Details about the use of AFM

[Zeudfish]

My job is to have different levels of noise images, but when I tried to use your method, I encountered two problems

1. When I used FFT according to Eq. 4 in the paper, the image could not be correctly restored, resulting in strip structure and color bias. I used the fft module of torch, is there anything I need to pay extra attention to?
2. When I use the idea of Mixup instead of FFT and proceed according to Eq. 9, I encounter a problem. After removing AFM, my model cannot properly de-noise hard noise, but must run the network twice and use mixup operation to de-noise correctly.

[dhryougit]

1. You're right, and when creating the hard image according to Eq. 4, slight distortions do occur. We utilized these distortions as a form of augmentation to make the denoising training more robust. However, if the distortions are too severe, it could indicate an implementation error.

2. Our method is designed to remove both the noise generated by mixup and the original noisy images. You might need to check if only the images generated by mixup were used in training, without training on the original noisy images.