgb.bindAxis('c', 1)
This gives the correct result irrespective of the axis ordering.
The index of the channel axis according to the axistags. For example, when axistags are ‘x y c’, the channel index is 2. If the axistags contain no channel axis, self.ndim is returned.
Create an iterator over the channels of the array. In each iteration, you get the array corresponding to a single channel. If the axistags contain no channel axis, there is only one iteration which yields the entire array. Example:
>>> rgb = vigra.RGBImage((200, 100))
>>> rgb.axistags
x y c
>>> red, green, blue = rgb.channelIter()
>>> red.axistags
x y
>>> red.shape
(200, 100)
The number of channels in this array (shape of the ‘c’ axis). If the axistags contain no channel axis, the number of channels is implicitly 1.
Copy the values of an array to another one. This is similar to:
self[...] = other
but will first transpose both arrays so that axistags are aligned. If there is no valid alignment, RuntimeError will be raised.
Get default axistags for the given specification ‘tagSpec’. TagSpec can be the number of dimensions of the array (array.ndim, must be <= 5) or a string containing a sequence of axis keys (only the default keys ‘x’, ‘y’, ‘z’, ‘t’, and ‘c’ are currently supported). The ‘order’ parameter determines the axis ordering, see the order definitions for details. If ‘noChannels’ is True, there will be no channel axis. Examples:
>>> vigra.VigraArray.defaultAxistags(3)
x y c
>>> vigra.VigraArray.defaultAxistags(4)
x y z c
>>> vigra.VigraArray.defaultAxistags(5)
x y z t c
>>> vigra.VigraArray.defaultAxistags(3, order='C')
y x c
>>> vigra.VigraArray.defaultAxistags(2, noChannels=True)
x y
>>> vigra.VigraArray.defaultAxistags(3, noChannels=True)
x y z
>>> vigra.VigraArray.defaultAxistags(4, noChannels=True)
x y z t
>>> vigra.VigraArray.defaultAxistags('xty')
x t y
>>> vigra.VigraArray.defaultAxistags('xty', order='V')
x y t
The depth of the array (shape of the ‘z’ axis). If the axistags contain no ‘z’ axis, RuntimeError will be raised.
Drop the channel axis when it is a singleton. This function is for easy transformation of an array shaped (width, height, 1) into (width, height). A RuntimeError is raised when there is more than one channel, unless ignoreMultiChannel=True, in which case ‘self’ is returned.
The number of time steps in the array (shape of the ‘t’ axis). If the axistags contain no ‘t’ axis, RuntimeError will be raised.
The height of the array (shape of the ‘y’ axis). If the axistags contain no ‘y’ axis, RuntimeError will be raised.
The index of the innermost nonchannel axis according to the axistags. The innermost axis is determined by the AxisInfo sorting rules (see the order definitions for details). For example, when axistags are ‘x y c’, the innerNonchannelIndex is 0.
Insert a singleton channel axis. This function is for easy transformation of an array shaped (width, height) into (width, height, 1). The ‘order’ parameter determines the position of the new axis: when order is ‘F’, it will become the first axis, otherwise it will become the last one. A RuntimeError is raised when the array already contains a channel axis.
Create the permutation that would transpose an array that is in normal (ascending) order into the axis order of this array. (e.g. ‘c x y’ into ‘x y c’).
Create the permutation that would transpose an array that is in numpy (descending) order into the axis order of this array. (e.g. ‘y x c’ into ‘x y c’).
Create the permutation that would transpose an array that is in VIGRA order (ascending spatial order, but with the channel axis at the last position) into the axis order of this array. (e.g. ‘x y c’ into ‘c x y’).
Create the permutation that would transpose this array to normal order (that is, from the current axis order into ascending order, e.g. ‘x y c’ into ‘c x y’). If ‘types’ is not ‘AxisType.AllAxes’, only the axes with the desired types are considered.
Create the permutation that would transpose this array to numpy order (that is, from the current axis order into descending order, e.g. ‘x y c’ into ‘y x c’).
Create the permutation that would transpose this array into an array view with the given order (where order can be ‘A’, ‘C’, ‘F’, ‘V’ with the usual meaning).
Create the permutation that would transpose this array to VIGRA order (that is, from the current axis order into ascending spatial order, but with the channel axis at the last position, e.g. ‘c x y’ into ‘x y c’).
Convert this image to a Qt QImage (mainly for display purposes). The present image must have 1, 2, 3, or 4 channels, and the resulting QImage will have QImage.Format_Indexed8 iff there was only one channel and QImage.Format_[A]RGB32 otherwise (with the last of 2/4 channels being used as alpha channel).
The parameter normalize can be used to normalize an image’s value range to 0..255:
Reconstruct an array that has been transferred via a ZMQ socket by a call to VigraArray.sendSocket(). This only works when the ‘zmq’ module is available. The meaning of the arguments is described in zmq.Socket.recv().
Send array and metadata over a ZMQ socket. Only works if the ‘zmq’ module is available. The meaning of the arguments is described in zmq.Socket.send().
Display this image in a vigra.pyqt.ImageWindow.
The channels are intepreted as follows: 1 channel = gray image, 2 channels = gray + alpha, 3 channels = RGB, 4 channels = RGB + alpha.
The parameter normalize can be used to normalize an image’s value range to 0..255:
Create an iterator over a single spatial axis of the array. In each iteration, you get the array corresponding to one coordinate along the axis given by ‘key’. For example, to iterate along the zaxis to get all xyslices in turn, you write:
>>> volume = vigra.Volume((width, height, depth))
>>> for slice in volume.sliceIter('z'):
... processSlice(slice)
Create an iterator over all the spatial coordinates in the array. In each iteration, you get the value corresponding to a single coordinate location. If the axistags contain no spatial axes, there is only one iteration which yields the entire array. Example:
>>> s = vigra.ScalarImage((2,2))
>>> s.ravel()[...] = range(4)
>>> for p in s.spaceIter():
.... print(p)
0.0
1.0
2.0
3.0
The number of spatial axes in the array. That is, the number of entries in the axistags where the flag ‘AxisType.Space’ is set.
Construct a subarray view from a pair of points. The first point denotes the start of the subarray (inclusive), the second its end (exclusive). For example,
a.subarray((1,2,3), (4,5,6)) # equivalent to a[1:4, 2:5, 3:6]
The given points must have the same dimension, otherwise an IndexError is raised. If only one point is given, it refers to the subarray’s end, and the start is set to the point (0, 0, ...) with appropriate dimension, for example
a.subarray((4,5,6)) # equivalent to a[:4, :5, :6]
The function transforms the given point pair into a tuple of slices and calls self.__getitem__() in it. If the points have lower dimension than the array, an Ellipsis (‘...’) is implicitly appended to the slicing, so that missing axes are left unaltered.
Return a view of the array with axis1 and axis2 interchanged.
Refer to numpy.swapaxes for full documentation.
numpy.swapaxes : equivalent function
Parameters ‘i’ and ‘j’ can also be ints (axis positions) or strings (axis keys).
If ‘keepsTags’ is False, axistags are swapped like the axes, otherwise they remain unchanged such that the swapped axes aquire a new meaning.
Create an iterator over the time points of the array. In each iteration, you get the array corresponding to a single time point. If the axistags contain no time axis, there is only one iteration which yields the entire array. Example:
>>> from vigra import *
>>> axistags = AxisTags(AxisInfo.t, AxisInfo.x, AxisInfo.y)
>>> timesteps, width, height = 2, 200, 100
>>> image_sequence = Image((timesteps, width, height), axistags=axistags)
>>> step1, step2 = image_sequence.timeIter()
Returns a view of the array with axes transposed.
For a 1D array, this has no effect. (To change between column and row vectors, first cast the 1D array into a matrix object.) For a 2D array, this is the usual matrix transpose. For an nD array, if axes are given, their order indicates how the axes are permuted (see Examples). If axes are not provided and a.shape = (i[0], i[1], ... i[n2], i[n1]), then a.transpose().shape = (i[n1], i[n2], ... i[1], i[0]).
axes : None, tuple of ints, or n ints
 None or no argument: reverses the order of the axes.
 tuple of ints: i in the jth place in the tuple means a‘s ith axis becomes a.transpose()‘s jth axis.
 n ints: same as an ntuple of the same ints (this form is intended simply as a “convenience” alternative to the tuple form)
ndarray.T : Array property returning the array transposed.
>>> a = np.array([[1, 2], [3, 4]])
>>> a
array([[1, 2],
[3, 4]])
>>> a.transpose()
array([[1, 3],
[2, 4]])
>>> a.transpose((1, 0))
array([[1, 3],
[2, 4]])
>>> a.transpose(1, 0)
array([[1, 3],
[2, 4]])
An additional keyword parameter ‘keepTags’ can be provided (it has to be passed as an explicit keyword parameter). If it is True, the axistags will remain unchanged such that the transposed axes aquire a new meaning.
Equivalent to self.transposeToOrder(VigraArray.defaultOrder).
Equivalent to self.transposeToOrder(‘F’).
Equivalent to self.transposeToOrder(‘C’).
Get a transposed view onto this array according to the given ‘order’. Possible orders are:
Equivalent to self.transposeToOrder(‘V’).
Create a 5dimensional view containing the standard tags ‘x’, ‘y’, ‘z’, ‘t’, ‘c’ in the desired ‘order’ (which can be ‘C’, ‘F’, and ‘V’ with the usual meaning). If ‘self’ has an axis key that is not among the five admissible keys, an exception is raised. Axes missing in ‘self’ are added as singleton axes with the appropriate tags.
The width of the array (shape of the ‘x’ axis). If the axistags contain no ‘x’ axis, RuntimeError will be raised.
This function creates a view whose axistags are standardized in a desired way. The standardization can be specified in two forms:
Provide axistags explicitly in any format understood by vigra.makeAxistags(). The original axistags are then transposed into the given order. When the original array contains axes not listed in the new specification, these axes are dropped if they are singletons (otherwise, an exception is raised). If requested axes is not present in the original array, singleton axes are inserted at the appropriate positions, provided the axis keys are among the predefined standard keys (‘x’, ‘y’, ‘z’, ‘t’, ‘c’, ‘n’, ‘e’, ‘fx’, ‘fy’, ‘fz’, ‘ft’). The function fails if the original array contains axes of unknown type (key ‘?’):
>>> array.axistags
x y c
>>> array.shape
(100, 50, 1)
>>> view = array.withAxes('tzyx')
>>> view.axistags
t z y x
>>> view.shape
(1, 1, 50, 100)
Provide keyword arguments order and (optionally) noChannels. The array is then transposed into the desired order (‘C’, ‘F’, or ‘V’). If noChannels=True, the channel axis is dropped if it is a singleton, otherwise an exception is raised:
>>> array.axistags
x y c
>>> array.shape
(100, 50, 1)
>>> view = array.withAxes(order='F')
>>> view.axistags
c x y
>>> view.shape
(1, 100, 50)
>>> view = array.withAxes(order='C', noChannels=True)
>>> view.axistags
y x
>>> view.shape
(50, 100)
The parameters axistags and order cannot be specified simultaneously.
Write the array to a HDF5 file. This is just a shortcut for vigra.impex.writeHDF5()
Write an image to a file. Consult vigra.impex.writeImage() for detailed documentation
Write a volume to a sequence of files. Consult vigra.impex.writeVolume() for detailed documentation.
Create a new singleton axis via the indexing operator. This works similar to numpy.newaxis, but allows to provide an AxisInfo object for the new axis. For example:
>>> s = vigra.ScalarImage((width, height))
>>> s.axistags # no channel axis
x y
>>> t = s[..., numpy.newaxis]
>>> t.axistags # with unknown axis type
x y ?
>>> t = s[..., vigra.newaxis(vigra.AxisInfo.c)]
>>> t.axistags # with channel axis
x y c
Create a view to the given array with type VigraArray and desired axistags.
You can either explicitly provide axistags to be imposed on the array (parameter axistags), or a general description for the desired axis ordering (parameters order and noChannels). It is an error to specify axistags and order simultaneously. In addition, the effect of taggedView() depends on whether array already has axistags or not.
If array has no axistags or force=True (i.e. existing axistags shall be ignored) and neither the axistags nor the order parameters are given, the function acts as if order="C" was specified (case 2 below).
If array has no axistags or force=True (i.e. existing axistags shall be ignored) and the order parameter is given, the function constructs appropriate axistags via defaultAxistags():
>>> view = array.view(VigraArray)
>>> view.axistags = VigraArray.defaultAxistags(view.ndim, order, noChannels)
If array has no axistags (or force=True) and the axistags parameter is given, the function transforms this specification into an object of type AxisTags and attaches the result to the view:
>>> view = array.view(VigraArray)
>>> view.axistags = makeAxistags(axistags)
If array has axistags (and force=False) and the order parameter is given, the function transposes the array into the desired order:
>>> view = array.transposeToOrder(order)
>>> if noChannels:
... view = view.dropChannelAxis()
If array has axistags (and force=False) and the axistags parameter is given, the function calls withAxes() to transforms the present axistags into the desired ones:
>>> view = array.withAxes(axistags)
The function raises a RuntimeError when the axistag specification is incompatible with the array.
Return the view created by array.dropChannelAxis() if the given array supports that function, or return array unchanged otherwise.
Factory function for a VigraArray representing an image (i.e. an array with two spatial axes ‘x’ and ‘y’ and optionally a channel axis ‘c’). Paramters are interpreted as in the VigraArray constructor, but an exception will be raised if the shape or axistags are not imagelike.
Factory function for a VigraArray representing a singleband image (i.e. an array with two spatial axes ‘x’ and ‘y’ and no channel axis). Paramters are interpreted as in the VigraArray constructor, but an exception will be raised if the shape or axistags are unsuitable for a singleband image.
Factory function for a VigraArray representing a 2band image (i.e. an array with two spatial axes ‘x’ and ‘y’ and channel axis ‘c’ with 2 channels). Paramters are interpreted as in the VigraArray constructor, but an exception will be raised if the shape or axistags are unsuitable for a 2band image.
Factory function for a VigraArray representing a 3band image (i.e. an array with two spatial axes ‘x’ and ‘y’ and channel axis ‘c’ with 3 channels). Paramters are interpreted as in the VigraArray constructor, but an exception will be raised if the shape or axistags are unsuitable for a 3band image.
Factory function for a VigraArray representing a 4band image (i.e. an array with two spatial axes ‘x’ and ‘y’ and channel axis ‘c’ with 4 channels). Paramters are interpreted as in the VigraArray constructor, but an exception will be raised if the shape or axistags are unsuitable for a 4band image.
Factory function for a VigraArray representing a RGB image (i.e. an array with two spatial axes ‘x’ and ‘y’ and channel axis ‘c’ with 3 channels). Paramters are interpreted as in the VigraArray constructor, but an exception will be raised if the shape or axistags are unsuitable for an RGB image.
Factory function for a VigraArray representing a volume (i.e. an array with three spatial axes ‘x’, ‘y’ and ‘z’ and optionally a channel axis ‘c’). Paramters are interpreted as in the VigraArray constructor, but an exception will be raised if the shape or axistags are not volumelike.
Factory function for a VigraArray representing a singleband volume (i.e. an array with three spatial axes ‘x’, ‘y’ and ‘z’ and no channel axis). Paramters are interpreted as in the VigraArray constructor, but an exception will be raised if the shape or axistags are unsuitable for a singleband volume.
Factory function for a VigraArray representing a 2band volume (i.e. an array with three spatial axes ‘x’, ‘y’ and ‘z’ and channel axis ‘c’ with 2 channels). Paramters are interpreted as in the VigraArray constructor, but an exception will be raised if the shape or axistags are unsuitable for a 2band volume.
Factory function for a VigraArray representing a 3band volume (i.e. an array with three spatial axes ‘x’, ‘y’ and ‘z’ and channel axis ‘c’ with 3 channels). Paramters are interpreted as in the VigraArray constructor, but an exception will be raised if the shape or axistags are unsuitable for a 3band volume.
Factory function for a VigraArray representing a 4band volume (i.e. an array with three spatial axes ‘x’, ‘y’ and ‘z’ and channel axis ‘c’ with 4 channels). Paramters are interpreted as in the VigraArray constructor, but an exception will be raised if the shape or axistags are unsuitable for a 4band volume.
Factory function for a VigraArray representing a 6band volume (i.e. an array with three spatial axes ‘x’, ‘y’ and ‘z’ and channel axis ‘c’ with 6 channels). Paramters are interpreted as in the VigraArray constructor, but an exception will be raised if the shape or axistags are unsuitable for a 6band volume.
Factory function for a VigraArray representing an RGB volume (i.e. an array with three spatial axes ‘x’, ‘y’ and ‘z’ and channel axis ‘c’ with 3 channels). Paramters are interpreted as in the VigraArray constructor, but an exception will be raised if the shape or axistags are unsuitable for an RGB volume.
Chunked arrays allow to allocate big data lazily, i.e. one chunk (rectangular block) at a time. Chunks which are currently not needed can be compressed or written to disk in order to free memory. This effectively allows VIGRA to work on data bigger than RAM.
Base class for chunked arrays, can only be created via factory functions like ChunkedArrayCompressed() or ChunkedArrayHDF5().
Raises an exception This class cannot be instantiated from Python
Read data from a chunked array with the usual index or slicing syntax:
value = chunked_array[5, 20]
roi = chunked_array[5:12, 10:19]
Note that the roi is not a slice view of the original array (as in numpy.ndarray), but a copy of the data.
Raises an exception This class cannot be instantiated from Python
Write data to a chunked array with the usual index or slicing syntax:
chunked_array[5, 20] = value
chunked_array[5:12, 10:19] = roi
the backend driver of this array.
get/set the size of the chunk cache.
checkoutSubarray(start, stop, res=None) => array
Obtain a copy of the subarray in the ROI ‘[start, stop)’. If ‘res’ is given, it must have matching shape and will be used to store the data instead of allocating new storage for ‘array’.
The index operator provides a shorthand for this function, e.g. for a 2dimensional array you can equivalently write:
roi = chunked_array.checkoutSubarray((5,10), (12,19))
roi = chunked_array[5:12, 10:19]
to read the ROI from ‘start=(5,10)’ to ‘stop=(12,19)’ (exclusive). Note that ‘roi’ is a copy, so overwriting it has no effect on the chunked array. Use ‘commitSubarray()’ to overwrite data.
shape of internal array of chunks.
shape of (interior) chunks.
commitSubarray(start, array)
Write the given ‘array’ at offset ‘start’. The index operator provides a shorthand for this function, e.g. for a 2dimensional array you can equivalently write:
chunked_array.commitSubarray((5,10), roi)
chunked_array[5:12, 10:19] = roi
to write an ROI with shape (5,7) starting at ‘start=(5,10)’.
size of the currently allocated part of the data.
size of the data of a single chunk.
the array’s value type
the array’s dimension
size of the overhead caused by chunked storage.
size of the overhead caused by chunked storage for a single chunk.
‘True’ if array values cannot be changed.
releaseChunks(start, stop, destroy=False)
release or destroy all chunks that are completely contained in [start, stop).
shape of the array.
number of elements of the array.
Bases: vigra.vigranumpycore.ChunkedArrayBase
Base class for HDF5based chunked arrays, can only be created via the factory function ChunkedArrayHDF5().
Raises an exception This class cannot be instantiated from Python
Flush data to disk and close the underlying HDF5 file.
Name of the dataset backend of this array.
Name of the file backend of this array.
Flush data to disk.
True if this array is readonly.
Enum to encode the type of compression for ChunkedArrayCompressed() and ChunkedArrayHDF5():
 Compression.ZLIB:
 ZLIB default compression
 Compression.ZLIB_NONE:
 ZLIB no compression (level = 0)
 Compression.ZLIB_FAST:
 ZLIB fast compression (level = 1)
 Compression.ZLIB_BEST:
 ZLIB best compression (level = 9)
 Compression.LZ4:
 LZ4 compression (very fast)
Enum to encode open mode for ChunkedArrayHDF5():
 HDF5Mode.Default:
 Use the default strategy (ReadOnly when file and dataset exist, New otherwise)
 HDF5Mode.New:
 Create new file (existing file will be deleted)
 HDF5Mode.ReadWrite:
 Open file (create when not existing) and allow creation of new datasets.
 Contents of existing datasets may be changed, but not their shape.
 HDF5Mode.ReadOnly:
 Open files and datasets readonly, fail when not existing.
 HDF5Mode.Replace:
 Like ReadWrite, but always replace exising datasets.
Create a chunked array (type ChunkedArrayHDF5Base) backed by a HDF5 file:
ChunkedArrayHDF5(file, dataset_name, shape=None, dtype=None,
mode=HDF5Mode.Default, compression=Compression.ZLIB_FAST,
chunk_shape=None, cache_max=1, fill_value=0, axistags=None)
Parameters ‘shape’, ‘dtype’, ‘compression’, ‘chunk_shape’, ‘fill_value’, and ‘axistags’ may only be provided when a new dataset is created.
‘file’ can be either a file name or a file ID as returned by h5py.File.id.id.
‘shape’ can be up to 5dimensional.
‘chunk_shape’ must have the same dimension as ‘shape’, and its elements must be powers of 2.
‘dtype’ can currently be uint8, uint32, and float32.
‘fill_value’ is returned for all array elements that have never been written.
‘compression’ can be any of the flags defined in the Compression enum except for LZ4.
‘cache_max’ specifies how many chunks may reside in memory at the same time. If it is ‘1’, vigra will choose a sensible default, but other values may better fit your data access patterns. This is a soft limit, i.e. may be exceeded temporarily if more chunks are needed simultaneously in a single operation.
‘mode’ defines the access rights to the file and may be any of the flags defined in the HDF5Mode enum. By default, you get read permission for an existing dataset and read/write permission for a new dataset.
For more details see ChunkedArray in the vigra C++ documentation.
Create a chunked array (type ChunkedArrayBase) backed by one plain array for each chunk (rectangular data block):
ChunkedArrayLazy(shape, dtype=float32, chunk_shape=None, fill_value=0, axistags=None)
The individual chunks are allocated lazily upon first write. Reads before the first write will simply return the ‘fill_value’ without allocating memory. All allocated chunks reside in memory.
‘shape’ can be up to 5dimensional.
‘chunk_shape’ must have the same dimension as ‘shape’, and its elements must be powers of 2.
‘dtype’ can currently be uint8, uint32, and float32.
‘fill_value’ is returned for all array elements that have never been written.
For more details see ChunkedArray in the vigra C++ documentation.
Create a chunked array (type ChunkedArrayBase) backed by one plain array for each chunk (rectangular data block):
ChunkedArrayCompressed(shape, compression=LZ4, dtype=float32, chunk_shape=None,
cache_max=1, fill_value=0, axistags=None)
The individual chunks are allocated lazily upon first write. Reads before the first write will simply return the ‘fill_value’ without allocating memory. All allocated chunks reside in memory, but may be compressed when not in use. This is especially beneficial for highly compressible data like label images.
‘shape’ can be up to 5dimensional.
‘chunk_shape’ must have the same dimension as ‘shape’, and its elements must be powers of 2.
‘dtype’ can currently be uint8, uint32, and float32.
‘fill_value’ is returned for all array elements that have never been written.
‘compression’ can be any of the flags defined in the Compression enum.
‘cache_max’ specifies how many uncompressed chunks may reside in memory at the same time. If it is ‘1’, vigra will choose a sensible default, but other values may better fit your data access patterns. This is a soft limit, i.e. may be exceeded temporarily if more chunks are needed simultaneously in a single operation.
For more details see ChunkedArray in the vigra C++ documentation.
Create a chunked array (type ChunkedArrayBase) backed by a temporary file:
ChunkedArrayTmpFile(shape, dtype=float32, chunk_shape=None, cache_max=1,
path=, fill_value=0, axistags=None)
The individual chunks are allocated lazily upon first write. Reads before the first write will simply return the ‘fill_value’ without allocating memory. Unused chunks will be moved to the file to free their memory. The file is automatically deleted when the object is deleted. Use ChunkedArrayHDF5() if you need persistent storage.
‘shape’ can be up to 5dimensional.
‘chunk_shape’ must have the same dimension as ‘shape’, and its elements must be powers of 2.
‘dtype’ can currently be uint8, uint32, and float32.
‘cache_max’ specifies how many uncompressed chunks may reside in memory at the same time. If it is ‘1’, vigra will choose a sensible default, but other values may better fit your data access patterns. This is a soft limit, i.e. may be exceeded temporarily if more chunks are needed simultaneously in a single operation.
‘fill_value’ is returned for all array elements that have never been written.
‘path’ is the directory where the file is located (default: the system’s TMP directory).
For more details see ChunkedArray in the vigra C++ documentation.
Create a chunked array (type ChunkedArrayBase) backed by a plain (consecutive) array:
ChunkedArrayFull(shape, dtype=float32, fill_value=0, axistags=None)
‘shape’ can be up to 5dimensional.
‘dtype’ can currently be uint8, uint32, and float32.
‘fill_value’ is returned for all array elements that have never been written.
For more details see ChunkedArray in the vigra C++ documentation.
The module vigra.impex defines read and write functions for image and volume data. Note that the contents of this module are automatically imported into the vigra module, so you may call ‘vigra.readImage(...)’ instead of ‘vigra.impex.readImage(...)’ etc.
Check whether the given file name contains image data:
isImage(filename) > bool
This function tests whether a file has a supported image format. It checks the first few bytes of the file and compares them with the “magic strings” of each recognized image format. If the image format is supported it returns True otherwise False.
Ask for the image file extensions that vigra.impex understands:
listExtensions() > string
This function returns a string containing the supported image file extensions for reading and writing with the functions readImage() and writeImage().
Ask for the image file formats that vigra.impex understands:
listFormats() > string
This function returns a string containing the supported image file formats for reading and writing with the functions readImage() and writeImage().
Check how many images the given file contains:
numberImages(filename) > int
This function tests how many images an image file contains(Values > 1 are only expected for the TIFF format to support multiimage TIFF).
Read an array from an HDF5 file.
‘filenameOrGroup’ can contain a filename or a group object referring to an already open HDF5 file. ‘pathInFile’ is the name of the dataset to be read, including intermediate groups. If the first argument is a group object, the path is relative to this group, otherwise it is relative to the file’s root group.
If the dataset has an attribute ‘axistags’, the returned array will have type VigraArray and will be transposed into the given ‘order’ (‘vigra.VigraArray.defaultOrder’ will be used if no order is given). Otherwise, the returned array is a plain ‘numpy.ndarray’. In this case, order=’F’ will return the array transposed into Fortran order.
Requirements: the ‘h5py’ module must be installed.
Read an image from a file:
readImage(filename, dtype='FLOAT', index=0, order='') > Image
When ‘dtype’ is ‘UINT8’, ‘INT16’, ‘UINT16’, ‘INT32’, ‘UINT32’, ‘FLOAT’, ‘DOUBLE’, or one of the corresponding numpy dtypes (numpy.uint8 etc.), the returned image will have the requested pixel type. If dtype is ‘NATIVE’ or ‘’ (empty string), the image is imported with the original type of the data in the file. By default, image data are returned as ‘FLOAT’ (i.e. numpy.float32). Caution: If the requested dtype is smaller than the original type in the file, values will be clipped at the bounds of the representable range, which may not be the desired behavior.
Individual images of sequential formats such as multiimage TIFF can be accessed via ‘index’. The number of images in a file can be checked with the function numberImages(). Alternatively, readVolume() can read an entire multipage TIFF in one go.
The ‘order’ parameter determines the axis ordering of the resulting array (allowed values: ‘C’, ‘F’, ‘V’). When order == ‘’ (the default), ‘vigra.VigraArray.defaultOrder’ is used.
Supported file formats are listed by the function listFormats(). When ‘filename’ does not refer to a recognized image file format, an exception is raised. The file can be checked beforehand with the function isImage().
Read a 3D volume from a directory:
readVolume(filename, dtype='FLOAT', order='') > Volume
If the filename refers to a multipage TIFF file, the images in the file are interpreted as the zslices of the volume.
If the volume is stored in a byslice manner (e.g. one file per zslice), the ‘filename’ can refer to an arbitrary image from the set. readVolume() then assumes that the slices are enumerated like:
name_base+[09]+name_ext
where name_base, the index, and name_ext are determined automatically. All slice files with the same name base and extension are considered part of the same volume. Slice numbers must be nonnegative, but can otherwise start anywhere and need not be successive. Slices will be read in ascending numerical (not lexicographic) order. All slices must have the same size.
Otherwise, readVolume() will try to read ‘filename’ as an info text file with the following keyvalue pairs:
name = [short descriptive name of the volume] (optional)
filename = [absolute or relative path to raw voxel data file] (required)
gradfile = [abs. or rel. path to gradient data file] (currently ignored)
description = [arbitrary description of the data set] (optional)
width = [positive integer] (required)
height = [positive integer] (required)
depth = [positive integer] (required)
datatype = [UNSIGNED_CHAR  UNSIGNED_BYTE] (default: UNSIGNED_CHAR)
Lines starting with # are ignored. When import_type is ‘UINT8’, ‘INT16’, ‘UINT16’, ‘INT32’, ‘UINT32’, ‘FLOAT’, ‘DOUBLE’, or one of the corresponding numpy dtypes (numpy.uint8 etc.), the returned volume will have the requested pixel type.
The order parameter determines the axis ordering of the resulting array (allowed values: ‘C’, ‘F’, ‘V’). When order == ‘’ (the default), vigra.VigraArray.defaultOrder is used.
For details see the help for readImage().
Write an array to an HDF5 file.
‘filenameOrGroup’ can contain a filename or a group object referring to an already open HDF5 file. ‘pathInFile’ is the name of the dataset to be written, including intermediate groups. If the first argument is a group object, the path is relative to this group, otherwise it is relative to the file’s root group. If the dataset already exists, it will be replaced without warning.
If ‘data’ has an attribute ‘axistags’, the array is transposed to numpy order before writing. Moreover, the axistags will be stored along with the data in an attribute ‘axistags’.
‘compression’ can be set to ‘gzip’, ‘szip’ or ‘lzf’ gzip (standard compression), szip (available if HDF5 is compiled with szip. Faster compression, limited types), lzf (very fast compression, all types). The ‘lzf’ compression filter is many times faster than ‘gzip’ at the cost of a lower compresion ratio.
Chunking is disabled by default. When ‘chunks’ is set to True chunking is enabled and a chunk shape is determined automatically. Alternatively a chunk shape can be specified explicitly by passing a tuple of the desired shape.
Requirements: the ‘h5py’ module must be installed.
Save an image to a file:
writeImage(image, filename, dtype='', compression='', mode='w')
Parameters:
 image:
 the image to be saved
 filename:
 the file name to save to. The file type will be deduced from the file name extension (see vigra.impexListExtensions() for a list of supported extensions).
 dtype:
the pixel type written to the file. Possible values:
 ‘’ or ‘NATIVE’:
 save with original pixel type, or convert automatically when this type is unsupported by the target file format
 ‘UINT8’, ‘INT16’, ‘UINT16’, ‘INT32’, ‘UINT32’, ‘FLOAT’, ‘DOUBLE’:
 save as specified, or raise exception when this type is not supported by the target file format (see list below)
 ‘NBYTE’:
 normalize to range 0...255 and then save as ‘UINT8’
 numpy.uint8, numpy.int16 etc.:
 behaves like the corresponding string argument
 compression:
how to compress the data (ignored when compression type is unsupported by the file format). Possible values:
 ‘’ or not given:
 save with the native compression of the target file format
 ‘RLE’, ‘RunLength’:
 use run length encoding (native in BMP, supported by TIFF)
 ‘DEFLATE’:
 use deflate encoding (only supported by TIFF)
 ‘LZW’:
 use LZW algorithm (only supported by TIFF with LZW enabled)
 ‘ASCII’:
 write as ASCII rather than binary file (only supported by PNM)
 ‘1’ ... ‘100’:
 use this JPEG compression level (only supported by JPEG and TIFF)
 mode:
support for sequential file formats such as multiimage TIFF. Possible values:
 ‘w’:
 create a new file (default)
 ‘a’:
 append an image to a file or create a new one if the file does not exist (only supported by TIFF to create multipage TIFF files)
Supported file formats are listed by the function vigra.impexListFormats(). The different file formats support the following pixel types:
 BMP:
 Microsoft Windows bitmap image file (pixel type: UINT8 as gray and RGB).
 GIF:
 CompuServe graphics interchange format; 8bit color (pixel type: UINT8 as gray and RGB).
 JPEG:
 Joint Photographic Experts Group JFIF format; compressed 24bit color (pixel types: UINT8 as gray and RGB). Only available if libjpeg is installed.
 PNG:
 Portable Network Graphic (pixel types: UINT8 and UINT16 with up to 4 channels). (only available if libpng is installed)
 PBM:
 Portable bitmap format (black and white).
 PGM:
 Portable graymap format (pixel types: UINT8, INT16, INT32 as gray scale).
 PNM:
 Portable anymap (pixel types: UINT8, INT16, INT32, gray and RGB)
 PPM:
 Portable pixmap format (pixel types: UINT8, INT16, INT32 as RGB)
 SUN:
 SUN Rasterfile (pixel types: UINT8 as gray and RGB).
 TIFF:
 Tagged Image File Format (pixel types: INT8, UINT8, INT16, UINT16, INT32, UINT32, FLOAT, DOUBLE with up to 4 channels). Only available if libtiff is installed.
 VIFF:
 Khoros Visualization image file (pixel types: UINT8, INT16 INT32, FLOAT, DOUBLE with arbitrary many channels).
Write a volume as a sequence of images:
writeVolume(volume, filename_base, filename_ext, dtype='', compression='')
The resulting image sequence will be enumerated in the form:
filename_base+[09]+filename_ext
Write a volume as a multipage tiff (filename_ext must be an empty string):
writeVolume(volume, filename, '', dtype='', compression='')
Parameters ‘dtype’ and ‘compression’ will be handled as in writeImage().
vigranumpy supports all arithmetic and algebraic functions defined in numpy.ufunc, but reimplements them in module vigra.ufunc to take full advantage of axistags.
The following mathematical functions are available in this module (refer to numpy for detailed documentation):
absolute absolute add arccos arccosh arcsin arcsinh arctan
arctan2 arctanh bitwise_and bitwise_or bitwise_xor ceil
conjugate conjugate copysign cos cosh deg2rad degrees divide
equal exp exp2 expm1 fabs floor floor_divide fmax
fmin fmod frexp greater greater_equal hypot invert invert
isfinite isinf isnan ldexp left_shift less less_equal log
log10 log1p log2 logaddexp logaddexp2 logical_and logical_not
logical_or logical_xor maximum minimum modf multiply negative
nextafter not_equal power rad2deg radians reciprocal remainder
remainder right_shift rint sign signbit sin sinh spacing
sqrt square subtract tan tanh true_divide trunc
Some of these functions are also provided as member functions of VigraArray:
__abs__ __add__ __and__ __div__ __divmod__ __eq__
__floordiv__ __ge__ __gt__ __invert__ __le__ __lshift__
__lt__ __mod__ __mul__ __ne__ __neg__ __or__ __pos__
__pow__ __radd__ __radd__ __rand__ __rdiv__ __rdivmod__
__rfloordiv__ __rlshift__ __rmod__ __rmul__ __ror__ __rpow__
__rrshift__ __rshift__ __rsub__ __rtruediv__ __rxor__ __sub__
__truediv__ __xor__
As usual, these functions are applied independently at each pixel.
Vigranumpy overloads the numpyversions of these functions in order to make their behavior more suitable for image analysis. In particular, we changed two aspects:
Default output types are thus determined according to the following rules:
The output type does not depend on the order of the arguments:
a + b results in the same type as b + a
 2.a With exception of logical functions and abs(), the output type
 does not depend on the function to be executed.
2.b The output type of logical functions is bool.
