This is a weakly supervised dark image enhancement paper of CVPR2023. It requires a data set with two dark images of different brightnesses and the same content for the same scene. However, the paper proposes that a dark image can be subjected to a sampling operation similar to neighbor2neighbor to obtain two images. get.
The network structure is shown in the figure below. It consists of 3 modules. P-net is responsible for denoising and removing artifacts on the image. L-Net and R-Net respectively correspond to the L component and R component estimation in the retinex model. g(L) is gamma correction.
It can be seen that there are three losses in the training process. One is the reconstruction loss of P-Net, which is actually just L2 Loss on the input and output (the reason why P-Net can work also needs to rely on the supervision of other losses), and then is the reconstruction loss of retinex (L*R=I), and then the R component generated by the two dark images must be the same loss, as follows:
The reconstruction loss of retinex is quite special and consists of 4 items. The first, third and fourth items are more conventional and are the common retinex prior losses. The second term is complementary, where stopgrad means that the gradient is not propagated into L.
The experimental results are not bad
Summary: I think the idea of the article is a bit strange. Since I can get two images of the same scene, why not one dark and one bright. However, it would be okay if we use neighbor2neighbor sampling to obtain two samples for training, but there are no experimental results of training in this way. In fact, the innovation points are average. The only one is that P-net uses retinex loss to denoise images, which is very interesting. The others are common losses of the retinex method.
