Low-Light Image Enhancement With Semi-Decoupled Decomposition

• This is a paper on traditional dark image enhancement in the 2020 TMM journal, which proposes GTV-based brightness layer estimation.

• Initially, it was the following formula, using Gaussian filter kernel and Gaussian function to normalize the gradient:
  
  But this is too difficult to optimize, so it is further reduced to the following form to facilitate optimization:
  
  Thus, the previous fraction is set to $\omega$ , the following final formula can be obtained, which is the GTV loss formula used in retinex papers in the past two years, where S is the original image:
  

• The optimal value can be found iteratively:
  
  
  

• After finding I, then find R. You can use the same method to find a smooth R to achieve denoising:
  

• I won’t go into details about the optimization formula here. The experimental results look pretty good according to NIQE and the pictures.
  
  
  It can also remove fog:
  

• The inspiration is to first estimate the smooth I, and then estimate the smooth R based on the smooth I and the original image, which can achieve denoising to a certain extent.