Kindling the Darkness: A Practical Low-light Image Enhancer paper reading notes

• This is a paper on supervised dark image enhancement from ACMMM2019, KinD
• Its network structure is shown in the figure below:
  
• First, a decomposition network decomposes the R and L components, then Restoration-Net and Adjustment-Net further process the R and L components respectively, and finally fuse the processed R and L components back. This is quite a routine process. There are some novel details. One is that the decomposition network uses the obtained R component to guide the extraction of the L component. One is a controllable brightness adjustment module

Decompose the network

• The loss function of the decomposition network is as follows: The first two losses are very common, namely the reconstruction loss and the constraint that dark and bright images have the same R. The third loss is the smoothing loss of the L component, but it is normalized with the gradient of the original image so that the edge area of ​​the dark image is preserved; the fourth loss is also the smoothing loss of the L component, here it is using a The curve smoothes out the middle part of the gradient value (noise)

• 

• 

• 

• $L_{is}^{LD}=$
  

• 

• 

• 

Reconstructing networks in R

• The loss function of restoration net is as follows, which is the restoration result of R of dark image and various distance measures of R of bright image:
  

Adjustment network of I

• The input to the illumination adjustment net is in addition to $L$ , and there is another one allα \alphaConcatenate$\alpha$$L$的feature map,$\alpha$ represents the adjustment factor, and the target$L$ divided by input$L$ goes to global average to get. The function of the network is to convert the input$L$ is adjusted to the LLof target$L.$ _ Compared with other retinex methods, this method directly targets$L\text{'}s$ method of gamma correction has better results.
• The loss function of this module is as follows. This loss must be calculated twice, once for $L$ as input, LLof bright image$L$ serves as target, once in reverse:

Experimental results

The visualization effect and NIQE of the method are very good, and the PSNR on LOL is also very high: