DCD-Net: Decoupling-Centric Decomposition Network for Low-Light Image Enhancement

This paper presents a Decoupling-Centric Decomposition network for Low-Light Image Enhancement (DCD-Net). The DCDNet addresses two key challenges: (1) existing methods center on how to design the enhancement network and ignore the decomposition network's critical role to decouple reflectance and illuminations as the first step and (2) existing decomposition networks process images directly (or with pre-denoising), ignore their compliance with the Retinex theory. Specifically, by centering on illumination-reflectance decoupling refinement, DCDNet operates without reliance on supplementary enhancement networks. It consists of a preprocessing network and a decomposition network. The preprocessing network adopts a self-supervised learning mechanism to suppress Retinex-incompatible features in the input image, thereby improving the quality of Retinex decomposition. Within the decomposition network, the reflectance net is designed to suppress the contamination of illumination on reflectance restoration by the Dual-Gated Directional Reflectance Module (DGD-RM) and Reflectance-Guided Multi-head Self-Attention (RG-MSA), while the illumination net utilizes DCT to achieve local-global illumination estimates. Comprehensive experiments were conducted on two benchmark datasets (LOL and MIT) and five unpaired datasets. The quantitative results on different datasets are as follows (measured by PSNR and SSIM): LOL v1 (20.87, 0.770) and MIT (21.66, 0.864). The average NIQE across five unpaired datasets is 3.420. Both qualitative and quantitative analyses demonstrated the superiority of DCDNet over state-of-the-art methods. Moreover, ablation studies demonstrated the effectiveness of each module in DCDNet.