Learning to Enhance Low-Light Images with Reliable Attention and Reinforced Distribution Alignment

Zhixin Cheng, Haodian Wang, Xiaotian Yin, Baoqun Yin, Tianzhu Zhang

Keywords: Attention Mechanism, Gaussian Mixture Model, Low-Light Image Enhancement

Abstract: Low-light image enhancement (LLIE) methods have recently adopted the HVI color space, which alleviates the entanglement between luminance and color and improves color fidelity through chrominance polarization and intensity compression. However, existing approaches may suffer from error accumulation during the interaction between luminance and chrominance components, and the lack of fine-grained modeling of color distribution can lead to unsatisfactory enhancement results. To address these challenges, we propose a novel low-light image enhancement framework, Learning to Enhance Low-Light Images with Reliable Attention and Reinforced Distribution Alignment. Specifically, we introduce two key modules: the Reliable Cross Attention (RCA) module, which aggregates luminance and chrominance features with reliable queries, and the Reinforced Distribution Alignment (RDA) module, which robustly fits the color distribution in a more fine-grained manner. These designs significantly improve the quality of enhanced images under low-light conditions. Extensive experiments on multiple benchmark datasets demonstrate that our method achieves state-of-the-art performance compared with existing approaches.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 1471