 2.c The output type of abs() follows general rules unless the
input contains complex numbers, in which case the output type is the corresponding float number type:
a + b results in the same type as a / b a == b => bool abs(complex128) => float64
If the inputs have the same type, the type is preserved:
uint8 + uint8 => uint8If (and only if) one of the inputs has at least 64 bits, the output will also have at least 64 bits:
int64 + uint32 => int64 int64 + 1.0 => float64If an array is combined with a scalar of the same kind (integer, float, or complex), the array type is preserved. If an integer array with at most 32 bits is combined with a float scalar, the result is float32 (and rule 4 kicks in if the array has 64 bits):
uint8 + 1 => uint8 uint8 + 1.0 => float32 float32 + 1.0 => float32 float64 + 1.0 => float64Integer expressions with mixed types always produce signed results. If the arguments have at most 32 bits, the result will be int32, otherwise it will be int64 (cf. rule 4):
int8 + uint8 => int32 int32 + uint8 => int32 int32 + uint32 => int32 int32 + int64 => int64 int64 + uint64 => int64In all other cases, the output type is equal to the highest input type:
int32 + float32 => float32 float32 + complex128 => complex128All defaults can be overridden by providing an explicit output array:
ufunc.add(uint8, uint8, uint16) => uint16
In order to prevent overflow, necessary upcasting is performed before the function is executed.
The module vigra.colors provides functions to adjust image brightness and contrast, and to transform between different color spaces. See Color Conversions in the C++ documentation for more information.
Convert the image (singleband) into a QImage of format Format_ARGB32_Premultiplied.
import qimage2ndarray qimg = QImage(a.shape[0], a.shape[1], QImage.Format_ARGB32_Premultiplied) normalize = numpy.asarray([10, 217], dtype=image.dtype) tintColor = numpy.asarray([1.0, 0.0, 0.0], dtype=numpy.float32) #RGB vigra.colors.alphamodulated2qimage_ARGB32Premultiplied(a, qimage2ndarray.byte_view(qimg), tintColor, normalize)
Applies a colortable to the given 2D valueImage.
Colortable must have 4 columns, each row represents a color (for example, RGBA). Values in valueImage are first taken modulo the length of the colortable. In the special case where the first color in the table is transparent, that value is NOT repeated for values outside the colortable length.
Returns: uint8 image with 4 channels
Adjust the brightness of a 2D scalar or multiband image. The function applies the formula:
out = image + 0.25 * log(factor) * (range[1]  range[0])to each element of the array. ‘factor’ and ‘range[1]  range[0]’ must be positive. Elements outside the given range are clipped at the range borders. If ‘range’ is None or “” or “auto”, the range is set to the actual range of ‘image’:
range = image.min(), image.max()
Adjust the contrast of an image or volume. The function applies the formula:
out = factor * image + (1.0  factor) * (range[1]  range[0]) / 2.0to each element of the array. ‘factor’ and ‘range[1]  range[0]’ must be positive. Elements outside the given range are clipped at the range borders. If ‘range’ is None or “” or “auto”, the range is set to the actual range of ‘image’:
range = image.min(), image.max()
Adjust gamma correction to an image or volume. The function applies the formula:
diff = range[1]  range[0] out = pow((image  range[0]) / diff, 1.0 / gamma) * diff + range[0]to each element of the array. ‘gamma’ and ‘range[1]  range[0]’ must be positive. Elements outside the given range are clipped at the range borders. If ‘range’ is None or “” or “auto”, the range is set to the actual range of ‘image’:
range = image.min(), image.max()
Convert the image (singleband) into a QImage of format Format_ARGB32_Premultiplied.
import qimage2ndarray qimg = QImage(a.shape[0], a.shape[1], QImage.Format_ARGB32_Premultiplied) normalize = numpy.asarray([10, 217], dtype=image.dtype) vigra.colors.gray2qimage_ARGB32Premultiplied(a, qimage2ndarray.byte_view(qimg), normalize)
Convert the intensity range of a 2D scalar or multiband image. The function applies a linear transformation to the intensities such that the value oldRange[0] is mapped onto newRange[0], and oldRange[1] is mapped onto newRange[1]. That is, the algorithm applies the formula:
oldDiff = oldRange[1]  oldRange[0] newDiff = newRange[1]  newRange[0] out = (image  oldRange[0]) / oldDiff * newDiff + newRange[0]to each element of the array. ‘oldDiff’ and ‘newDiff’ must be positive. If ‘oldRange’ is None or “” or “auto” (the default), the range is set to the actual range of ‘image’:
range = image.min(), image.max()If ‘newRange’ is None or “” or “auto”, it is set to (0, 255.0). If ‘out’ is explicitly passed, it must be a uint8 image.
Convert the colors of the given ‘image’ using Lab2RGBFunctor.
For details see Lab2RGBFunctor in the C++ documentation.
Convert the colors of the given ‘image’ using Lab2RGBPrimeFunctor.
For details see Lab2RGBPrimeFunctor in the C++ documentation.
Convert the colors of the given ‘image’ using Lab2XYZFunctor.
For details see Lab2XYZFunctor in the C++ documentation.
Convert the colors of the given ‘image’ using Luv2RGBFunctor.
For details see Luv2RGBFunctor in the C++ documentation.
Convert the colors of the given ‘image’ using Luv2RGBPrimeFunctor.
For details see Luv2RGBPrimeFunctor in the C++ documentation.
Convert the colors of the given ‘image’ using Luv2XYZFunctor.
For details see Luv2XYZFunctor in the C++ documentation.
Convert the colors of the given ‘image’ using RGB2LabFunctor.
For details see RGB2LabFunctor in the C++ documentation.
Convert the colors of the given ‘image’ using RGB2LuvFunctor.
For details see RGB2LuvFunctor in the C++ documentation.
Convert the colors of the given ‘image’ using RGB2RGBPrimeFunctor.
For details see RGB2RGBPrimeFunctor in the C++ documentation.
Convert the colors of the given ‘image’ using RGB2XYZFunctor.
For details see RGB2XYZFunctor in the C++ documentation.
Convert the colors of the given ‘image’ using RGB2sRGBFunctor.
For details see RGB2sRGBFunctor in the C++ documentation.
Convert the colors of the given ‘image’ using RGBPrime2LabFunctor.
For details see RGBPrime2LabFunctor in the C++ documentation.
Convert the colors of the given ‘image’ using RGBPrime2LuvFunctor.
For details see RGBPrime2LuvFunctor in the C++ documentation.
Convert the colors of the given ‘image’ using RGBPrime2RGBFunctor.
For details see RGBPrime2RGBFunctor in the C++ documentation.
Convert the colors of the given ‘image’ using RGBPrime2XYZFunctor.
For details see RGBPrime2XYZFunctor in the C++ documentation.
Convert the colors of the given ‘image’ using RGBPrime2YPrimeCbCrFunctor.
For details see RGBPrime2YPrimeCbCrFunctor in the C++ documentation.
Convert the colors of the given ‘image’ using RGBPrime2YPrimeIQFunctor.
For details see RGBPrime2YPrimeIQFunctor in the C++ documentation.
Convert the colors of the given ‘image’ using RGBPrime2YPrimePbPrFunctor.
For details see RGBPrime2YPrimePbPrFunctor in the C++ documentation.
Convert the colors of the given ‘image’ using RGBPrime2YPrimeUVFunctor.
For details see RGBPrime2YPrimeUVFunctor in the C++ documentation.
Convert the colors of the given ‘image’ using XYZ2LabFunctor.
For details see XYZ2LabFunctor in the C++ documentation.
Convert the colors of the given ‘image’ using XYZ2LuvFunctor.
For details see XYZ2LuvFunctor in the C++ documentation.
Convert the colors of the given ‘image’ using XYZ2RGBFunctor.
For details see XYZ2RGBFunctor in the C++ documentation.
Convert the colors of the given ‘image’ using XYZ2RGBPrimeFunctor.
For details see XYZ2RGBPrimeFunctor in the C++ documentation.
Convert the colors of the given ‘image’ using YPrimeCbCr2RGBPrimeFunctor.
For details see YPrimeCbCr2RGBPrimeFunctor in the C++ documentation.
Convert the colors of the given ‘image’ using YPrimeIQ2RGBPrimeFunctor.
For details see YPrimeIQ2RGBPrimeFunctor in the C++ documentation.
Convert the colors of the given ‘image’ using YPrimePbPr2RGBPrimeFunctor.
For details see YPrimePbPr2RGBPrimeFunctor in the C++ documentation.
Convert the colors of the given ‘image’ using YPrimeUV2RGBPrimeFunctor.
For details see YPrimeUV2RGBPrimeFunctor in the C++ documentation.
Convert the colors of the given ‘image’ using sRGB2RGBFunctor.
For details see sRGB2RGBFunctor in the C++ documentation.
The module vigra.filters provides operators that consider a window around each pixel, compute one or several numbers from the values in the window, and store the results in the corresponding pixel of the output image. This includes convolution, nonlinear diffusion, morphological operators, feature detectors (such as the structure tensor) etc.
Generic 1 dimensional convolution kernel.
This kernel may be used for convolution of 1 dimensional signals or for separable convolution of multidimensional signals. The kernel’s size is given by its left() and right() methods. The desired border treatment mode is returned by getBorderTreatment(). The different init functions create a kernel with the specified properties. For more details, see Kernel1D in the C++ documentation.
Standard constructor:
Kernel1D()
Creates an identity kernel.
Copy constructor:
Kernel1D(other_kernel)
Return current border treatment mode.
Init kernel as an averaging filter with given radius (i.e. window size 2*radius+1). ‘norm’ denotes the sum of all bins of the kernel.
Kernel construction and initialization can be performed in one step by calling the factory function ‘averagingKernel()’.
Init kernel as a binomial filter with given radius (i.e. window size 2*radius+1). ‘norm’ denotes the sum of all bins of the kernel.
Kernel construction and initialization can be performed in one step by calling the factory function ‘binomialKernel()’.
Init kernel as a 5tap smoothing filter of the form:
[ a, 0.25, 0.5  2*a, 0.25, a]
Kernel construction and initialization can be performed in one step by calling the factory function ‘burtFilterKernel()’.
Init kernel as Lindeberg’s discrete analog of the Gaussian function. The radius of the kernel is always 3*std_dev. ‘norm’ denotes the desired sum of all bins of the kernel.
Kernel construction and initialization can be performed in one step by calling the factory function ‘discreteGaussianKernel()’.
Init the kernel with explicit values from ‘contents’, which must be a 1D numpy.ndarray. ‘left’ and ‘right’ are the boundaries of the kernel (inclusive). If ‘contents’ contains the wrong number of values, a runtime error results. It is, however, possible to give just one initializer. This creates an averaging filter with the given constant. The norm is set to the sum of the initializer values.
Kernel construction and initialization can be performed in one step by calling the factory function ‘explicitlyKernel()’.
Init kernel as a sampled Gaussian function. The radius of the kernel is always 3*std_dev. ‘norm’ denotes the desired sum of all bins of the kernel (i.e. the kernel is corrected for the normalization error introduced by windowing the Gaussian to a finite interval). However, if norm is 0.0, the kernel is normalized to 1 by the analytic expression for the Gaussian, and no correction for the windowing error is performed.
Kernel construction and initialization can be performed in one step by calling the factory function ‘gaussianKernel()’.
Init kernel as a Gaussian derivative of order ‘order’. The radius of the kernel is always 3*std_dev + 0.5*order. ‘norm’ denotes the norm of the kernel. Thus, the kernel will be corrected for the error introduced by windowing the Gaussian to a finite interval. However, if norm is 0.0, the kernel is normalized to 1 by the analytic expression for the Gaussian derivative, and no correction for the windowing error is performed.
Kernel construction and initialization can be performed in one step by calling the factory function ‘gaussianDerivativeKernel()’.
Init kernel as a 3tap second difference filter of the form:
[ 1, 2, 1]
Kernel construction and initialization can be performed in one step by calling the factory function ‘secondDifference3Kernel()’.
Init kernel as a symmetric difference filter of the form:
[ 0.5 * norm, 0.0 * norm, 0.5 * norm]
Kernel construction and initialization can be performed in one step by calling the factory function ‘symmetricDifferenceKernel()’.
Left border of kernel (inclusive).
Return the norm of kernel.
Set a new norm and normalize kernel, use the normalization formula for the given derivativeOrder.
Right border of kernel (inclusive).
Set border treatment mode.
Number of kernel elements (right()  left() + 1).
Generic 2 dimensional convolution kernel.
This kernel may be used for convolution of 2 dimensional signals. The desired border treatment mode is returned by borderTreatment().(Note that the 2D convolution functions don’t currently support all modes.) The different init functions create a kernel with the specified properties. For more details, see Kernel2D in the C++ documentation.
Standard constructor:
Kernel2D()
Creates an identity kernel.
Copy constructor:
Kernel2D(other_kernel)
Return current border treatment mode.
Vertical kernel size (lowerRight()[1]  upperLeft()[1] + 1).
Init the 2D kernel as a circular averaging filter. The norm will be calculated as 1 / (number of nonzero kernel values).
Precondition:
radius > 0
Kernel construction and initialization can be performed in one step by calling the factory function ‘diskKernel2D()’.
Init the kernel with explicit values from ‘contents’, which must be a 2D numpy.ndarray. ‘upperLeft’ and ‘lowerRight’ are the boundaries of the kernel (inclusive), and must be 2D tuples. If ‘contents’ contains the wrong number of values, a runtime error results. It is, however, possible to give just one initializer. This creates an averaging filter with the given constant. The norm is set to the sum of the initializer values.
Kernel construction and initialization can be performed in one step by calling the factory function ‘explicitlyKernel2D()’.
Init kernel as a sampled 2D Gaussian function. The radius of the kernel is always 3*std_dev. ‘norm’ denotes the desired sum of all bins of the kernel (i.e. the kernel is corrected for the normalization error introduced by windowing the Gaussian to a finite interval). However, if norm is 0.0, the kernel is normalized to 1 by the analytic expression for the Gaussian, and no correction for the windowing error is performed.
Kernel construction and initialization can be performed in one step by calling the factory function ‘gaussianKernel2D()’.
Init the 2D kernel as the cartesian product of two 1D kernels of type Kernel1D. The norm becomes the product of the two original norms.
Kernel construction and initialization can be performed in one step by calling the factory function ‘separableKernel2D()’.
Lower right border of kernel (inclusive).
Return the norm of the kernel.
Set the kernel’s norm and renormalize the values.
Set border treatment mode.
Upper left border of kernel (inclusive).
Horizontal kernel size (lowerRight()[0]  upperLeft()[0] + 1).
Compute the Euclidean distance transform of all regions in a 2D or 3D label array with respect to the region boundaries. The ‘boundary’ parameter must be one of the following strings:
 ‘OuterBoundary’: compute distance relative to outer regin boundaries
 ‘InterpixelBoundary’: compute distance relative to interpixel boundaries (default)
 ‘InnerBoundary’: compute distance relative to inner region boundaries
where the outer boundary consists of the pixels touching a given region from the outside and the inner boundary are the pixels adjacent to the region’s complement. If ‘array_border_is_active=True’, the external border of the array (i.e. the border between the image and the infinite region) is also used. Otherwise (default), regions touching the array border are treated as if they extended to infinity.
For more details see boundaryMultiDistance in the vigra C++ documentation.
Calculate the boundary tensor for a scalar valued 2D image.For details see boundaryTensor in the vigra C++ documentation.
Compute the Euclidean distance transform of all regions in a 2D or 3D label array with respect to the region boundaries and return, in each pixel, the difference vector to the nearest boundary point. The ‘boundary’ parameter must be one of the following strings:
 ‘OuterBoundary’: compute distance relative to outer regin boundaries
 ‘InterpixelBoundary’: compute distance relative to interpixel boundaries (default)
 ‘InnerBoundary’: compute distance relative to inner region boundaries
where the outer boundary consists of the pixels touching a given region from the outside and the inner boundary are the pixels adjacent to the region’s complement. If ‘array_border_is_active=True’, the external border of the array (i.e. the border between the image and the infinite region) is also used. Otherwise (default), regions touching the array border are treated as if they extended to infinity.
For more details see boundaryDistanceTransform() and boundaryVectorDistance in the vigra C++ documentation.
Convolve an array (up to 5D) with the given ‘kernel’ (or kernels). If the input has multiple channels, the filter is applied to each channel independently. The function can be used in 3 different ways:
For details see separableConvolveMultiArray and convolveImage in the vigra C++ documentation.
Convolve a single dimension of a scalar or multiband array with up to five dimensions. ‘dim’ denotes the dimension to be convolved. ‘kernel’ must be an instance of Kernel1D.
For details see convolveMultiArrayOneDimension in the vigra C++ documentation.
Apply a closing filter with disc of given radius to image.
This is an abbreviation for applying a dilation and an erosion filter in sequence. This function also works for multiband images, it is then executed on every band.
See discRankOrderFilter in the C++ documentation for more information.
Apply dilation (maximum) filter with disc of given radius to image.
This is an abbreviation for the rank order filter with rank = 1.0. This function also works for multiband images, it is then executed on every band.
See discDilation in the C++ documentation for more information.
Apply erosion (minimum) filter with disc of given radius to image.
This is an abbreviation for the rank order filter with rank = 0.0. This function also works for multiband images, it is then executed on every band.
See discErosion in the C++ documentation for more information.
Apply median filter with disc of given radius to image.
This is an abbreviation for the rank order filter with rank = 0.5. This function also works for multiband images, it is then executed on every band.
See discMedian in the C++ documentation for more information.
Apply a opening filter with disc of given radius to image.
This is an abbreviation for applying an erosion and a dilation filter in sequence. This function also works for multiband images, it is then executed on every band.
See discRankOrderFilter in the C++ documentation for more information.
Apply rank order filter with disc structuring function to a float image.
The pixel values of the source image must be in the range 0...255. Radius must be >= 0. Rank must be in the range 0.0 <= rank <= 1.0. The filter acts as a minimum filter if rank = 0.0, as a median if rank = 0.5, and as a maximum filter if rank = 1.0. This function also works for multiband images, it is then executed on every band.
For details see discRankOrderFilter in the C++ documentation.
Apply rank order filter with disc structuring function to a float image using a mask.
The pixel values of the source image must be in the range 0...255. Radius must be >= 0.Rank must be in the range 0.0 <= rank <= 1.0. The filter acts as a minimum filter if rank = 0.0,as a median if rank = 0.5, and as a maximum filter if rank = 1.0.
