ksize has no effect in skimage.filters.laplace

[grlee77]

Description

The discrete Laplacian kernel used internally by skimage.filters.laplace is unaffected by the setting of ksize. The actual kernel generated is always shape (3, ) * ndim. The shape parameter used by skimage.restoration.laplacian only effects the FFT size when creating a Fourier representation of the kernel, but that argument is not even used within skimage.filters.laplace.

We should either deprecate this ksize argument or implement other sizes.

Ideally the spatial laplacian kernel generation should be split into a separate function than the Fourier transfrom used inskimage.restoration.uft.

[grlee77]

Specifically, ksize is passed as a shape parameter here:

scikit-image/skimage/filters/edges.py

Line 690 in a86701c

_, laplace_op = laplacian(image.ndim, (ksize,) * image.ndim)

but that shape is only used in the argument to the FFT for the first argument returned by laplacian while the line above only uses the second argument:

scikit-image/skimage/restoration/uft.py

Line 453 in a86701c

return ir2tf(impr, shape, is_real=is_real), impr

[grlee77]

OpenCV supports odd sizes 3, 5, 7, ... where size = 3 correspond to [-1, 2, -1] and summing the result of applying this along each axis

For ksize = 5 the kernel would be np.convolve([-1, 2, -1], [-1, 2, -1]) with additional repeated convolutions for larger sizes

For larger ksize, I am not sure where the algorithm OpenCV uses to compute the length ksize 2nd order stencils comes from.

For ksize = 1, the specific non-separable 2D kernel

[[0,  1,  0],
 [1, -4,  1],
 [0,  1,  4]]

is used. This ksize = 1 case is what we have implemented regardless of the user's ksize input.

[themnvrao76]

can I solve this issue?

[themnvrao76]

I run code and I saw that every time returns (3,3,3) laplace_op, first, i passed k_size as 3 it is returning (3,3,3) second time I passed k size as 5 still it is returning (3,3,3) so I am going to solve this bug

[grlee77]

See #5552 (comment) for a specific example showing exactly what OpenCV does.