Modified Diff-Retinex Model for Robust Low-Light Image Enhancement

Abstract

Improving images captured under low light is a common topic in computer vision. Applications range from surveillance, self-driving cars, and photography. However, classical Retinex-based algorithms suffer from over-enhancement, noise amplification, and color distortion. This is because they suppose that the resultant image can be separated into two components: illumination and reflectance. To overcome these issues, we propose a Modified Diff-Retinex Model, which combines the Retinex theory with a diffusion-based regularization term to achieve strong and visually pleasant image enhancement. We also use an adaptive edge-aware diffusion to estimate illumination, preserving structural details while eliminating noise. Unlike conventional Retinex, which assumes fixed priors or logarithmic transformations of the image, our model adapts to local changes in lighting and contrast in real time. We restore the original colors and brightness of the image via a post-enhancement refinement stage that reduces artifacts and improves visual quality. Experiments on publicly available low-light image datasets show that our model outperforms the state-of-the-art models and works better in terms of visual quality and quantitative metrics such as contrast, color, and noise reduction. The proposed method achieves an average PSNR of 16.2890 dB, SSIM of 0.6344, and LPIPS of 0.2628, demonstrating superior performance compared to existing low-light enhancement methods.