Adding other padding methods

[MathiasHurst]

Hi,
is it possible to add different padding methods to ufo? Very helpful would be a padding mode as in python numpy.pad mode=median.
I am aware that speed will be reduced compared to the default opencl "padding" modes.
Bests,
Mathias

[tfarago]

Which ones would you like to have? Also specify the essence of it (median of the whole image, of the rows/cols...). Do you need them in the pad task or also in the reconstructors?

[MathiasHurst]

It is needed for Image correlation. Therefore implementation in the pad task should be sufficient. The opencl padding methods (clamp, camp_to_edge) do not allow to find the maximum of the correlation by a simple theresholding due to background effects. I tested the stuff with np.median, this uses the median of each row / column and for padding, what worked fine. However I was not able to find out which value is written into padded region where there is no column / row end see attached picture. I do not think that this plays a important role, for this we can just use the mean or median value of the image.

[tfarago]

BTW, do you have a use case when this is necessary?

[tfarago]

I am asking because for correlation just zero-padding usually works fine.

[MathiasHurst]

Hi Tomas, I correlated rescaled fresnel diffraction patterns as described e.g. in https://aip.scitation.org/doi/pdf/10.1063/1.2719653

I also tested zero padding. There was a bright gradient which was stronger than the correlation peak it self.
I attached some images to show you

Image 1 zero padded data
Image 2 Correlation output using python zero padding
Image 3 Correlation output using python median padding
Image 4 Correlation output using ufo clamp to edge padding

I can also try to zero pad in ufo but i did not find the command on read the docs there is only the link to opencls sampler_t documentation.

[tfarago]

And what happens when you apply sobel filter before the correlation?

[MathiasHurst]

much better, but there is still a stronger peak (see attached picture). I think this can be handled by cropping. But also the maximum is affected by on Pixel after edge detection filter.

[tfarago]

Don't do LR sobel, do really a 2D magnitude of the derivative and take a look at the peak.

[tfarago]

see e.g. how scikit-image does 2D sobel

[MathiasHurst]

i used this scipy.ndimage.sobel but will also try the one you proposed. LR is only the filename of the inital image

[tfarago]

Don't use scipy, it doesn't do the 2D version, or use scipy and calculate the 2D magnitude yourself.

[tfarago]

magnitude = np.sqrt(sobel(im, axis=0) ** 2 + sobel(im, axis=1) ** 2)

[MathiasHurst]

Thanks for help but it is still not sufficent :D Do you have another idea.

[tfarago]

Please send me an e-mail with a link to the two images you are trying to correlate, I will check it.

[tfarago]

It's better to first subtract the means of the two images (https://en.wikipedia.org/wiki/Cross-correlation#Zero-normalized_cross-correlation_(ZNCC).

ifft2(fft2(LR - LR.mean()) * np.conjugate(fft2(HR - HR.mean(), s=LR.shape))).real

gives the correct result below (with just zero-padding, the default when padding directly in numpy's fft2 with the s argument).

[MathiasHurst]

Thank you! Do you know how to implement this in UFO:

e.g. read path=2196Mn_LR_0p625_region1/rescaled/*.tif number=2500 ! SUBTRACT MEAN ! pad x=0 y=0 width=8192 height=8192 ! ...

I found measure metric=mean and calculate but was not able to pass the result of measure to calculate.

[tfarago]

Unfortunately, that's not possible right now.

[MathiasHurst]

Can this be implemented, or do you have an idea for a workaround?

[tfarago]

It certainly can be implemented, the question is how to make it in such a way that it's general enough for other operations (what if you want to multiply with standard deviation tomorrow?) 😉 I will need to think about it a little. In the meantime, can you remove the mean in python or fiji (both excellent scripting options)?