Partially decoupled low-light image enhancement for brightening and denoising

The low-light enhancement problem is challenging due to the coupling of brightness and noise interference. Simple end-to-end models require the network to simultaneously possess strong denoising and brightness enhancement capabilities, making them relatively inefficient. The mainstream solution is Retinex-based decomposition, which separates the image into illumination and reflection components, with each component addressing brightness and noise issues, respectively, achieving full decoupling. However, this approach faces two key issues: firstly, the lack of ground truth labels for the illumination and reflection components makes the Retinex decomposition process difficult; secondly, after full decoupling, the reflection component’s corresponding network does not adjust brightness, limiting the network’s potential. To address these challenges, we propose a simple partially decoupled model for efficient low-light enhancement. Our model includes a straightforward brightness enhancement module to recover image brightness, followed by a denoising adjustment module for denoising and fine-tuning brightness. Experimental results demonstrate the efficiency of our approach.