DT-Retinex: low-light enhancement network based on diffuse denoising and light enhancement

Abstract

Low-light images often suffer from insufficient brightness, blurred details, and noise interference, which degrade visual quality and reduce the accuracy of computer vision tasks. To address these challenges, this paper proposes a low-light image enhancement model named DT-Retinex. The method improves image quality through three stages: image decomposition, reflectance denoising, and illumination enhancement. First, the decomposition network decouples the input image into reflectance and illumination components while preserving structural features. Then, a diffusion model is introduced to progressively denoise the reflectance component, with a customized denoising loss designed to enhance detail restoration. Finally, DT-Retinex adopts an encoder-decoder architecture for illumination enhancement: the encoder extracts multi-level features and leverages the LIT module to model global illumination, while the decoder incorporates CBAM attention to emphasize key regions and adaptively adjust lighting information during spatial reconstruction. Experimental results show that DT-Retinex outperforms existing methods on several benchmark datasets, achieving excellent performance on PSNR, SSIM, and LPIPS, as well as better perceptual naturalness and consistency under no-reference metrics such as NIQE and BRISQUE. Overall, DT-Retinex provides a robust and high-quality solution for low-light image enhancement tasks.