SCAUNet: Step-Size-Consistent ADMM Unfolding Network for Low-Light Image Enhancement

Low-light image enhancement aims to restore visually pleasing normal-light images from degraded low-light observations. Most existing methods handle luminance variation from the enhancement perspective. As a result, the degradation process from a normal-light image to a low-light observation is usually not explicitly characterized. In addition, degradation-oriented optimization is often computationally expensive due to repeated iterative updates. To address these issues, based on the alternating direction method of multipliers (ADMM), a degradation-oriented step-size-consistent unfolding network SCAUNet is proposed. Specifically, a low-light image is modeled as the element-wise product of a normal-light image and a luminance degradation operator, together with additive noise. Based on this formulation, low-light enhancement is converted into the joint estimation of the target image and the degradation operator. Then, a state-based one-step ADMM solver is developed, and a step-size consistency constraint is introduced to improve the reliability of one-step unfolding. Extensive experiments on LOL-v1 and LOL-v2 demonstrate the effectiveness of the proposed SCAUNet. Compared with existing state-of-the-art methods, SCAUNet yields better enhancement quality, especially in preserving image structures, correcting illumination, and suppressing artifacts. Strong generalization ability is also verified on four no-reference low-light datasets, and promising results are obtained on single-image exposure correction.