The mask is only applied to the input image, i.e. the function generates an output wherever the current disc contains at least one pixel with nonzero mask value. Source pixels with mask value zero are ignored during the calculation of the rank order.
This function also works for multiband images, it is then executed on every band. If the mask has only one band, it is used for every image band. If the mask has the same number of bands, as the image the bands are used for the corresponding image bands.
For details see discRankOrderFilterWithMask in the C++ documentation.
Compute the Euclidean distance transform of a scalar array (up to 3D).
All pixels with a value of 0.0 are considered background, while all pixels with a nonzero value are considered foreground. The parameter ‘background’ is a Boolean scalar that specifies whether to compute the distance of all background pixels to the nearest foreground pixel (if it is ‘True’, default) or vice versa (if it is ‘False’). Hence in the destination array, for background==True all background elements will be assigned their distance value, while all foreground elements will be assigned 0. For background==False, it is exactly the other way around.
If ‘pixel_pitch’ is given, it must contain the pixel distance along the three axes. They are then used to compute the distance anisotropically. If no ‘pixel_pitch’ is given, the data is treated isotropically with unit distance between pixels.
For more details see separableMultiDistance in the vigra C++ documentation.
Compute the distance transform of a 2D scalar float image. All pixels with a value of 0.0 are considered to be background pixels, while all pixels with a nonzero value are considered to be foreground pixels. The parameter ‘background’ is a Boolean scalar that specifies whether to compute the distance of all background pixels to the nearest foreground pixels (if it is ‘True’, default) or vice versa (if it is ‘False’). Hence in the destination image, for background==True all background pixels will be assigned their distance value, while all foreground pixels will be assigned 0. For background==False, it is exactly the other way around.
The ‘norm’ parameter gives the distance norm to use (0: infinity norm, 1: L1 norm, 2: Euclidean norm).
If ‘pixel_pitch’ is given, it must contain the pixel distance along the two axes. They are then used to compute the distance anisotropically. If no ‘pixel_pitch’ is given, the data is treated isotropically with unit distance between pixels. The anisotropic distance transform is only supported for norm =2 (Euclidean).
For details see distanceTransform in the vigra C++ documentation.
Compute a list holding the eccentricity center of each region in a label array (2D or 3D).
For more details see eccentricityCenters in the vigra C++ documentation.
Compute the eccentricity transform of a label array (2D or 3D).
For more details see eccentricityTransformOnLabels in the vigra C++ documentation.
Compute the eccentricity transform and eccentricity centers of a label array (2D and 3D).
Returns the tuple (ecc_image, centers). See eccentricityTransform() and eccentricityCenters().
Convolve ‘array’ with a Gaussian derivate kernel of the given ‘orders’. ‘orders’ must contain a list of integers >= 0 for each nonchannel axis. Each channel of the array is treated independently. If ‘sigma’ is a single value, the kernel size is equal in each dimension. If ‘sigma’ is a tuple or list of values of appropriate length, a different size is used for each axis.
‘window_size’ specifies the ratio between the filter scale and the size of the filter window. Use values around 2.0 to speedup the computation for the price of increased cutoff error, and values >= 4.0 for very accurate results. The window size is automatically determined for the default value 0.0.
For the first and second derivatives, you can also use gaussianGradient() and hessianOfGaussian().
Compute the divergence of a 2D or 3D vector field with a first derivative of Gaussian at the given scale. The array must have as many channels as spatial dimensions.
If ‘sigma’ is a single value, an isotropic filter at this scale is applied (i.e., each dimension is filtered in the same way). If ‘sigma’ is a tuple or list of values, the amount of smoothing will be different for each spatial dimension. The optional ‘sigma_d’ (single, tuple, or list) denotes the PSF standard deviation per axis, the optional ‘step_size’ (single, tuple, or list) the distance between two adjacent pixels for each dimension. The length of the tuples or lists must be equal to the number of spatial dimensions.
‘window_size’ and ‘roi’ have the same meaning as in gaussianSmoothing().
For details see gaussianDivergenceMultiArray and ConvolutionOptions in the vigra C++ documentation.
Calculate the gradient vector by means of a 1st derivative of Gaussian filter at the given scale for a scalar array (up to 4D).
If ‘sigma’ is a single value, an isotropic filter at this scale is applied (i.e., each dimension is filtered in the same way). If ‘sigma’ is a tuple or list of values, the amount of smoothing will be different for each spatial dimension. The optional ‘sigma_d’ (single, tuple, or list) denotes the PSF standard deviation per axis, the optional ‘step_size’ (single, tuple, or list) the distance between two adjacent pixels for each dimension. The length of the tuples or lists must be equal to the number of spatial dimensions.
‘window_size’ and ‘roi’ have the same meaning as in gaussianSmoothing().
For details see gaussianGradientMultiArray and ConvolutionOptions in the vigra C++ documentation.
Calculate the gradient magnitude by means of a 1st derivative of Gaussian filter at the given scale for a scalar or multiband array with up to 5 dimensions.
If ‘accumulate’ is True (the default), the gradients are accumulated (in the L2norm sense) over all channels of a multichannel array. Otherwise, a separate gradient magnitude is computed for each channel.
If ‘sigma’ is a single value, an isotropic filter at this scale is applied (i.e., each dimension is filtered in the same way). If ‘sigma’ is a tuple or list of values, the amount of smoothing will be different for each spatial dimension. The optional ‘sigma_d’ (single, tuple, or list) denotes the PSF standard deviation per axis, the optional ‘step_size’ (single, tuple, or list) the distance between two adjacent pixels for each dimension. The length of the tuples or lists must be equal to the number of spatial dimensions.
‘window_size’ and ‘roi’ have the same meaning as in gaussianSmoothing().
For details see gaussianGradientMultiArray and ConvolutionOptions in the vigra C++ documentation.
Perform sharpening function with gaussian filter.
For details see gaussianSharpening in the vigra C++ documentation.
Perform Gaussian smoothing of an array with up to five dimensions.
If the array has multiple channels, each channel is smoothed independently.
If ‘sigma’ is a single value, an isotropic filter at this scale is applied (i.e., each dimension is filtered in the same way). If ‘sigma’ is a tuple or list of values, the amount of smoothing will be different for each spatial dimension. The optional ‘sigma_d’ (single, tuple, or list) denotes the PSF standard deviation per axis, the optional ‘step_size’ (single, tuple, or list) the distance between two adjacent pixels for each dimension. The length of the tuples or lists must be equal to the number of spatial dimensions.
‘window_size’ specifies the ratio between the effective filter scale and the size of the filter window. Use a value around 2.0 to speedup the computation by increasing the error resulting from cutting off the Gaussian. For the default 0.0, the window size is automatically determined.
If ‘roi’ is not None, it must specify the desired regionofinterest as a pair ‘(first_point, beyond_last_point)’ (e.g. ‘roi=((10,20), (200,250))’). As usual, the second point is the first point outside the ROI, and the ROI must not be outside the input array dimensions. The coordinates refer only to nonchannel axes  if your array has an explicit channel axis, the ROI dimension must be one less than the array dimension. If you pass in an explicit ‘out’ array and specify an ROI, the ‘out’ array must have the shape of the ROI.
For details see gaussianSmoothing and ConvolutionOptions in the vigra C++ documentation.
Calculate the Hessian matrix by means of 2nd derivative of Gaussian filters at the given scale for scalar arrays up to 4D. The result has N*(N+1)/2 channels representing the flattened upper triangular part of the Hessian (N is the dimension of the input).
If ‘sigma’ is a single value, an isotropic filter at this scale is applied (i.e., each dimension is filtered in the same way). If ‘sigma’ is a tuple or list of values, the amount of smoothing will be different for each spatial dimension. The optional ‘sigma_d’ (single, tuple, or list) denotes the PSF standard deviation per axis, the optional ‘step_size’ (single, tuple, or list) the distance between two adjacent pixels for each dimension. The length of the tuples or lists must be equal to the number of spatial dimensions.
‘window_size’ and ‘roi’ have the same meaning as in gaussianSmoothing().
For details see hessianOfGaussianMultiArray in the vigra C++ documentation.
Compute the eigenvalues of the Hessian of Gaussian at the given scale for a scalar image or volume.
Calls hessianOfGaussian() and tensorEigenvalues().
Anisotropic tensor smoothing with the hourglass filter.
For details see hourGlassFilter in the vigra C++ documentation.
Filter a 2D or 3D scalar array with the Laplacian of Gaussian operator at the given scale. Multiple channels are filtered independently.
If ‘sigma’ is a single value, an isotropic filter at this scale is applied (i.e., each dimension is filtered in the same way). If ‘sigma’ is a tuple or list of values, the amount of smoothing will be different for each spatial dimension. The optional ‘sigma_d’ (single, tuple, or list) denotes the PSF standard deviation per axis, the optional ‘step_size’ (single, tuple, or list) the distance between two adjacent pixels for each dimension. The length of the tuples or lists must be equal to the number of spatial dimensions.
‘window_size’ and ‘roi’ have the same meaning as in gaussianSmoothing().
For details see laplacianOfGaussianMultiArray and ConvolutionOptions in the vigra C++ documentation.
Binary closing on a scalar or multiband array (up to 3D, uint8 or bool). Multiple channels are treated independently.
This function applies a flat circular closing operator (sequential dilation and erosion) with a given radius. The operation is isotropic. The input is a uint8 or boolean multidimensional array where nonzero elements represent foreground and zero elements represent background.
For details see vigra C++ documentation (multiBinaryDilation and multiBinaryErosion).
Binary dilation on a scalar or multiband array (up to 3D, uint8 or bool). Multiple channels are treated independently.
This function applies a flat circular dilation operator with a given radius. The operation is isotropic. The input is a uint8 or boolean multidimensional array where nonzero elements represent foreground and zero elements represent background.
For details see multiBinaryDilation in the C++ documentation.
Binary erosion on a scalar or multiband array (up to 3D, uint8 or bool). Multiple channels are treated independently.
This function applies a flat circular erosion operator with a given radius. The operation is isotropic. The input is a uint8 or boolean multidimensional array where nonzero elements represent foreground and zero elements represent background.
For details see multiBinaryErosion in the C++ documentation.
Binary opening on a scalar or multiband array (up to 3D, uint8 or bool). Multiple channels are treated independently.
This function applies a flat circular opening operator (sequential erosion and dilation) with a given radius. The operation is isotropic. The input is a uint8 or boolean multidimensional array where nonzero elements represent foreground and zero elements represent background.
For details see vigra C++ documentation (multiBinaryDilation and multiBinaryErosion).
Parabolic grayscale closing on a scalar or multiband array (up to 3D). Multiple channels are treated independently.
This function applies a parabolic closing (sequential dilation and erosion) operator with a given spread ‘sigma’ on a grayscale array. The operation is isotropic.
For details see multiGrayscaleDilation and multiGrayscaleErosion in the C++ documentation.
Parabolic grayscale dilation on a scalar or multiband array (up to 3D). Multiple channels are treated independently.
This function applies a parabolic dilation operator with a given spread ‘sigma’ on a grayscale array. The operation is isotropic.
For details see multiGrayscaleDilation in the C++ documentation.
Parabolic grayscale erosion on a scalar or multiband array (up to 3D). Multiple channels are treated independently.
This function applies a parabolic erosion operator with a given spread ‘sigma’ on a grayscale array. The operation is isotropic.
For details see multiGrayscaleErosion in the C++ documentation.
Parabolic grayscale opening on a scalar or multiband array (up to 3D). Multiple channels are treated independently.
This function applies a parabolic opening (sequential erosion and dilation) operator with a given spread ‘sigma’ on a grayscale array. The operation is isotropic.
For details see multiGrayscaleDilation and multiGrayscaleErosion in the C++ documentation.
loop over an image and do something with each pixels
Args:
image : input imagereturns an an image with the same shape as the input image
loop over an image and do something with each pixels
Args:
image : input image
returns an an image with the same shape as the input image
loop over an image and do something with each pixels
Args:
image : input image
returns an an image with the same shape as the input image
loop over an image and do something with each pixels
Args:
image : input image
returns an an image with the same shape as the input image
loop over an image and do something with each pixels
Args:
image : input imagereturns an an image with the same shape as the input image
loop over an image and do something with each pixels
Args:
image : input image
returns an an image with the same shape as the input image
loop over an image and do something with each pixels
Args:
image : input imagereturns an an image with the same shape as the input image
loop over an image and do something with each pixels
Args:
image : input image
returns an an image with the same shape as the input image
Perform edgepreserving smoothing at the given scale.
For details see nonlinearDiffusion in the vigra C++ documentation.
Perform normalized convolution of an image. If the image has multiple channels, every channel is convolved independently. The ‘mask’ tells the algorithm whether input pixels are valid (nonzero mask value) or not. Invalid pixels are ignored in the convolution. The mask must have one channel (which is then used for all channels input channels) or as many channels as the input image.
For details, see normalizedConvolveImage in the C++ documentation.
Find centers of radial symmetry in an 2D image.
This algorithm implements the Fast Radial Symmetry Transform according to [G. Loy, A. Zelinsky: “A Fast Radial Symmetry Transform for Detecting Points of Interest”, in: A. Heyden et al. (Eds.): Proc. of 7th European Conf. on Computer Vision, Part 1, pp. 358368, Springer LNCS 2350, 2002]
For details see radialSymmetryTransform in the vigra C++ documentation.
Perform 2D convolution with a firstorder recursive filter with parameter ‘b’ and given ‘borderTreatment’. ‘b’ must be between 1 and 1.
For details see recursiveFilterX and recursiveFilterY (which this function calls in succession) in the vigra C++ documentation.
Perform 2D convolution with a secondorder recursive filter with parameters ‘b1’ and ‘b2’. Border treatment is always BORDER_TREATMENT_REFLECT.
For details see recursiveFilterX and recursiveFilterY (which this function calls in succession) in the vigra C++ documentation.
Compute a fast approximate Gaussian smoothing of a 2D scalar or multiband image.
This function uses the thirdorder recursive filter approximation to the Gaussian filter proposed by Young and van Vliet. Each channel of the array is smoothed independently. If ‘sigma’ is a single value, an isotropic Gaussian filter at this scale is applied (i.e. each dimension is smoothed in the same way). If ‘sigma’ is a tuple of values, the amount of smoothing will be different for each spatial dimension. The length of the tuple must be equal to the number of spatial dimensions.
For details see recursiveGaussianFilterLine in the vigra C++ documentation.
Compute the gradient of a scalar image using a recursive (exponential) filter at the given ‘scale’. The output image (if given) must have two channels.
For details see recursiveSmoothLine and recursiveFirstDerivativeLine (which this function calls internally) in the vigra C++ documentation.
Compute the gradient of a 2D scalar or multiband image using a recursive (exponential) filter at the given ‘scale’. The output image (if given) must have as many channels as the input.
For details see recursiveSmoothLine and recursiveSecondDerivativeLine (which this function calls internally) in the vigra C++ documentation.
Calls recursiveFilter2D() with b = exp(1/scale), which corresponds to smoothing with an exponential filter exp(abs(x)/scale).
For details see recursiveSmoothLine in the vigra C++ documentation.
Calculate Riesz transforms of the Laplacian of Gaussian.
For details see rieszTransformOfLOG in the vigra C++ documentation.
Perform edgepreserving smoothing at the given scale.
For details see shockFilter_ in the vigra C++ documentation.
Perform simple sharpening function.
For details see simpleSharpening in the vigra C++ documentation.
Skeletonize all regions in the given label image. Each skeleton receives the label of the corresponding region, unless ‘length’ or ‘salience’ are requested, in which case the skeleton points hold real numbers. Nonskeleton points always have the value zero. When the input image contains label zero, it is always considered background and therefore ignored. The ‘mode’ must be one of the following strings:
‘DontPrune’: don’t remove any branches
 ‘ReturnLength’: mark each pixel with the length of the longest branch
it belongs to
 ‘PruneLength’: remove all branches that are shorter than the given
‘pruning_threshold’
 ‘PruneLengthRelative’: remove all branches that are shorter than the
fraction specified in ‘pruning_threshold’ of the longest branch in the present region
 ‘ReturnSalience’: mark each pixel with the salience of the longest branch
it belongs to
 ‘PruneSalience’: remove all branches whose salience is less than the given
‘pruning_threshold’
 ‘PruneSalienceRelative’: remove all branches whose salience is less than the
fraction specified in ‘pruning_threshold’ of the most salient branch in the present region (default with pruning_threshold=0.2)
 ‘PruneTopology’: prune all branches that are not essential for the topology,
but keep the skeleton center
‘PruneAggressive’: like ‘PruneTopology’, but don’t necessarily preserve the center
For details see skeletonizeImage in the vigra C++ documentation.
Calculate the structure tensor of an array (up to 5D) by means of Gaussian (derivative) filters at the given scales. If the input has multiple channels, the structure tensors of each channel are added to get the result. The result has N*(N+1)/2 channels representing the flattened upper triangular part of the structure tensor (N is the dimension of the input).
If ‘innerScale’ and ‘outerScale’ are single values, isotropic filters at these scales are applied (i.e., each dimension is filtered in the same way). SIf ‘innerScale’ and/or ‘outerScale’ are are tuples or lists of values, the amount of smoothing will be different for each spatial dimension. The optional ‘sigma_d’ (single, tuple, or list) denotes the PSF standard deviation per axis, the optional ‘step_size’ (single, tuple, or list) the distance between two adjacent pixels for each dimension. The length of the tuples or lists must be equal to the number of spatial dimensions.
‘window_size’ and ‘roi’ have the same meaning as in gaussianSmoothing().
For details see structureTensorMultiArray and ConvolutionOptions in the vigra C++ documentation.
Compute the eigenvalues of the structure tensor at the given scales for a scalar or multichannel image or volume.
Calls structureTensor() and tensorEigenvalues().
Calculate gradient of a scalar 2D image using symmetric difference filters. The optional tuple or list ‘step_size’ denotes the distance between two adjacent pixels for each dimension; its length must be equal to the number of spatial dimensions.
‘roi’ has the same meaning as in gaussianSmoothing().
For details see symmetricGradientMultiArray and ConvolutionOptions in the vigra C++ documentation.
Calculate the elementwise determinant of an array which stores the flattened upper triangular part of a symmetric tensor in each element (e.g. the output of structureTensor()).
