Low-light image enhancement with luminance duality

Xingguo Lv, Xingbo Dong, Jiewen Yang, Lei Zhao, Bin Pu, Zhe Jin, Yudong Zhang

Abstract: Images captured under low-light conditions often suffer from high noise levels and a loss of details. Existing Low-light enhancement approaches often assume illumination as a global factor under Retinex theory, which may need to be more balanced with the complexities of real-world scenes with diverse lighting conditions. We propose a novel enhancement approach rooted in the Direct Perception (DP) theory. Through empirical evidence from real-world scenes, we illustrate the phenomenon of the duality of luminance in DP model, highlighting that luminance can exhibit both global and regional variations. Motivated by the above, we propose a novel low-light image enhancement framework, namely DPNet, that considers global and local luminance differences. Central to our approach is the introduction of two key modules: the Lumimator, a luminance estimator that leverages both local and global attention mechanisms, and the NLRestorer, a normal-light restoration network that effectively fuses color and luminance information for image restoration. Extensive experiments validate the efficacy of our framework, demonstrating significant enhancements in image quality metrics over state-of-the-art methods.

External IDs:doi:10.1016/j.knosys.2025.114420