RIRO: From Retinex-Inspired Reconstruction Optimization Model to Deep Low-Light Image Enhancement Unfolding Network

Low contrast, noise pollution and color distortion of low-light images tremendously affect human visual perception. The Retinex and its variant models are widely used for low-light image enhancement (LLIE). However, the performances of traditional Retinex algorithms are limited by intrinsic non-learnable characteristic. Recently, the latest LLIE methods directly unfold Retinex model as the popular networks such as URetinex-Net and RAUNA to resolve the black-box problem of conventional neural networks. Different from these methods focusing on the unfolding of image decomposition, we treat the classic LLIE as an image reconstruction task. Built upon Retinex theory, we propose a Retinex-Inspired Reconstruction Optimization (RIRO) model, which is unrolled as the RIRO network. This network consists of Low-light Decomposition and Enhancement Sub-Network (LDE Sub-Net) and Image Reconstruction Unrolling Sub-Network (IRU Sub-Net). The LDE Sub-Net is leveraged for the input initialization of the IRU Sub-Net. In RIRO model, we introduce a Dual-Domain Proximal (DDP) block to replace classic proximal operator, in which Fourier transform is utilized to transform spatial domain information into frequency domain information so as to simultaneously extract dual features on both spatial and frequency domains. Besides, we design a residual-aware weighted dual-fusion module and an adaptive weighted triple-fusion module to fuse different kinds of features. Numerous experiments on benchmark datasets have shown that the proposed method outperforms many advanced LIE methods.