Padding and cropping the images - Sanity check

[aziz-ayed]

Description

When it came to integrating our denoising solution to the MCCD algorithm, we encountered an error. In fact, CFIS data are of size (51x51), whereas our networks' architectures could only work with some specific powers of two. Thus, we added a first zero-padding to (64x64) images layer at the entry of our network, and a cropping to (51x51) images layer before outputting the results of our denoising. We ran some tests to verify that this new network could achieve similar results than the previous ones.

Results

The results seem to be quite similar to the ones without the patch:

Model	SNR	Train RMSE	Test RMSE	e1 RMSE	e2 RMSE	R2 RMSE
U-Nets without patch	0-200	5.349e-05	6.086e-05	6.821e-03	4.510e-03	8.081e-03
U-Nets with patch	0-200	6.327e-05	7.214e-05	8.372e-03	3.855e-03	6.928e-03

Model	e1 error	e2 error	R2 error
U-Nets without patch	-5.588e-03	1.816e-03	4.804e-03
U-Nets with patch	7.522e-03	1.164e-03	1.168e-03

Visual results

Conclusion

Our quick patch doesn't seem to harm significantly the accuracy of our model!

[tobias-liaudat]

Thanks @aziz-ayed :)

This results are from the padding and cropping included in the network or done outside the network?

[aziz-ayed]

Hi @tobias-liaudat , they are done inside the network (first and last layer), but it's not the final version yet: this one isn't responsive to different input formats of images, it only works for 51x51 images

[tobias-liaudat]

@aziz-ayed cool, thanks :)

[tobias-liaudat]

You should push them to the repo and maybe edit the issue to add a reference to the new patch network.

[tobias-liaudat]

@aziz-ayed have you pushed this fix to the repo?

[aziz-ayed]

Yes! They are defined in the convert_and_pad and crop_and_convert functions that are applied in the op function of the class Learnlets here: Link