Retinex With Progressive Decomposition and Multi-Scale Denoising for Low Light Image Enhancement

Previous approaches based on Retinex Theory enhance low light images by carefully decomposing an input into reflectance and illumination. However, due to unnatural lighting, color distortion, and noise in the dark regions, the visibility of recovered images is drastically reduced. If the ill-posed separation of reflectance and illumination cannot be properly addressed, the application of the Retinex-based framework will be significantly limited. To settle this issue, we propose a progressive deep network that employs a deep unfolding scheme with step-wise cross-compensation in the decomposition stage to disentangle the illumination and reflectance information from the original image. Then, following the divide-and-conquer principle, we design a multi-scale denoising network that hierarchically aggregates local and global features to eliminate the noise remaining in the dark regions of the reflectance map. Additionally, we design an adjustment network to dynamically modify the pixel values of the illumination for further improving the perceptibility of the regions with inadequate lightness. We comprehensively study the inherent properties of each proposed module. Both qualitative and quantitative experiments indicate that our framework is superior to state-of-the-art (SOTA) methods on a wide range of real and synthetic low light datasets, especially for images with non-uniform lightness.