Low-light Image Enhancement Based on Retinex

Noise and underexposure in low-light images present challenges to image enhancement, often resulting in distortion in enhanced images. This paper proposes an improved R2RNet enhancement algorithm tailored for low-light images. The denoising module of the original network has been removed, and four Attention Feature Modules (AFM) have been adopted in the decomposition network for multi-scale feature extraction and denoising, thereby producing high-dimensional illumination and reflection feature maps. These feature maps are then utilized in the enhancement network to enhance brightness and contrast, generating the final enhanced image. Experimental results demonstrate that the proposed algorithm excels in reducing noise, maintaining high texture and detail, and effectively restoring realistic scene colors. Compared to various other enhancement algorithms, objective metrics such as PSNR and SSIM show improvements.