Multimodal Low-Light Image Enhancement using Retinex-based Decomposition and Transformer Networks

Low-light image enhancement is a critical problem, impacting various applications such as surveillance and autonomous driving. Conventional methods often lead to overexposure or fail to preserve natural colors and fine details, producing overexposed or under enhanced results. In this work, we present a novel multimodal enhancement framework that integrates Retinex-based decomposition with Transformer networks. Retinex decomposition separates illumination and reflectance, allowing targeted corrections. A Transformer-based network models global dependencies, enhancing color and texture. An attention guided fusion mechanism integrates RGB and infrared (IR) modalities for robust enhancement. Experiments on standard datasets demonstrate superior performance in PSNR and SSIM, with visually improved brightness, contrast, and texture preservation.