-
This is a paper on supervised dark image enhancement published in the Journal of Area 2 of the Chinese Academy of Sciences in 2022.
-
The network structure is shown in the figure below:
The ARM module is shown in the figure below:
The CAB module is shown in the figure below:
The LKA module actually puts in some large convolution kernels:
The AFB module is shown in the figure below:
-
There is nothing special about these network structures. It is meaningless to make them so complicated one after another. The final predicted result is the sum of the two branches:
-
The loss function is the weighted sum of the smooth L1 loss of the enhancement result and GT and the smooth L1 loss of the inverse image enhancement result and GT inverse image:
-
Judging from the experimental results, the color tone is better, and the PSNR and SSIM are quite high:
-
Summary: It’s really not interesting to build a network structure. There is also some work being done on the dual-branch front and back graphs. The only thing is that this smooth L1 loss seems to be used in several tasks. You can try it in the future.