WaveletFilterBank multiplication with FFT at highest band

InsightSoftwareConsortium / ITKIsotropicWavelets

External Module for ITK, implementing Isotropic Wavelets and Riesz Filter for multiscale phase analysis.

Apache License 2.0

13 stars 11 forks source link

WaveletFilterBank multiplication with FFT at highest band #73

Closed phcerdan closed 5 years ago

phcerdan commented 6 years ago

Why is the last slice showing traces of the first slice? Is there any default boundary condition applied? Where?

Filters involved: Check script: https://github.com/phcerdan/ITKfilters/blob/master/scripts-cpp/rieszWaveletPhaseAnalysis.cxx

ForwardWavelet
Monogenic
PhaseAnalysis
InverseWavelet

Original image, first and last slices: original_first original_last

After phase analysis:

phase_first phase_last

phcerdan commented 6 years ago

Not related to Boundary Conditions. It is related to the WaveletFilterBank in ForwardWavelet. To reproduce, use a --levels 1, --bands 1, and just visualize the high frequency output (index 0 for example). After the shrink, the artifacts are gone in the rest of outputs.

To explore:

Are we using the right frequency layout in the FilterBankGenerator? The artifacts seem frequency related?... The forwardWavelet output is the bank multiplied by the fft of a real image. Maybe there is a miss match in the factors of negative and positive frequencies?
The origin/spacing of input image (different than default {{0}}) shouldn't be involved, but I haven't completely discarded it.

phcerdan commented 6 years ago

It's not happening in ctest -R WaveletFrequencyForwardTest -E Multi -E 2D -V with tests images, so leaning into origin,spacing, buffered region issues.

phcerdan commented 6 years ago

Not origin issues, this image has default spacing/origin and still show it.

carr_tile1_64_64_32.zip

phcerdan commented 6 years ago

I just don't understand how it is working with test images and not with others. What is the difference between those?

Working test mage (in .nrrd, used to be .tiff) collagen_64x64x16.zip

phcerdan commented 5 years ago

This might not be a bug, but an unavoidable artifact for working in the FFT domain.