For details see tensorDeterminantMultiArray in the vigra C++ documentation.
Calculate eigen representation of a symmetric 2x2 tensor.
For details see tensorEigenRepresentation in the vigra C++ documentation.
Calculate the eigenvalues in every element of an array which stores the flattened upper triangular part of a symmetric tensor in each element (e.g. the output of structureTensor()). The result has as many channels (= eigenvalues) as the spatial dimension of the input.
For details see tensorEigenvaluesMultiArray in the vigra C++ documentation.
Calculate the elementwise trace of an array which stores the flattened upper triangular part of a symmetric tensor in each element (e.g. the output of structureTensor()).
For details see tensorTraceMultiArray in the vigra C++ documentation.
Perform total variation filter on 2D single band images.
For details see totalVariationFilter in the vigra C++ documentation.
Perform weighted total variation filter on 2D single band images.
For details see totalVariationFilter in the vigra C++ documentation.
Compute the Euclidean vector distance transform of a scalar array (up to 3D). The function returns an array with a many channels as the input dimension.
In contrast to the plain distance transform, this function returns the difference vector of each background pixel to the nearest foreground pixel (when ‘background=True’, the default), or the other way around (when ‘background=False’). Otherwise, this function behaves like distanceTransform() (which just returns the magnitude of the difference vectors).
For more detailed documentation, see distanceTransform2D() and separableVectorDistance in the vigra C++ documentation.
Turn a vector valued 2D or 3D array (e.g. the gradient array) into a tensor array by computing the outer product in every pixel.
For details see vectorToTensorMultiArray in the vigra C++ documentation.
The module vigra.sampling contains methods to change the number and/or location of the image sampling points, such as resizing, rotation, and interpolation.
Resample an image by the given ‘factor’
The ‘out’ parameter must have, if given, the according dimensions. This function also works for multiband images, it is then executed on every band.
For more details, see resampleImage in the vigra C++ documentation.
Resample image using a gaussian filter:
resamplingGaussian(image,
sigmaX=1.0, derivativeOrderX=0, samplingRatioX=2.0, offsetX=0.0,
sigmaY=1.0, derivativeOrderY=0, samplingRatioY=2.0, offsetY=0.0,
out=None)
This function utilizes resamplingConvolveImage with a Gaussianfilter (see the vigra C++ documentation for details).
Resize image or volume using Bspline interpolation.
The spline order is given in the parameter ‘order’. The desired shape of the output array is taken either from ‘shape’ or ‘out’. If both are given, they must agree. This function also works for multichannel data, it is then executed on every channel independently.
For more details, see resizeImageSplineInterpolation and resizeMultiArraySplineInterpolation in the vigra C++ documentation.
resize( (object)image [, (object)shape=None [, (int)order=3 [, (object)out=None]]]) > object
Resize image using the Catmull/Rom interpolation function.
The desired shape of the output image is taken either from ‘shape’ or ‘out’. If both are given, they must agree. This function also works for multiband images, it is then executed on every band.
For more details, see resizeImageCatmullRomInterpolation in the vigra C++ documentation.
Resize image using the Coscot interpolation function.
The desired shape of the output image is taken either from ‘shape’ or ‘out’. If both are given, they must agree. This function also works for multiband images, it is then executed on every band.
For more details, see resizeImageCoscotInterpolation in the vigra C++ documentation.
Resize image using linear interpolation. The function uses the standard separable bilinear interpolation algorithm to obtain a good compromise between quality and speed.
The desired shape of the output image is taken either from ‘shape’ or ‘out’. If both are given, they must agree. This function also works for multiband images, it is then executed on every band.
For more details, see resizeImageLinearInterpolation in the vigra C++ documentation.
Resize image by repeating the nearest pixel values.
The desired shape of the output image is taken either from ‘shape’ or ‘out’. If both are given, they must agree. This function also works for multiband images, it is then executed on every band.
For more details, see resizeImageNoInterpolation in the vigra C++ documentation.
Resize image using Bspline interpolation.
The spline order is given in the parameter ‘order’. The desired shape of the output image is taken either from ‘shape’ or ‘out’. If both are given, they must agree. This function also works for multiband images, it is then executed on every band.
For more details, see resizeImageSplineInterpolation in the vigra C++ documentation.
Resize volume using Bspline interpolation.
The spline order is given in the parameter ‘order’. The dimensions of the output volume is taken either from ‘shape’ or ‘out’. If both are given, they must agree. This function also works for multiband volumes, it is then executed on every band.
For more details, see resizeMultiArraySplineInterpolation in the vigra C++ documentation.
Rotate an image by an arbitrary angle using splines for interpolation around its center.
The angle may be given in degree (parameter degree). The parameter ‘splineOrder’ indicates the order of the splines used for interpolation. If the ‘out’ parameter is given, the image is cropped for it’s dimensions. If the ‘out’ parameter is not given, an output image with the same dimensions as the input image is created.
For more details, see GeometricTransformations.rotationMatrix2DDegrees in the vigra C++ documentation.
Rotate an image by an arbitrary angle around its center using splines for interpolation.
The angle may be given in radiant (parameter radiant). The parameter ‘splineOrder’ indicates the order of the splines used for interpolation. If the ‘out’ parameter is given, the image is cropped for it’s dimensions. If the ‘out’ parameter is not given, an output image with the same dimensions as the input image is created.
For more details, see GeometricTransformations.rotationMatrix2DRadians in the vigra C++ documentation.
Rotate an image by a multiple of 90 degrees.
The ‘orientation’ parameter (which must be one of CLOCKWISE, COUNTER_CLOCKWISE and UPSIDE_DOWN indicates the rotation direction. The ‘out’ parameter must, if given, have the according dimensions. This function also works for multiband images, it is then executed on every band.
For more details, see rotateImage in the vigra C++ documentation.
Spline image views implement an interpolated view for an image which can be accessed at realvalued coordinates (in contrast to the plain image, which can only be accessed at integer coordinates). Module vigra.sampling defines:
SplineImageView0
SplineImageView1
SplineImageView2
SplineImageView3
SplineImageView4
SplineImageView5
The number denotes the spline interpolation order of the respective classes. Below, we describe SplineImageView3 in detail, but the other classes work analogously. See SplineImageView in the C++ documentation for more detailed information.
Construct a SplineImageView for the given image:
SplineImageView(image, skipPrefilter = False)
Currently, 'image' can have dtype numpy.uint8, numpy.int32, and numpy.float32. If 'skipPrefilter' is True, image values are directly used as spline coefficients, so that the view performs approximation rather than interploation.
__init__( (object)arg1, (object)arg2) > object
__init__( (object)arg1, (object)arg2) > object
__init__( (object)arg1, (object)arg2, (bool)arg3) > object
__init__( (object)arg1, (object)arg2, (bool)arg3) > object
__init__( (object)arg1, (object)arg2, (bool)arg3) > object
Return first derivative in x direction at a realvalued coordinate.
SplineImageView.dx(x, y) > value
Return third derivative in x direction at a realvalued coordinate.
SplineImageView.dx3(x, y) > value
Like dx3(), but returns an entire image with the given sampling factors. For example,
SplineImageView.dx3Image(2.0, 2.0) > image
creates an derivative image with twofold oversampling in both directions.
Like dx(), but returns an entire image with the given sampling factors. For example,
SplineImageView.dxImage(2.0, 2.0) > image
creates an derivative image with twofold oversampling in both directions.
Return second derivative in x direction at a realvalued coordinate.
SplineImageView.dxx(x, y) > value
Like dxx(), but returns an entire image with the given sampling factors. For example,
SplineImageView.dxxImage(2.0, 2.0) > image
creates an derivative image with twofold oversampling in both directions.
Return mixed third derivative at a realvalued coordinate.
SplineImageView.dxxy(x, y) > value
Like dxxy(), but returns an entire image with the given sampling factors. For example,
SplineImageView.dxxyImage(2.0, 2.0) > image
creates an derivative image with twofold oversampling in both directions.
Return mixed second derivative at a realvalued coordinate.
SplineImageView.dxy(x, y) > value
Like dxy(), but returns an entire image with the given sampling factors. For example,
SplineImageView.dxyImage(2.0, 2.0) > image
creates an derivative image with twofold oversampling in both directions.
Return mixed third derivative at a realvalued coordinate.
SplineImageView.dxyy(x, y) > value
Like dxyy(), but returns an entire image with the given sampling factors. For example,
SplineImageView.dxyyImage(2.0, 2.0) > image
creates an derivative image with twofold oversampling in both directions.
Return first derivative in y direction at a realvalued coordinate.
SplineImageView.dy(x, y) > value
Return third derivative in y direction at a realvalued coordinate.
SplineImageView.dy3(x, y) > value
Like dy3(), but returns an entire image with the given sampling factors. For example,
SplineImageView.dy3Image(2.0, 2.0) > image
creates an derivative image with twofold oversampling in both directions.
Like dy(), but returns an entire image with the given sampling factors. For example,
SplineImageView.dyImage(2.0, 2.0) > image
creates an derivative image with twofold oversampling in both directions.
Return second derivative in y direction at a realvalued coordinate.
SplineImageView.dyy(x, y) > value
Like dyy(), but returns an entire image with the given sampling factors. For example,
SplineImageView.dyyImage(2.0, 2.0) > image
creates an derivative image with twofold oversampling in both directions.
SplineImageView.facetCoefficients(x, y) > matrix
Return the facet coefficient matrix so that spline values can be computed explicitly. The matrix has size (order+1)x(order+1), where order is the order of the spline. The matrix must be multiplied from left and right with the powers of the local facet x and ycoordinates respectively (note that local facet coordinates are in the range [0,1] for odd order splines and [0.5, 0.5] for even order splines).
Usage for odd spline order:
s = SplineImageView3(image) c = s.coefficients(10.1, 10.7) x = matrix([1, 0.1, 0.1**2, 0.1**3]) y = matrix([1, 0.7, 0.7**2, 0.7**3]) assert abs(x * c * y.T  s[10.1, 10.7]) < smallNumber
Usage for even spline order:
s = SplineImageView2(image) c = s.coefficients(10.1, 10.7) x = matrix([1, 0.1, 0.1**2]) y = matrix([1, 0.3, (0.3)**2]) assert abs(x * c * y.T  s[10.1, 10.7]) < smallNumber
Return gradient squared magnitude at a realvalued coordinate.
SplineImageView.g2(x, y) > value
Like g2(), but returns an entire image with the given sampling factors. For example,
SplineImageView.g2Image(2.0, 2.0) > image
creates an derivative image with twofold oversampling in both directions.
Return first derivative in x direction of the gradient squared magnitude at a realvalued coordinate.
SplineImageView.g2x(x, y) > value
Like g2x(), but returns an entire image with the given sampling factors. For example,
SplineImageView.g2xImage(2.0, 2.0) > image
creates an derivative image with twofold oversampling in both directions.
Return first derivative in y direction of the gradient squared magnitude at a realvalued coordinate.
SplineImageView.g2y(x, y) > value
Like g2y(), but returns an entire image with the given sampling factors. For example,
SplineImageView.g2yImage(2.0, 2.0) > image
creates an derivative image with twofold oversampling in both directions.
The height of the underlying image.
Return an interpolated image or derivative image with the given sampling factors and derivative orders. For example, we get a twofold oversampled image with the xderivatives in each pixel by:
SplineImageView.interpolatedImage(2.0, 2.0, 1, 0) > image
Check if a coordinate is inside the underlying image.
SplineImageView.isInside(x, y) > bool
Check if a coordinate is within the valid range of the SplineImageView.
SplineImageView.isValid(x, y) > bool
Thanks to reflective boundary conditions, the valid range is three times as big as the size of the underlying image.
The shape of the underlying image.
The width of the underlying image.
Bases: list
Create a new pyramid. The new pyramid levels range from ‘lowestLevel’ to ‘highestLevel’ (inclusive), and the given ‘image’ is copied to ‘copyImageToLevel’. The images at other levels are filled with zeros and sized so that the shape is reduced by half when going up (to higher levels), and doubled when going down.
This class can handle multichannel images, but only when image.channelIndex exists and returns image.ndim1 (i.e. the image must have axistags, and the channel axis must correspond to the last index, as in C or Vorder).
The axistags of the images in this pyramid.
The channel dimension of the images in this pyramid. If the images have no axistags, or no channel axis is specified, this defaults to ‘ndim’.
Make sure that ‘level’ exists. If ‘level’ is outside the current range of levels, empty images of the appropriate shape are inserted into the pyramid.
The pixel type of the images in this pyramid.
Expand the image at ‘srcLevel’ to ‘destLevel’, using the Burt smoothing filter with the given ‘centerValue’. srcLevel must be larger than destLevel.
For more details, see pyramidExpandBurtFilter in the C++ documentation.
Expand the image at ‘srcLevel’ to ‘destLevel’, using the Burt smoothing filter with the given ‘centerValue’, and reconstruct the images for the levels srcLevel1 ... destLevel from their Laplacian images. srcLevel must be larger than destLevel.
For more details, see pyramidExpandBurtLaplacian in the C++ documentation.
The pyramids highest level (inclusive).
The pyramids lowest level.
The dimension of the images in this pyramid.
Reduce the image at ‘srcLevel’ to ‘destLevel’, using the Burt smoothing filter with the given ‘centerValue’. srcLevel must be smaller than destLevel.
For more details, see pyramidReduceBurtFilter in the C++ documentation.
Reduce the image at ‘srcLevel’ to ‘destLevel’, using the Burt smoothing filter with the given ‘centerValue’, and compute Laplacian images for the levels srcLevel ... destLevel1. srcLevel must be smaller than destLevel.
For more details, see pyramidReduceBurtLaplacian in the C++ documentation.
The module vigra.fourier contains functions for Fourier transforms, Cosine/Sine transforms, and Fourierdomain filters.
Calculate sensible angular sigma for given parameters.
Create a 2dimensional gabor filter in frequency space.
Perform 2dimensional Fourier transformation of a scalar float32 image.If the input array has multiple channels, each channel is transformed separately.
Perform 2dimensional inverse Fourier transformation of a complex64 array.If the input array has multiple channels, each channel is transformed separately.
Calculate sensible radial sigma for given parameters.
The module vigra.analysis contains segmentation algorithms (e.g. watershed), edge and corner detection, localization of maxima and minima etc.
Represent an Edgel at a particular subpixel position (x, y), having given ‘strength’ and ‘orientation’.
For details, see Edgel in the vigra C++ documentation.
Standard constructor:
Edgel()
Constructor:
Edgel(x, y, strength, orientation)
The edgel’s orientation.
The edgel’s strength.
The edgel’s x position.
The edgel’s y position.
An instance of this accumulator class is returned by extractFeatures(). The object contains the computed features (i.e. the selected features and their dependencies).
Raises an exception This class cannot be instantiated from Python
Returns a list with the names of all computed features.
Create an empty accumulator with the same active features as ‘self’. This is useful for merging.
Returns True if ‘feature’ has been computed and False otherwise.
Returns a list with the names of all computed features.
Merge features with the features from accumulator ‘other’. Raises a TypeError when ‘other’ is incompatible with ‘self’.
Returns a list of all supported features for the given input data array.
An instance of this accumulator class is returned by extractRegionFeatures() and contains the computed global and perregion features.
Raises an exception This class cannot be instantiated from Python
Returns a list with the names of all selected features.
Create an empty accumulator with the same active features as ‘self’. This is useful for merging.
Returns True if ‘feature’ has been computed and False otherwise.
Returns a list with the names of all selected features.
Return the highest region label in this accumulator.
Merge features with the features from accumulator ‘other’. ‘self’ and ‘other’ must have the same maxRegionLabel`(), or ‘self’ must be an empty accumulator (as returned by `create).
Returns a list with the names of all supported features for the given input arrays.
Map all values in labels to new values using the given mapping (a dict). Useful for maps with large values, for which a numpy index array would need too much RAM. To relabel inplace, set out=labels.
labels: ndarray mapping: dict of {old_label : new_label} allow_incomplete_mapping: If True, then any voxel values in the original data that are missing
from the mapping dict will be copied (and casted) into the output. Otherwise, an IndexError will be raised if the map is incomplete for the input data.
 out: ndarray to hold the data. If None, it will be allocated for you.
 The dtype of out is allowed to be smaller (or bigger) than the dtype of labels.
Note: As with other vigra functions, you should provide accurate axistags for optimal performance.
applyMapping( (object)src, (dict)mapping [, (bool)allow_incomplete_mapping=False [, (object)out=None]]) > object
applyMapping( (object)src, (dict)mapping [, (bool)allow_incomplete_mapping=False [, (object)out=None]]) > object
applyMapping( (object)src, (dict)mapping [, (bool)allow_incomplete_mapping=False [, (object)out=None]]) > object
applyMapping( (object)src, (dict)mapping [, (bool)allow_incomplete_mapping=False [, (object)out=None]]) > object
applyMapping( (object)src, (dict)mapping [, (bool)allow_incomplete_mapping=False [, (object)out=None]]) > object
applyMapping( (object)src, (dict)mapping [, (bool)allow_incomplete_mapping=False [, (object)out=None]]) > object
applyMapping( (object)src, (dict)mapping [, (bool)allow_incomplete_mapping=False [, (object)out=None]]) > object
applyMapping( (object)src, (dict)mapping [, (bool)allow_incomplete_mapping=False [, (object)out=None]]) > object
applyMapping( (object)src, (dict)mapping [, (bool)allow_incomplete_mapping=False [, (object)out=None]]) > object
applyMapping( (object)src, (dict)mapping [, (bool)allow_incomplete_mapping=False [, (object)out=None]]) > object
applyMapping( (object)src, (dict)mapping [, (bool)allow_incomplete_mapping=False [, (object)out=None]]) > object
applyMapping( (object)src, (dict)mapping [, (bool)allow_incomplete_mapping=False [, (object)out=None]]) > object
applyMapping( (object)src, (dict)mapping [, (bool)allow_incomplete_mapping=False [, (object)out=None]]) > object
applyMapping( (object)src, (dict)mapping [, (bool)allow_incomplete_mapping=False [, (object)out=None]]) > object
applyMapping( (object)src, (dict)mapping [, (bool)allow_incomplete_mapping=False [, (object)out=None]]) > object
applyMapping( (object)src, (dict)mapping [, (bool)allow_incomplete_mapping=False [, (object)out=None]]) > object
applyMapping( (object)src, (dict)mapping [, (bool)allow_incomplete_mapping=False [, (object)out=None]]) > object
applyMapping( (object)src, (dict)mapping [, (bool)allow_incomplete_mapping=False [, (object)out=None]]) > object
applyMapping( (object)src, (dict)mapping [, (bool)allow_incomplete_mapping=False [, (object)out=None]]) > object
applyMapping( (object)src, (dict)mapping [, (bool)allow_incomplete_mapping=False [, (object)out=None]]) > object
applyMapping( (object)src, (dict)mapping [, (bool)allow_incomplete_mapping=False [, (object)out=None]]) > object
applyMapping( (object)src, (dict)mapping [, (bool)allow_incomplete_mapping=False [, (object)out=None]]) > object
applyMapping( (object)src, (dict)mapping [, (bool)allow_incomplete_mapping=False [, (object)out=None]]) > object
applyMapping( (object)src, (dict)mapping [, (bool)allow_incomplete_mapping=False [, (object)out=None]]) > object
applyMapping( (object)src, (dict)mapping [, (bool)allow_incomplete_mapping=False [, (object)out=None]]) > object
applyMapping( (object)src, (dict)mapping [, (bool)allow_incomplete_mapping=False [, (object)out=None]]) > object
Beautify crack edge image for visualization.
