Two-Stage Denoising Diffusion Model for Low-Light Image Enhancement

Images captured under weak illumination typically suffer from low brightness and contrast, severe color distortion, and significant noise contamination, which not only degrade human visual perception but also hinder the performance of high-level vision tasks. Low-light image enhancement aims to improve visual quality and provide favorable conditions for subsequent image processing. To address the challenges of non-uniform illumination and loss of details in dark regions, we propose a two-stage denoising diffusion model (two-stage DDM). Specifically, we design a convolution-based Retinex decomposition module to achieve fast and robust image decomposition, followed by a two-stage diffusion-based denoising process that further enhances global image details, brightness, and contrast. In addition, we introduce a feature enhancement module to strengthen the representational capacity of the reflectance component. To evaluate the robustness and generalization ability of the proposed model, extensive experiments are conducted on the LOLv1, LOLv2-real, and LSRW datasets. Experimental results demonstrate that the proposed two-stage DDM achieves competitive performance with state-of-the-art methods, producing enhanced images with more natural and spatially uniform brightness, along with noticeably improved visual quality.