search

imatge-upc / 3D-GAN-superresolution

3D super-resolution using Generative Adversarial Networks
160 stars 46 forks source link

LR image #4

Open johnnylili opened 5 years ago

johnnylili commented 5 years ago

How do low-resolution images are simulated from high-resolution images? Is it using cubic interpolation?

vagomundo21 commented 4 years ago

It uses FFT and IFFT.

johnnylili commented 4 years ago

The super-resolution scale in this article is in three dimensions or two dimensions. Can the Fourier transform reduce the resolution simultaneously in three dimensions

vagomundo21 commented 4 years ago

Yes, it can reduce in 3 dimensions. I've implemented it.

johnnylili commented 4 years ago

Can you show me this part of the code? Many of the documents I have seen are only degrading the resolution by zeroing the outer part of the 3D k-space along two axes representing two MR phase encoding directions.

vagomundo21 commented 4 years ago

Here. Let me know if there is a better method.

johnnylili commented 4 years ago

I looked at your method and it seems to be truncated in both phase encoding directions. imgfft = imgfft [:, y_center-x: y_center + x, z_center-x: z_center + x] . This code should be used for K-space truncation. I want to know whether it can be truncated in all three dimensions.

johnnylili commented 4 years ago

zeroing the outer part of the 3D k-space along two axes representing two MR phase encoding directions, then inverse Fourier transform will be better. The resulting image will be the same size as the hr image