For details see beautifyCrackEdgeImage in the vigra C++ documentation.
Detect and mark edges in an edge image using Canny’s algorithm.
For details see cannyEdgeImage in the vigra C++ documentation.
Detect and mark edges in an edge image using Canny’s algorithm.
For details see cannyEdgeImage in the vigra C++ documentation.
Detect and mark edges in an edge image using Canny’s algorithm.
For details see cannyEdgeImageWithThinning in the vigra C++ documentation.
Return a list of Edgel objects whose strength is at least ‘threshold’.
The function comes in two forms:
cannyEdgelList(gradient, threshold) > list cannyEdgelList(image, scale, threshold) > listThe first form expects a gradient image (i.e. with two channels) to compute edgels, whereas the second form expects a scalar image and computes the gradient internally at ‘scale’.
For details see cannyEdgelList in the vigra C++ documentation.
Return a list of Edgel objects whose strength is at least ‘threshold’.
The function comes in two forms:
cannyEdgelList3x3(gradient, threshold) > list cannyEdgelList3x3(image, scale, threshold) > listThe first form expects a gradient image (i.e. with two channels) to compute edgels, whereas the second form expects a scalar image and computes the gradient internally at ‘scale’. The results are slightly better than those of cannyEdgelList().
For details see cannyEdgelList3x3 in the vigra C++ documentation.
Close onepixel wide gaps in a cell grid edge image.
For details see closeGapsInCrackEdgeImage in the vigra C++ documentation.
Find corners in a scalar 2D image using the method of Beaudet at the given ‘scale’.
For details see beaudetCornerDetector in the vigra C++ documentation.
Find corners in a scalar 2D image using the boundary tensor at the given ‘scale’.
Specifically, the cornerness is defined as twice the small eigenvalue of the boundary tensor.
For details see boundaryTensor in the vigra C++ documentation.
Find corners in a scalar 2D image using the method of Foerstner at the given ‘scale’.
For details see foerstnerCornerDetector in the vigra C++ documentation.
Find corners in a scalar 2D image using the method of Harris at the given ‘scale’.
For details see cornerResponseFunction in the vigra C++ documentation.
Find corners in a scalar 2D image using the method of Rohr at the given ‘scale’.
For details see rohrCornerDetector in the vigra C++ documentation.
Find local maxima and maximal plateaus in an image and mark them with the given ‘marker’. Parameter ‘neighborhood’ specifies the pixel neighborhood to be used and can be 4 or 8 (default).
For details see extendedLocalMaxima in the vigra C++ documentation.
Find local maxima and maximal plateaus in a volume and mark them with the given ‘marker’. Parameter ‘neighborhood’ specifies the pixel neighborhood to be used and can be 6 (default) or 26 .
For details see extendedLocalMaxima3D in the vigra C++ documentation.
Find local minima and minimal plateaus in an image and mark them with the given ‘marker’. Parameter ‘neighborhood’ specifies the pixel neighborhood to be used and can be 4 or 8 (default).
For details see extendedLocalMinima in the vigra C++ documentation.
Find local minima and minimal plateaus in a volume and mark them with the given ‘marker’. Parameter ‘neighborhood’ specifies the pixel neighborhood to be used and can be 6(default) or 26 .
For details see extendedLocalMinima3D in the vigra C++ documentation.
Extract global features (e.g. Mean, Variance, Minimum, etc.) from the input array (‘image’ or ‘volume’). An accumulator object of type FeatureAccumulator is returned that holds the computed features.
The overloaded function extractFeatures() supports 2D or 3D arrays with arbitrary many channels. The element type of the input array must be dtype=numpy.float32. The set of available features depends on the input array. The ‘Histogram’ feature, for example, is only supported for singleband arrays. Call supportedFeatures() with the same input array to get a list of all available features for this input.
The argument ‘features’ can take the following values:
‘all’: compute all supported features (default)
name: compute a single feature (and its dependencies)
[name1, name2,...]: compute the given features plus dependencies
 None or ‘’: return an empty accumulator, whose method
supportedFeatures() tells you the list of supported features for the given input array.
To compute perregion features, use extractRegionFeatures().
This overload is called for 2D input arrays two or more than four channels. Histograms and quantiles are not supported for this input.
For further details about the meaning of the features, see Feature Accumulators in the vigra C++ documentation.
extractFeatures( (object)image [, (object)features=’all’ [, (object)histogramRange=’globalminmax’ [, (int)binCount=64]]]) > FeatureAccumulator :
This overload of extractFeatures() computes global statistics for a 2D scalar input array, e.g. vigra.ScalarImage
Features ‘Histogram’ and ‘Quantiles’ are supported for this input. Options are:
histogramRange: lower and upper bound of the histogram
‘globalminmax’: compute and use global minimum/maximum (default)
 [lower, upper]: provide explicit bounds (float numbers),
useful to ensure that merge will be allowed.
binCount: number of bins (default: 64).
Histogram options are ignored when the histogram feature is not selected. Quantiles (0%, 10%, 25%, 50%, 75%, 90%, 100%) are computed from the specified histogram.
Extract region features from an input array with dtype=numpy.float32 and return a RegionFeatureAccumulator object.
Membership of the array elements (pixels) to regions is specified by a ‘labels’ array with element type dtype=uint32.
The set of available features depends on the input array. Call supportedRegionFeatures() with the same input and label arrays to get a list of all available features for these inputs.
The argument ‘features’ can take the following values:
‘all’: compute all supported features (default)
name: compute a single feature (and its dependencies)
[name1, name2,...]: compute the given features plus dependencies
 None or ‘’: return an empty accumulator, whose method
supportedFeatures() tells you the list of supported features for the given input array.
When the feature name starts with ‘Global’, the feature is computed globally, i.e. without considering region membership.
The argument ‘ignoreLabel’ is useful when the label array contains a background region (usually label 0) that should be ignored during feature computation. If ‘ignoreLabel’ is None (the default), all region labels are used.
This overload is called for 2D input arrays with two or more than four channels. Histograms and quantiles are not supported for this input.
For further details about the meaning of the features, see Feature Accumulators in the vigra C++ documentation.
extractRegionFeatures( (object)image, (object)labels [, (object)features=’all’ [, (object)histogramRange=’globalminmax’ [, (int)binCount=64 [, (object)ignoreLabel=None]]]]) > RegionFeatureAccumulator :
This overload of extractRegionFeatures() computes region statistics for a scalar 2D input array, e.g. vigra.ScalarImage.
Features ‘Histogram’ and ‘Quantiles’ are supported for this input. Options are:
histogramRange: lower and upper bound of the histogram
‘globalminmax’: compute and use global minimum/maximum (default)
‘regionminmax’: use minimum/maximum within each region
 [lower, upper]: provide explicit bounds (float numbers),
useful to ensure that merge will be allowed.
binCount: number of bins (default: 64).
Histogram options are ignored when Histogram feature is not selected. Quantiles (0%, 10%, 25%, 50%, 75%, 90%, 100%) are computed from the specified histogram.
Extract skeleton features for each region of a labeled 2D image (with dtype=numpy.uint32) and return a dictionary holding the resulting feature arrays. Label 0 is always considered background and therefore skipped. The skeleton is computed using mode ‘PruneSalienceRelative’ with the given ‘pruning_threshold’.
The result dictionary holds the following keys:
‘Diameter’: the longest path between two terminals of the skeleton
‘Center’: the center point of this path
‘Terminal1’: first end point of this path
‘Terminal2’: second end point of this path
‘EuclideanDiameter’: the Euclidean distance between Terminal1 and Terminal2
‘TotalLength’: total length of the (pruned) skeleton
‘AverageLength’: the average length of the skeleton’s branches after pruning
‘BranchCount’: the number of skeleton branches (i.e. end points after pruning)
 ‘HoleCount’: the number of cycles in the skeleton
(i.e. the number of cavities in the region)
Find the connected components of a segmented image. Parameter ‘neighborhood’ specifies the pixel neighborhood to be used and can be 4 (default) or 8.
For details see labelMultiArray in the vigra C++ documentation.
Find the connected components of a segmented image, excluding the background from labeling, where the background is the set of all pixels with the given ‘background_value’. Parameter ‘neighborhood’ specifies the pixel neighborhood to be used and can be 4 (default) or 8.
For details see labelMultiArrayWithBackground in the vigra C++ documentation.
Find the connected components of a segmented multidimensional array (supported dimensions: 2 to 5). Parameter ‘neighborhood’ specifies the pixel neighborhood to be used and can be ‘direct’ (default) or ‘indirect’ or the exact number of neighbors (2D: 4 or 8, 3D: 6 or 26, 4D: 8 or 80, 5D: 10 or 242).
For details see labelMultiArray in the vigra C++ documentation.
Find the connected components of a segmented multidimensional array (supported dimensions: 2 to 5), excluding the background from labeling, where background is the set of all pixels with the given ‘background_value’. Parameter ‘neighborhood’ specifies the pixel neighborhood to be used and can be ‘direct’ (default) or ‘indirect’ or the exact number of neighbors (2D: 4 or 8, 3D: 6 or 26, 4D: 8 or 80, 5D: 10 or 242).
For details see labelMultiArrayWithBackground in the vigra C++ documentation.
Find the connected components of a segmented volume. Parameter ‘neighborhood’ specifies the pixel neighborhood to be used and can be 6 (default) or 26.
For details see labelMultiArray in the vigra C++ documentation.
Find the connected components of a segmented volume, excluding the background from labeling, where the background is the set of all pixels with the given ‘background_value’. Parameter ‘neighborhood’ specifies the pixel neighborhood to be used and can be 6 (default) or 26.
For details see labelMultiArrayWithBackground in the vigra C++ documentation.
Find local maxima in an image and mark them with the given ‘marker’. Parameter ‘neighborhood’ specifies the pixel neighborhood to be used and can be 4 or 8 (default). If ‘allowAtBorder’ is set to ‘True’ local maxima at image border will be detected. If ‘allowPlateaus’ is set to ‘True’ regions of constant gray value whose neighbors are all lower than the value of the region will be detected.
For details see localMaxima in the vigra C++ documentation.
Find local maxima and maximal plateaus in a volume and mark them with the given ‘marker’. Parameter ‘neighborhood’ specifies the pixel neighborhood to be used and can be 6(default) or 26. If ‘allowAtBorder’ is set to ‘True’ local maxima at the volume border will be detected. If ‘allowPlateaus’ is set to ‘True’ regions of constant gray value whose neighbors are all lower than the value of the region will be detected.
For details see localMaxima in the vigra C++ documentation.
Find local minima in an image and mark them with the given ‘marker’. Parameter ‘neighborhood’ specifies the pixel neighborhood to be used and can be 4 or 8 (default). If ‘allowAtBorder’ is true local minima at the image border will be detected. If ‘allowPlateaus’ is true regions of constant gray value whose neighbors are all higher than the value of the region will be detected.
For details see localMinima in the vigra C++ documentation.
Find local minima in a volume and mark them with the given ‘marker’. Parameter 6 (default) or 26. If ‘allowAtBorder’ is set to ‘True’ local minima at the volume border will be detected. If ‘allowPlateaus’ is set to ‘True’ regions of constant gray value whose neighbors are all higher than the value of the region will be detected.
For details see localMinima in the vigra C++ documentation.
Transform a labeled uint32 image into a crack edge image.
For details see regionImageToCrackEdgeImage in the vigra C++ documentation.
Transform a labeled uint32 image into an edge image.
For details see regionImageToEdgeImage in the vigra C++ documentation.
Relabel the given label image to have consecutive label values. Note: The relative order between label values will not necessarily be preserved.
labels: ndarray start_label: The lowest label of the output array. keep_zeros: Don’t relabel zerovalued items. out: ndarray to hold the data. If None, it will be allocated for you.
A combination of uint64 labels and uint32 ‘out’ is permitted.Returns a tuple of (newlabels, maxlabel, mapping), where: maxlabel is the maximum label of the new labels, and mapping is a dict showing how the old labels were converted to the new label values.
Note: As with other vigra functions, you should provide accurate axistags for optimal performance.
relabelConsecutive( (object)labels [, (int)start_label=1 [, (bool)keep_zeros=True [, (object)out=None]]]) > tuple
relabelConsecutive( (object)labels [, (int)start_label=1 [, (bool)keep_zeros=True [, (object)out=None]]]) > tuple
relabelConsecutive( (object)labels [, (int)start_label=1 [, (bool)keep_zeros=True [, (object)out=None]]]) > tuple
relabelConsecutive( (object)labels [, (int)start_label=1 [, (bool)keep_zeros=True [, (object)out=None]]]) > tuple
relabelConsecutive( (object)labels [, (int)start_label=1 [, (bool)keep_zeros=True [, (object)out=None]]]) > tuple
relabelConsecutive( (object)labels [, (int)start_label=1 [, (bool)keep_zeros=True [, (object)out=None]]]) > tuple
relabelConsecutive( (object)labels [, (int)start_label=1 [, (bool)keep_zeros=True [, (object)out=None]]]) > tuple
relabelConsecutive( (object)labels [, (int)start_label=1 [, (bool)keep_zeros=True [, (object)out=None]]]) > tuple
relabelConsecutive( (object)labels [, (int)start_label=1 [, (bool)keep_zeros=True [, (object)out=None]]]) > tuple
relabelConsecutive( (object)labels [, (int)start_label=1 [, (bool)keep_zeros=True [, (object)out=None]]]) > tuple
relabelConsecutive( (object)labels [, (int)start_label=1 [, (bool)keep_zeros=True [, (object)out=None]]]) > tuple
Remove short edges from an edge image.
For details see removeShortEdges in the vigra C++ documentation.
shrink / ungrow a labeling / segmentation
Detect and mark edges in a crack edge image using the Shen/Castan zerocrossing detector.
For details see differenceOfExponentialCrackEdgeImage in the vigra C++ documentation.
Detect and mark edges in an edge image using the Shen/Castan zerocrossing detector.
For details see differenceOfExponentialEdgeImage in the vigra C++ documentation.
replace every occurance of each number in the array ‘seg’ with zeros if this number occures less than ‘sizeLimit’ times in the array. If ‘checkAtBorder’ is false (default) segments that touch the border of the array will not be changed. ‘maxLabel’ is the maximum label in seg
Compute Slic superpixels for a 2D or 3D image.
Parameters:
 array:
 The array on which the superpixels will be calculated. Accepts single and threeband images/volumes.
 intensityScaling:
 Scale (divide) color/intensity difference by this parameter before comparing to spatial distance.
 seedDistance:
 specify the radius of the window around each seed in which the algorithm looks for potential members of the corresponding superpixel thus limiting the superpixel size. The grid spacing for seed placement is determined by this parameter.
 minSize:
 Minimum size for superpixels. By default the algorithm merges all regions smaller than a quarter of the average superpixel size.
 iterations:
 Specify number of iterations. The default is 10.
 out:
 The label image (with dtype=numpy.uint32) to be filled by the algorithm. It will be allocated by the slicSuperpixels function if not provided)
The function returns a Python tuple (labelImage, maxRegionLabel)
Return a list of Convex Hull feature names that are available for the given label array. These Convex Hull feature names are the valid inputs to a call with extract2DConvexHullFeatures() or extract3DConvexHullFeatures. E.g., to compute just the first two features in the list, use:
f = vigra.analysis.supportedConvexHullFeatures(labels)
print("Computing Convex Hull features:", f[:2])
r = vigra.analysis.extract2DConvexHullFeatures(labels, features=f[:2])
Return a list of feature names that are available for the given array. These feature names are the valid inputs to a call of extractFeatures(). E.g., to compute just the first two features in the list, use:
f = vigra.analysis.supportedFeatures(array)
print("Computing features:", f[:2])
r = vigra.analysis.extractFeatures(array, features=f[:2])
Return a list of feature names that are available for the given array and label array. These feature names are the valid inputs to a call of extractRegionFeatures(). E.g., to compute just the first two features in the list, use:
f = vigra.analysis.supportedRegionFeatures(array, labels)
print("Computing features:", f[:2])
r = vigra.analysis.extractRegionFeatures(array, labels, features=f[:2])
Return a list of Skeleton feature names that are available for the given 2D label array. These Skeleton feature names are the valid inputs to a call of extractSkeletonFeatures(). E.g., to compute just the first two features in the list, use:
f = vigra.analysis.supportedSkeletonFeatures(labels)
print("Computing Skeleton features:", f[:2])
r = vigra.analysis.extractSkeletonFeatures(labels, features=f[:2])
Find unique values in the given label array. If sort is True, then the output is sorted. Much faster then numpy.unique().
