This code is the unofficial implementation of this work. Specifically, I coded the pipeline of the whole degradation model according to the configuration given in the paper. Note that some details are different from the original paper (For the sake of simplicity, I've removed the processed camera sensor noise, but this part may be key to the generalization of the network). I chose HiFaceGAN as the model for restoring the degraded images. I used the code degrade.py to generate 10000 degraded face images in advance. After training with two V100 on hifacegan for one day, the reconstruction effect on the real images I found is much better than the general bicubic degradation method.
I added the code of the processed camera sensor noise, thanks for this work, but I have not tested the effectiveness of this method. Note that my implementation of this method is slightly different from the BSRGAN author's paper.
The degraded images below are all real images from social networks.
Images layout is LR | mixed degradation SR
Images layout is LR | bicubic degradation SR | mixed degradation SR
Compared with the original paper, the size of current training set of my method is relatively small and the training time is relatively short, though it has achieved much better results than the bicubic degradation method. In some real images, there are still some noises that can't be removed, which may be because I didn't use the processed camera sensor noise. In some test images, it is also found that the reconstruction effect of the model for the face area is acceptable, but the reconstruction effect of the background and text is relatively poor, which may be caused by the lack of diversity of the data set itself.