A Retinex unfolding network for low-light image enhancement

Low-light image enhancement is a fundamental problem in computational imaging, where insufficient illumination severely degrades visual quality and adversely affects downstream vision tasks. Existing approaches are typically dominated either by purely data-driven deep networks with limited interpretability or by handcrafted variational models with restricted representation capability. In this work, we propose an optimization-inspired unfolding framework for low-light image enhancement under the Retinex model. The proposed framework decomposes the enhancement process into illumination restoration and reflectance refinement, which are handled using two complementary mechanisms according to their distinct structural characteristics. Specifically, the illumination component is restored through an efficient element-wise transform-domain restoration module, while the reflectance component is recovered using an ISTA-inspired learned proximal operator for adaptive sparse feature refinement. By integrating model-based optimization with learnable feature restoration in a unified iterative architecture, the proposed framework achieves a favorable balance between interpretability and enhancement capability. Extensive experiments on paired low-light benchmark datasets demonstrate that the proposed method achieves competitive quantitative and visual enhancement performance.