Compute the watersheds of a 2D image.
 watersheds(image, neighborhood=4, seeds = None, methods = ‘RegionGrowing’,
 terminate=CompleteGrow, threshold=0, out = None) > (labelimage, max_ragion_label)
Parameters:
 image:
 the image or volume containing the boundary indicator values (high values = high edgeness, dtype=numpy.uint8 or numpy.float32).
 neighborhood:
the pixel neighborhood to be used. Feasible values depend on the dimension and method:
 2dimensional data:
 4 (default) or 8.
 3dimensional data:
 6 (default) or 26
 seeds:
 a label image specifying region seeds, only supported by methods ‘RegionGrowing’ and ‘Turbo’ (with dtype=numpy.uint32).
 method:
the algorithm to be used for watershed computation. Possible values:
 ‘Turbo’:
 (default if input dtype == uint8) use fastSeededRegionGrowing() (in 2D) or tws() (in 3D)
 ‘RegionGrowing’:
 (default if input dtype != uint8) use seededRegionGrowing or seededRegionGrowing3D respectively
 ‘UnionFind:
 use watershedsUnionFind or watersheds3D respectively
 terminate:
when to stop growing. Possible values:
 CompleteGrow:
 (default) grow until all pixels are assigned to a region
 KeepCountours:
 keep a 1pixel wide contour between all regions, only supported by method ‘RegionGrowing’
 StopAtThreshold:
 stop when the boundary indicator values exceed the threshold given by parameter ‘max_cost’, only supported by method ‘RegionGrowing’
 KeepCountours  StopAtThreshold:
 keep 1pixel wide contour and stop at given ‘max_cost’, only supported by method ‘RegionGrowing’
 max_cost:
 terminate growing when boundary indicator exceeds this value (ignored when ‘terminate’ is not StopAtThreshold or method is not ‘RegionGrowing’)
 out:
 the label image (with dtype=numpy.uint32) to be filled by the algorithm. It will be allocated by the watershed function if not provided)
The function returns a Python tuple (labelImage, maxRegionLabel)
Compute the watersheds of a 2D image.
 watersheds(image, neighborhood=4, seeds = None, methods = ‘RegionGrowing’,
 terminate=CompleteGrow, threshold=0, out = None) > (labelimage, max_ragion_label)
Parameters:
 image:
 the image or volume containing the boundary indicator values (high values = high edgeness, dtype=numpy.uint8 or numpy.float32).
 neighborhood:
the pixel neighborhood to be used. Feasible values depend on the dimension and method:
 2dimensional data:
 4 (default) or 8.
 3dimensional data:
 6 (default) or 26
 seeds:
 a label image specifying region seeds, only supported by methods ‘RegionGrowing’ and ‘Turbo’ (with dtype=numpy.uint32).
 method:
the algorithm to be used for watershed computation. Possible values:
 ‘Turbo’:
 (default if input dtype == uint8) use fastSeededRegionGrowing() (in 2D) or tws() (in 3D)
 ‘RegionGrowing’:
 (default if input dtype != uint8) use seededRegionGrowing or seededRegionGrowing3D respectively
 ‘UnionFind:
 use watershedsUnionFind or watersheds3D respectively
 terminate:
when to stop growing. Possible values:
 CompleteGrow:
 (default) grow until all pixels are assigned to a region
 KeepCountours:
 keep a 1pixel wide contour between all regions, only supported by method ‘RegionGrowing’
 StopAtThreshold:
 stop when the boundary indicator values exceed the threshold given by parameter ‘max_cost’, only supported by method ‘RegionGrowing’
 KeepCountours  StopAtThreshold:
 keep 1pixel wide contour and stop at given ‘max_cost’, only supported by method ‘RegionGrowing’
 max_cost:
 terminate growing when boundary indicator exceeds this value (ignored when ‘terminate’ is not StopAtThreshold or method is not ‘RegionGrowing’)
 out:
 the label image (with dtype=numpy.uint32) to be filled by the algorithm. It will be allocated by the watershed function if not provided)
The function returns a Python tuple (labelImage, maxRegionLabel)
Likewise, compute watersheds of a volume.
graphbased watershed
Compute watersheds of an image using the union find algorithm. If ‘neighborhood’ is ‘None’, it defaults to 8neighborhood for 2D inputs and 6neighborhood for 3D inputs.
Calls watersheds() with parameters:
watersheds(image, neighborhood=neighborhood, method='UnionFind', out=out)
A probability map ‘pmap’ is provided and thresholded using pmin. This results in a mask. Every connected component which has fewer pixel than ‘minMembraneSize’ is deleted from the mask. The mask is used to calculate the signed distance transformation.
From this distance transformation the segmentation is computed using a seeded watershed algorithm. The seeds are placed on the local maxima of the distanceTrafo after smoothing with ‘sigmaMinima’.
The weights of the watershed are defined by the inverse of the signed distance transform smoothed with ‘sigmaWeights’.
‘minSegmentSize’ determines how small the smallest segment in the final segmentation is allowed to be. If there are smaller ones the corresponding seeds are deleted and the watershed is done again.
If ‘cleanCloseSeeds’ is True, multiple seed points that are clearly in the same neuron will be merged with a heuristik that ensures that no seeds of two different neurons are merged.
The module vigra.geometry contains geometric primitives (such as polygons) and related algorithms.
Compute the convex hull of a point set.
For details see convexHull in the vigra C++ documentation.
The module vigra.optimization provides functions for constrained and unconstrained linear regression.
Perform linear regression with L1 regularization.
If ‘nonNegative’ is ‘True’, the solution will be constrained to nonnegative values, otherwise values may have arbitrary sign (the default). If ‘lsq’ is ‘True’, the algorithm will return the least squares solution for each active set. If ‘lasso’ is ‘True’, the LASSO solution will be returned for each active set. Both may be ‘True’ simultaneously. If ‘maxSolutionCount’ is nonzero, atr most so many active sets will be computed.
The algorithm returns a tuple:
(numActiveSets, activeSets, lsqSolutions, lassoSolutions)
where ‘numActiveSets’ specifies how many active sets have been computed, ‘activeSets’ is the list of all active sets (ordered by decreasing regularization), and ‘lsqSolutions’ and ‘lassoSolutions’ are lists of the corresponding solutions for each active set (‘lsqSolutions’ and ‘lassoSolutions’ will be ‘None’ when the corresponding function argument was ‘False’). An active set is a list of indices of all variables whose values are nonzero in the corresponding solution.
For details see leastAngleRegression in the vigra C++ documentation.
Perform plain linear regression.
For details see leastSquares in the vigra C++ documentation.
Perform linear regression where the solution is constrained to be nonnegative.
For details see nonnegativeLeastSquares in the vigra C++ documentation.
Perform linear regression with L2 regularization.
‘lambda’ is the regularization parameter  the larger it is, the more biased towards zero the solution will become.
For details see ridgeRegression in the vigra C++ documentation.
The module vigra.learning will eventually provide a wide range of machine learning tools. Right now, it only contains an implementation of the random forest classifier and probabilistic latent semantic analysis (pLSA) as an example for unsupervised learning.
Perform probabilistic latent semantic analysis.
The imput matrix ‘features’ must have shape (nFeatures*nSamples). PCA will reduce it to a smaller matrix ‘C’ with shape (nComponents*nSamples) that preserves as much information as possible. Specifically, the call:
P, C = pLSA(features, 3)
returns a projection matrix ‘P’ with shape (nComponents*nFeatures) such that the matrix f = numpy.dot(P, C) is a ranknComponents matrix that approximates the matrix ‘features’ well under the pLSA criterion. Note that the result of pLSA() is not unique, since the algorithm uses random initialization.
See pLSA in the C++ documentation for more detailed information.
Perform principal component analysis.
The imput matrix ‘features’ must have shape (nFeatures*nSamples). PCA will reduce it to a smaller matrix ‘C’ with shape (nComponents*nSamples) that preserves as much variance as possible. Specifically, the call:
P, C = principalComponents(features, 3)
returns a projection matrix ‘P’ with shape (nComponents*nFeatures) such that C = numpy.dot(numpy.transpose(P), features). Conversely, the matrix f = numpy.dot(P, C) is the best possible ranknComponents approximation to the matrix ‘features’ under the leastsquares criterion.
See principalComponents in the C++ documentation for more detailed information.
be specified explicitly, otherwise the argument will be interpreted as the number of trees to be used):
RandomForest(file_id=id, pathInFile='/path/to/dataset')
__init__( (object)arg1, (str)filename [, (str)pathInFile=’‘]) > object :
Load from HDF5 file:
RandomForest(filename, pathInFile)
__init__( (object)arg1 [, (int)treeCount=255 [, (int)mtry=1 [, (int)min_split_node_size=1 [, (int)training_set_size=0 [, (float)training_set_proportions=1.0 [, (bool)sample_with_replacement=True [, (bool)sample_classes_individually=False [, (bool)prepare_online_learning=False [, (object)labels=[]]]]]]]]]]) > object :
Construct a new random forest:
RandomForest(treeCount = 255, mtry=RF_SQRT, min_split_node_size=1, training_set_size=0, training_set_proportions=1.0, sample_with_replacement=True, sample_classes_individually=False, prepare_online_learning=False)
 treeCount:
 controls the number of trees that are created.
 labels:
 is a list specifying the permitted labels. If empty (default), the labels are automatically determined from the training data. A nonempty list is useful when some labels lack training examples.
See RandomForest_ and RandomForestOptions_ in the C++ documentation for the meaning of the other parameters.
Returns the number of features the RandomForest works with.
Returns the number of labels, the RandomForest knows.
Trains a random Forest using ‘trainData’ and ‘trainLabels’.
and returns the OOB. See the vigra documentation for the meaning af the rest of the parameters.
Train a random Forest using ‘trainData’ and ‘trainLabels’.
and returns the OOB and the Variable importanceSee the vigra documentation for the meaning af the rest of the paremeters.
Learn online.
Works only if forest has been created with prepare_online_learning=true. Needs the old training data and the new appened, starting at startIndex.
Predict labels on ‘testData’.
If a ‘nanLabel’ is provided, it will be returned for all rows of the ‘testData’ that contain an NaN value. Otherwise, an exception is thrown whenever Nan is encountered.
The output is an array containing a label for every test samples.
Predict probabilities for different classes on ‘testData’.
The output is an array containing a probability for every test sample and class.
Relearn one tree of the forest using ‘trainData’ and ‘trainLabels’.
and returns the OOB. This might be helpful in an online learning setup to improve the classifier.
Returns the ‘treeCount’, that was set when constructing the RandomForest.
Store the random forest in the given HDF5 file ‘filename’ under the internal path ‘pathInFile’.
For more information, refer to RandomForest_ in the C++ documentation.
Constructor:
RandomForestOld(trainData, trainLabels,
treeCount = 255, mtry=0, min_split_node_size=1,
training_set_size=0, training_set_proportions=1.0,
sample_with_replacement=True, sample_classes_individually=False,)
Construct and train a RandomForest using ‘trainData’ and ‘trainLabels’. ‘treeCount’ controls the number of trees that are created.
See RandomForest_ and RandomForestOptions_ in the C++ documentation for the meaning of the other parameters.
Returns the number of features the RandomForest works with.
Returns the number of labels, the RanfomForest knows.
Predict labels on ‘testData’.The output is an array containing a labels for every test samples.
Predict probabilities for different classes on ‘testData’.The output is an array containing a probability for every test sample and class.
Returns the ‘treeCount’, that was set when constructing the RandomForest.
The module vigra.noise provides noise estimation and normalization according to a method proposed by Foerstner.
Noise normalization by means of an estimated linear noise model.
For details see linearNoiseNormalization in the vigra C++ documentation.
Determine the noise variance as a function of the image intensity and cluster the results. This operator first calls noiseVarianceEstimation() to obtain a sequence of intensity/variance pairs, which are then clustered using the median cut algorithm. Then the cluster centers (i.e. average variance vs. average intensity) are determined and returned in the result sequence.
Since the length of the resulting array is not known beforhand, it cannot be written into an preallocated array (the “out” argument in most other vigra python functions) . For details see the vigra documentation noiseVarianceClustering.
Determine the noise variance as a function of the image intensity.
Returns an array with the means in the first column and the variances in the second column. Since the length of the resulting array is not known beforhand, it can not be written into an preallocated array (the “out” argument in most other vigra python functions.
For details see the vigra documentation noiseVarianceEstimation.
Noise normalization by means of an estimated nonparametric noise model.
For details see nonparametricNoiseNormalization in the vigra C++ documentation.
Noise normalization by means of an estimated quadratic noise model.
For details see quadraticNoiseNormalization in the vigra C++ documentation.
The module vigra.histogram provides histograms and channel representation
 C++ signature :
 vigra::NumpyAnyArray gaussianCoHistogram(vigra::NumpyArray<2u, float, vigra::StridedArrayTag>,vigra::NumpyArray<2u, float, vigra::StridedArrayTag>,vigra::TinyVector<float, 2>,vigra::TinyVector<float, 2>,vigra::TinyVector<int, 2>,vigra::TinyVector<float, 3> [,vigra::NumpyArray<4u, float, vigra::StridedArrayTag>=None])
gaussianCoHistogram( (object)imageA, (object)imageB, (object)minVals, (object)maxVals, (object)bins, (object)sigma [, (object)out=None]) > object :
 C++ signature :
 vigra::NumpyAnyArray gaussianCoHistogram(vigra::NumpyArray<3u, float, vigra::StridedArrayTag>,vigra::NumpyArray<3u, float, vigra::StridedArrayTag>,vigra::TinyVector<float, 2>,vigra::TinyVector<float, 2>,vigra::TinyVector<int, 2>,vigra::TinyVector<float, 3> [,vigra::NumpyArray<5u, float, vigra::StridedArrayTag>=None])
 C++ signature :
 vigra::NumpyAnyArray gaussianHistogram_(vigra::NumpyArray<2u, vigra::TinyVector<float, 1>, vigra::StridedArrayTag>,vigra::TinyVector<float, 1>,vigra::TinyVector<float, 1> [,unsigned long=30 [,float=3.0 [,float=2.0 [,vigra::NumpyArray<4u, float, vigra::StridedArrayTag>=None]]]])
gaussianHistogram_( (object)image, (object)minVals, (object)maxVals [, (int)bins=30 [, (float)sigma=3.0 [, (float)sigmaBin=2.0 [, (object)out=None]]]]) > object :
 C++ signature :
 vigra::NumpyAnyArray gaussianHistogram_(vigra::NumpyArray<2u, vigra::TinyVector<float, 3>, vigra::StridedArrayTag>,vigra::TinyVector<float, 3>,vigra::TinyVector<float, 3> [,unsigned long=30 [,float=3.0 [,float=2.0 [,vigra::NumpyArray<4u, float, vigra::StridedArrayTag>=None]]]])
gaussianHistogram_( (object)image, (object)minVals, (object)maxVals [, (int)bins=30 [, (float)sigma=3.0 [, (float)sigmaBin=2.0 [, (object)out=None]]]]) > object :
 C++ signature :
 vigra::NumpyAnyArray gaussianHistogram_(vigra::NumpyArray<3u, vigra::TinyVector<float, 1>, vigra::StridedArrayTag>,vigra::TinyVector<float, 1>,vigra::TinyVector<float, 1> [,unsigned long=30 [,float=3.0 [,float=2.0 [,vigra::NumpyArray<5u, float, vigra::StridedArrayTag>=None]]]])
gaussianHistogram_( (object)image, (object)minVals, (object)maxVals [, (int)bins=30 [, (float)sigma=3.0 [, (float)sigmaBin=2.0 [, (object)out=None]]]]) > object :
 C++ signature :
 vigra::NumpyAnyArray gaussianHistogram_(vigra::NumpyArray<3u, vigra::TinyVector<float, 3>, vigra::StridedArrayTag>,vigra::TinyVector<float, 3>,vigra::TinyVector<float, 3> [,unsigned long=30 [,float=3.0 [,float=2.0 [,vigra::NumpyArray<5u, float, vigra::StridedArrayTag>=None]]]])
gaussianHistogram_( (object)image, (object)minVals, (object)maxVals [, (int)bins=30 [, (float)sigma=3.0 [, (float)sigmaBin=2.0 [, (object)out=None]]]]) > object :
 C++ signature :
 vigra::NumpyAnyArray gaussianHistogram_(vigra::NumpyArray<3u, vigra::TinyVector<float, 10>, vigra::StridedArrayTag>,vigra::TinyVector<float, 10>,vigra::TinyVector<float, 10> [,unsigned long=30 [,float=3.0 [,float=2.0 [,vigra::NumpyArray<5u, float, vigra::StridedArrayTag>=None]]]])
The module vigra.graphs provides graphs and graph algorithms
undirected adjacency list graph
addNode( (AdjacencyListGraph)arg1, (int)arg2) > NodeAdjacencyListGraph
get the arc descriptor from the given id
number of arcs (2*edgeNum for undirected graphs)
get the edge descriptor from the given id
number of edges within the graph
find an edge between node u and v
get the id of a given node
maximum id of a valid arc in the graph
maximum id of a valid node in the graph
maximum id of a valid edge in the graph
get the node descriptor from the given id
number of nodes within the graph
number of integers needed to serialize graph
get the u node of an edge
uIds( (AdjacencyListGraph)arg1, (object)edgeIds [, (object)out=None]) > object
uvId( (AdjacencyListGraph)arg1, (int)arg2) > tuple
uvIds( (AdjacencyListGraph)arg1, (object)edgeIds [, (object)out=None]) > object
geht the v node of an edge
vIds( (AdjacencyListGraph)arg1, (object)edgeIds [, (object)out=None]) > object
agglomerative hierarchicalClustering Keyword Arguments :
graph : input graph
edgeWeights : edge weights / indicators (default : None)
 edgeLengths : length / weight of each edge (default : None)
Since we do weighted mean agglomeration, a length/weight is needed for each edge to merge 2 edges w.r.t. weighted mean. If no edgeLengths is given, ‘getEdgeLengths’ is called.
 nodeFeatures : a feature vector for each node (default: None)
A feature vector as RGB values,or a histogram for each node. Within the agglomeration, an additional edge weight will be computed from the “difference” between the features of two adjacent nodes. The metric specified in the keyword ‘metric’ is used to compute this difference
 nodeSizes : size / weight of each node (default : None)
Since we do weighted mean agglomeration, a size / weight is needed for each node to merge 2 edges w.r.t. weighted mean. If no nodeSizes is given, ‘getNodeSizes’ is called.
nodeNumStop : stop the agglomeration at a given nodeNum (default : graph.nodeNum/2)
 beta : weight between edgeWeights and nodeFeatures based edgeWeights (default:0.5) :
0.0 means only edgeWeights (from keyword edge weights) and 1.0 means only edgeWeights from nodeFeatures differences
metric : metric used to compute node feature difference (default : ‘l1’)
wardness : 0 means do not apply wards critrion, 1.0 means fully apply wards critrion (default : 1.0)
out : preallocated nodeMap for the resulting labeling (default : None)
Returns:
A node labele map encoding the segmentation
cyclesEdges( (GridGraphUndirected2d)graph, (object)graph [, (object)out=None]) > object
cyclesEdges( (GridGraphUndirected3d)graph, (object)graph [, (object)out=None]) > object
convert an image with shape = graph.shape OR shape = graph.shape *2 1 to an edge weight array
convert an image with shape = graph.shape*2  1 to an edge weight array
edge weighted seeded watersheds
Keyword Arguments :
graph : input graph
edgeWeights : evaluation weights
 seeds : node map with seeds .
