Decouple and reconstruct: An end-to-end enhancement framework for low-light and backlight images via DLS-SSNet

To address severe underexposure, loss of detail, noise amplification, and color distortion in images under complex lighting conditions, this paper proposes a deep learning model, DLS-SSNet, specifically for low-light and backlight image enhancement. Based on Retinex theory, this model employs an innovative decoupled architecture comprising decomposition, optimization, enhancement, and reconstruction to separate and target-process the reflectivity and illumination components. To further improve network performance, DLS-SSNet introduces an ECABAM channel-spatial attention mechanism incorporating high-frequency detail priors, a multi-scale detail processing module, and an adaptive saturation adjustment mechanism, effectively balancing global illumination correction and local texture enhancement. Furthermore, this paper designs a composite loss function (DLSLoss) containing seven core loss components to jointly constrain and optimize multiple dimensions of image detail, color, edge, and perceptual quality. Extensive comparative and ablation experiments on public datasets such as LOL-v1, LOL-v2, LSRW, and BAID demonstrate that DLS-SSNet performs excellently in both subjective and objective evaluation metrics, including PSNR, SSIM, and LPIPS. It solves the common problems of local artifacts and color shifts in existing methods and achieves image texture restoration, global brightness equalization, and extremely high visual perception quality under complex lighting conditions.