FCNet: Frequency Co-Interaction Network With Color Restoration for Lowlight Image Enhancement

Haoxiang Lu, Xiaowei Zhao, Zhenbing Liu, Peixian Zhuang, Zheng Liang, Weidong Zhang

Abstract: Enhancing low-light images is paramount in surveillance, photography, and other real-world applications. Unfortunately, numerous existing low-light image enhancement methods (LLIMs) tend to introduce color deviation and inherent noise. To address these challenges, we design an innovative frequency co-interaction dual-branch method, called FCNet, which consists of a content and details restoration branch as well as a color restoration branch. Specifically, a texture-guided multi-head self-attention mechanism is developed to achieve the interaction of high- and low-frequency features, and a spatial-dconv feed-forward module is used to aggregate the spatial correlation information of adjacent pixels. Additionally, the color restoration branch relying on the self-attention color strategy is introduced to restore the color in a pixel-by-pixel manner. Extensive experiments suggest our FCNet superiors several state-of-the-art approaches in terms of color restoration, light enhancement, and detail boosting. Furthermore, experiments further indicate that our method exhibits the potential benefits in object recognition and segmentation for low-illumination images.

External IDs:dblp:journals/tce/LuZLZLZ25