For at least one node, seeds must be nonzero
backgroundLabel : a specific backgroundLabel (default : None)
backgroundBias : backgroundBias (default : None)
felzenszwalbs segmentation method
Keyword Arguments :
graph : input graph
edgeWeights : edge weights / indicators
 nodeSizes : size of each node (default: None)
If nodeSizes is None, ‘getNodeSizes’ will be called
 k : free parameter in felzenszwalbs algorithms (default : 1.0)
(todo: write better docu)
 nodeNumStop : stop the agglomeration at a given nodeNum (default :None)
If nodeNumStop is None, the resulting number of nodes does depends on k.
backgroundBias : backgroundBias (default : None)
find3Cycles( (GridGraphUndirected2d)arg1) > object
find3Cycles( (GridGraphUndirected3d)arg1) > object
find3CyclesEdges( (GridGraphUndirected2d)arg1) > object
find3CyclesEdges( (GridGraphUndirected3d)arg1) > object
get lengths/sizes of edges:
This functions will try to call ‘graph.edgeLength()’ . If this fails, an edge map filled with 1.0 will be returned
Keyword Arguments:
 graph : input graph
get size of nodes:
This functions will try to call ‘graph.nodeSize()’ . If this fails, a node map filled with 1.0 will be returned
Keyword Arguments:
 graph : input graph
Return a graph map for a given graph item ('node' , 'edge' or 'arc').
Parameters:
graph : graph to get a graph map for
item : 'node' , 'edge' or 'arc'
dtype : desired dtype
channels : number of channels (default: 1)
 addChannelDim – add an explicit channelDim :(default: False)
only useful if channels == 1
Returns:
 graphmap as numpy.ndarray / VigraArray
Return a grid graph with certain shape.
Parameters:
 shape – shape of the image
 directNeighborhood – use 4 (True) or 8 (False) neighborhood (default: True)
Returns:
 grid graph
use:
>>> # 4connected
>>> g = vigra.graps.gridGraph(shape=[10,20])
>>> g.nodeNum
200
>>> # 8connected
>>> g = vigra.graps.gridGraph(shape=[10,20],directNeighborhood=False)
get a region adjacency graph and a grid graph from a labeling.
This function will call ‘graphs.gridGraph’ and ‘graphs.regionAdjacencyGraph’
implicitMeanEdgeMap( (GridGraphUndirected2d)arg1, (object)arg2) > ImplicitMEanEdgeMap_2d_float_float
Intrinsic shape of node/edge/arcmap for a given graph.
Node edge and arc maps are stored in numpy arrays by default. The instric shape may not be confused with the number of nodes/edges/arcs. The instric shape is used to allocate a numpy are which can store data for nodes/arcs/edgeSizes of a given graph.
Parameters:
 graph : input graph to get the shape for
 item : item must be 'node' , 'edge' or 'arc'
Returns:
 shape as tuple
check if obj is gridGraph
check if obj is gridGraph
Return an empty directed graph
Parameters :
 nodes : number of nodes to reserveEdges
 edges : number of edges to reserve
Returns :
 graph
get a merge graph from input graph.
A merge graph might be usefull for hierarchical clustering
compute an edge indicator from node features .
graph : input graph
nodeFeatures : node map with feature vector for each node
an edge weight
nodeGtToEdgeGt( (GridGraphUndirected2d)graph, (object)nodeGt, (int)ignoreLabel [, (object)out=None]) > object
nodeGtToEdgeGt( (GridGraphUndirected3d)graph, (object)nodeGt, (int)ignoreLabel [, (object)out=None]) > object
node weighted seeded watersheds
Keyword Arguments :
graph : input graph
nodeWeights : node height map / evaluation weights
 seeds : node map with seeds (default: None)
If seeds are None, ‘nodeWeightedWatershedsSeeds’ will be called
generate watersheds seeds
Keyword Arguments :
 graph : input graph
 nodeWeights : node height map
 out : seed map
recursive graph smoothing to smooth node features. Each node feature is smoothed with the features of neighbor nodes. The strength of the smoothing is computed from:
“edgeIndicator > edgeThreshold ? 0 : exp(1.0*gamma*edgeIndicator)*scale”
Therefore this filter is edge preserving.
Keyword Arguments :
 graph : input graph
 nodeFeatures : node features which should be smoothed
 edgeIndicator : edge indicator
 gamma : scale edgeIndicator by gamma bevore taking the negative exponent
 scale : how much should a node be mixed with its neighbours per iteration
 iteration : how often should recursiveGraphSmoothing be called recursively
Return a region adjacency graph for a labeld graph.
Parameters:
 graph – input graph
 lables – nodemap with labels for each nodeSumWeights
 ignoreLabel – label to ingnore (default: None)
 reserveEdges – reverse a certain number of edges (default: 0)
If graph is a GridGraph or a GridRegionAdjacencyGraph, a GridRegionAdjacencyGraph will be returned. Otherwise a RegionAdjacencyGraph will be returned
node weighted seeded watersheds
Keyword Arguments :
graph : input graph
edgeWeights : edge weight map
nodeWeights : node weight map
 seeds : node map with seeds (default: None)
If seeds are None, ‘nodeWeightedWatershedsSeeds’ will be called
__init__( (object)arg1, (object)arg2, (bool)arg3) > object
get the arc descriptor from the given id
number of arcs (2*edgeNum for undirected graphs)
get the edge descriptor from the given id
node map filled with 1.0
number of edges within the graph
find an edge between node u and v
get the id of a given node
maximum id of a valid arc in the graph
maximum id of a valid node in the graph
maximum id of a valid edge in the graph
get the node descriptor from the given id
number of nodes within the graph
node map filled with 1.0
shape of grid graph
get the u node of an edge
uIds( (GridGraphUndirected2d)arg1, (object)edgeIds [, (object)out=None]) > object
uvId( (GridGraphUndirected2d)arg1, (int)arg2) > tuple
uvIds( (GridGraphUndirected2d)arg1, (object)edgeIds [, (object)out=None]) > object
geht the v node of an edge
vIds( (GridGraphUndirected2d)arg1, (object)edgeIds [, (object)out=None]) > object
__init__( (object)arg1, (object)arg2, (bool)arg3) > object
get the arc descriptor from the given id
number of arcs (2*edgeNum for undirected graphs)
get the edge descriptor from the given id
node map filled with 1.0
number of edges within the graph
find an edge between node u and v
get the id of a given node
maximum id of a valid arc in the graph
maximum id of a valid node in the graph
maximum id of a valid edge in the graph
get the node descriptor from the given id
number of nodes within the graph
node map filled with 1.0
shape of grid graph
get the u node of an edge
uIds( (GridGraphUndirected3d)arg1, (object)edgeIds [, (object)out=None]) > object
uvId( (GridGraphUndirected3d)arg1, (int)arg2) > tuple
uvIds( (GridGraphUndirected3d)arg1, (object)edgeIds [, (object)out=None]) > object
geht the v node of an edge
vIds( (GridGraphUndirected3d)arg1, (object)edgeIds [, (object)out=None]) > object
undirected adjacency list graph
addNode( (AdjacencyListGraph)arg1, (int)arg2) > NodeAdjacencyListGraph
get the arc descriptor from the given id
number of arcs (2*edgeNum for undirected graphs)
get the edge descriptor from the given id
number of edges within the graph
find an edge between node u and v
get the id of a given node
maximum id of a valid arc in the graph
maximum id of a valid node in the graph
maximum id of a valid edge in the graph
get the node descriptor from the given id
number of nodes within the graph
number of integers needed to serialize graph
get the u node of an edge
uIds( (AdjacencyListGraph)arg1, (object)edgeIds [, (object)out=None]) > object
uvId( (AdjacencyListGraph)arg1, (int)arg2) > tuple
uvIds( (AdjacencyListGraph)arg1, (object)edgeIds [, (object)out=None]) > object
geht the v node of an edge
vIds( (AdjacencyListGraph)arg1, (object)edgeIds [, (object)out=None]) > object
Region adjacency graph
Attributes:
labels : labels passed in constructor
ignoreLabel : ignoreLabel passed in constructor
 baseGraphLabels : labels passed in constructor
(fixme,dublicated attribute (see labels) )
baseGraph : baseGraph is the graph passed in constructor
 affiliatedEdges : for each edge in the region adjacency graph,
a vector of edges of the baseGraph is stored in affiliatedEdges
accumulate edge features from base graphs edges features
Keyword Argument:
edgeFeatures : edge features of baseGraph
 acc : used accumulator (default: ‘mean’)
Currently only ‘mean’ and ‘sum’ are implemented
out : preallocated edge map
accumulate edge features from base graphs edges features
Keyword Argument:
nodeFeatures : node features of baseGraph
 acc : used accumulator (default: ‘mean’)
Currently only ‘mean’ and ‘sum’ are implemented
out : preallocated node map (default: None)
project labels from current graph to baseGraph and repeat this recursively
Keyword Arguments:
steps : how often should the labels be projected back
 labels : labels for the current graph (default: None)
If labels is None, each node gets its own label
project node labels from this graph, to the base graph of this graph.
Keyword Arguments:
 labels : node labels for this graph (default: None)
If labels is None, each node gets its own label
out : preallocated node map of baseGraph (default: None)
Returns :
project node features from this graph, to the base graph of this graph.
Keyword Arguments:
 features : node feautres for this graph
 out : preallocated node map of baseGraph (default: None)
Grid Region adjacency graph
A region adjaceny graph,where the base graph should be a grid graph or a GridRegionAdjacencyGraph.
Attributes :
labels : labels passed in constructor
ignoreLabel : ignoreLabel passed in constructor
 baseGraphLabels : labels passed in constructor
(fixme,dublicated attribute (see labels) )
baseGraph : baseGraph is the graph passed in constructor
 affiliatedEdges : for each edge in the region adjacency graph,
a vector of edges of the baseGraph is stored in affiliatedEdges
 shape : shape of the grid graph which is a base graph in the
complete graph chain.
get the geometric length of the edges
get the geometric size of the nodes
project labels of this graph to the underlying grid graph.
Keyword Arguments :
 labels : node labeling of this graph (default: None)
If labels is None, each node gets its own label
project features of this graph to the underlying grid graph. Therefore project the features to an image.
Keyword Arguments :
 features : nodeFeatures of the current graph
shape of the underlying grid graph
show the graph given an RGB image
img : RGB image
If labels is None, each node gets its own label
Do not use values bigger than 1 in edgeColor.
0 means no transparency,1 means full transparency.
show the complet graph chain / hierarchy given an RGB image
img : RGB image
If labels is None, each node gets its own label
shortest path computer
Keyword Arguments:
 graph : input graph
get distance from source to target
If target node is None, the target specified by ‘run’ is used.
return the full distance map
get the shortest path from source to target
Keyword Arguments:
weights : edge weights encoding distance from two adjacent nodes
source : source node
 target : target node (default: None)
If target node is None, the target specified by ‘run’ is used.
pathType : ‘coordinates’ or ‘ids’ path (default: ‘coordinates’)
return the full predecessors map
run shortest path search
Keyword Arguments:
weights : edge weights encoding distance from two adjacent nodes
source : source node
 target : target node (default: None)
If target node is None, the shortest path to all nodes!=source is computed
run shortest path search, nodes with all edge weights larger than val will be ignored
Keyword Arguments:
 weights : edge weights encoding distance from two adjacent nodes
 source : source node
 val : upper bound
The module vigra.utilities provides utilities and tools like priority queues with changeable priorities
When you want to write your own vigranumpy extension modules, first make sure that you compile and link with the same versions of numpy and boost_python that your current vigranumpy installation uses. Otherwise, communication between new and existing modules will not work (and even crash). Then follow these steps:
Create the main module source file. This file contains the module’s ‘init’ function. Let’s assume that the module will be called ‘my_module’, and the file is ‘my_module.cxx’. A stub for ‘my_module.cxx’ typically looks like this:
// define PY_ARRAY_UNIQUE_SYMBOL (required by the numpy CAPI)
#define PY_ARRAY_UNIQUE_SYMBOL my_module_PyArray_API
// include the vigranumpy C++ API
#include <Python.h>
#include <boost/python.hpp>
#include <vigra/numpy_array.hxx>
#include <vigra/numpy_array_converters.hxx>
... // your includes
... // implementation of your wrapper functions and classes
using namespace boost::python;
// the argument of the init macro must be the module name
BOOST_PYTHON_MODULE_INIT(my_module)
{
// initialize numpy and vigranumpy
vigra::import_vigranumpy();
// export a function
def("my_function", &my_function,
(arg("arg1"), arg("arg2"), ...),
"Documentation");
// export a class and its member functions
class_<MyClass>("MyClass",
"Documentation")
.def("foo", &MyClass::foo,
(arg("arg1"), arg("arg2"), ...),
"Documentation")
;
... // more module functionality (refer to boost_python documentation)
}
When your module uses additional C++ source files, they should start with the following defines:
// this must define the same symbol as the main module file (numpy requirement)
#define PY_ARRAY_UNIQUE_SYMBOL my_module_PyArray_API
#define NO_IMPORT_ARRAY
Implement your wrapper functions. Numpy ndarrays are passed to C++ via the wrapper classes NumpyArray and NumpyAnyArray. You can influence the conversion from Python to C++ by using different instantiations of NumpyArray, as long as the Python array supports the axistags attribute (refer to axis order definitions for the meaning of the term ‘ascending order’):
// We add a 'using' declaration for brevity of our examples.
// In actual code, you should probably prefer explicit namespace qualification.
using namespace vigra;
// Accept any array type and return an arbitrary array type.
// Returning NumpyAnyArray is always safe, because at that point
// C++ no longer cares about the particular type of the array.
NumpyAnyArray foo(NumpyAnyArray array);
// Accept a 3dimensional float32 array and transpose it
// into ascending axis order ('F' order).
void foo(NumpyArray<3, float> array);
// Accept a 2dimensional float32 array with an arbitrary number of channels and
// transpose the axes into VIGRA ('V') order (channels are last, other axes ascending).
// Note that the NumpyArray dimension is 3 to account for the channel dimension.
// If the original numpy array has no channel axis, vigranumpy will automatically
// insert a singleton axis.
void foo(NumpyArray<3, Multiband<float> > array);
// Accept a 2dimensional float32 array that has only a single channel
// (that is, 'array.channels == 1' must hold on the Python side).
// Nonchannel axes are transposed into ascending order.
// Note that the NumpyArray dimension is now 2.
void foo(NumpyArray<2, Singleband<float> > array);
// Accept a float32 array that has 2 nonchannel dimensions and
// exactly 3 channels (i.e. 'array.channels == 3' on the Python side).
// Nonchannel axes are transposed into ascending order.
// Note that the NumpyArray dimension is again 2, but the pixel type is
// now a vector.
// The conversion will only succeed if the channel axis is unstrided on
// the Python side (that is, the following expression is True:
// array.strides[array.channelIndex] == array.dtype.itemsize).
void foo(NumpyArray<2, TinyVector<float, 3> > array);
void foo(NumpyArray<2, RGBValue<float> > array);
Or course, these functions can also be templated.
When your functions return newly allocated arrays, it is usually desirable to transfer the input’s axistags to the output (otherwise, vigranumpy will use defaultAxistags() as a fallback). There is a standard vigranumpy idiom for this task which assumes that the wrapped function has an optional parameter ‘output’ for a possibly preallocated output array. The axistags are then transferred by reshaping the output array with a taggedShape() (which is a combination of a shape and axistags):
NumpyAnyArray
foo(NumpyArray<3, Multiband<float32> > input,
NumpyArray<3, Multiband<float32> > output = boost::python::object())
{
// Reshape only if the output array was not explicitly passed in.
// Otherwise, use the output array as is.
output.reshapeIfEmpty(input.taggedShape(),
"error message when shape is unsuitable.");
... // your algorithm
}
It is also possible to modify the tagged shape before it is applied to the output array:
input.taggedShape()
.resize(Shape2(new_width, new_height))
.setChannelCount(new_channel_count)
.setChannelDescription("a description")
The C++ code can be multithreaded when you unlock Python’s global interpreter lock. After unlocking, your wrapper code must not call any Python functions, so the unlock statement should go after output.reshapeIfEmpty():
NumpyAnyArray
foo(NumpyArray<3, Multiband<float32> > input,
NumpyArray<3, Multiband<float32> > output = boost::python::object())
{
output.reshapeIfEmpty(input.taggedShape(), "Message.");
// Allow parallelization from here on. The destructor of
// _pythread will automatically regain the global interpreter lock
// just before this function returns to Python.
PyAllowThreads _pythread;
... // your algorithm
}
Export your wrapped functions. boost::python::def is called in its usual way, with one simple extension: Since vigranumpy does not know which NumpyArray variants you are going to use, appropriate converter functions between Python and C++ must be registered on demand. You do this by enclosing your function pointer into a call to the ‘registerConverters()’ function:
// in the module's init function
def("my_function", vigra::registerConverters(&my_function),
(arg("arg1"), ...),
"Documentation");
If you need more information, it is always a good idea to look at the source code of the existing vigranumpy modules.