Closed Tiger1994 closed 6 years ago
Thanks for mentioning this. Actually I don't know why right now. It might be relevant to the data preprocessing, which I convert pictures in YCbCr
format and use Y
as the only channel. I will look into it later this week.
Thanks for your answer. I find the problem is caused by learning rate. I have tried a few of learning rates, the high learning rates (e.g. lr=0.01) can reconstruct color image, but the loss goes wild after dozens of epochs. And the low learning rates ( e.g. lr=0.0001) have the problem I mentioned before.
I am sorry to bother you, but I has a problem in using FSRCNN. The reconstructed SR images are tend to be gray image. If the FSRCNN can only reconstruct gray image, and why? Thank you very much.