Closed huanzheng551803 closed 2 years ago
why adding noise each iteration on Guassian case? I think it should be only adding once and training with them.
The same noise level generates non-deterministic noise, therefore it must be regenerated every time it is created.
then the problem become supervised training. Since we can get the same image with different noise.
Also in real case, we can not access the same image with different noise
Maybe you have some mistakes in understanding the code.
its not about the code,but about the problem setting
In unsupervised denoising, unsupervised means we have no clean image, only noisy image itself. But if you add noise every iteration, it means we can access the clean image by averaging them, this does not meet the problem setting
Since the loss function does not include clean images and only uses noisy images, our method is fully self-supervised.
i admit your method is fully self-supervised. But i think the experiment should mimic the real case, where we only have one noisy image for each image
No noise is added to SIDD or FMDD images; instead, we train directly on raw noisy images.
why adding noise each iteration on Guassian case? I think it should be only adding once and training with them.