This is an article on self-supervised dark image enhancement posted on arxiv in 2020. Many subsequent unsupervised dark image enhancement papers will quote and compare it with it.
The article analyzes the formula of retinex from the perspective of Bayesian probability as follows: Therefore, the maximization of the three probabilities on the right side of the equation can be guided by three loss functions respectively: The first term is defined in the article as L1 loss, which is very common: The second term It is defined as the brightness component of the dark image (V in the HSV color space) histogram equalization result F(S) and the L1 distance of the brightness component of R and the smoothing loss of R: The third term is defined as follows: The network structure is as follows As shown in the figure:
The article proposes that a network that downsamples first and then upsamples can denoise, but it will make the picture blurry.
Training with 485 real images of LOL, the patch size is 48*48
The article also mentioned that in fact, different training times cannot achieve the same results.
At the same time, the article also proposed that training can only use one picture. On the 15 test sets of LOL, select one as the training set, train for 10,000 epochs, and then test on these 15 images, the effect is still very good. Moreover, when training with a single image, as the number of iterations increases, no artifacts appear in the results produced by the network, indicating that the artifacts are not caused by the network design (so the finger points to stochastic gradient descent and the data set?): The visuals look good too:
