Collaborative Dual-Branch Spatial-Frequency Enhancement Network for Low-Light Images
Tao He, Tiecheng Song, Yin Liu, Feng Yang, Ruiyuan Chen, Zhixin Li
Abstract
Low-light images are commonly present due to imaging factors such as insufficient light, night shooting and back lit. Existing low-light image enhancement (LLIE) methods typically rely on a low-light input image for enhancement, which seldom leverage information contained in its high-light counterpart to restore image structures and handle complex lighting, leading to unsatisfactory image quality. In view of this, in this paper we propose a Collaborative Dual-Branch Spatial-Frequency Enhancement Network (CDSE-Net). Specifically, we apply the inversion operation to low-light images to self-generate high-light images and build a collaborative dual-branch network which enhances images sequentially in spatial and frequency domains. In the spatial domain, we leverage adaptive curve estimation and multi-direction convolutions to restore lightness and structure information, respectively. In the frequency domain, we perform amplitude interactions on dual-branch images and 1×1 convolution on phase features to adjust image lightness and structures, respectively. Finally, we introduce an illumination-aware attention module to fuse two branches. Experiments on several widely used datasets quantitatively and qualitatively demonstrate the advantages of our network over state-of-the-art methods for LLIE.
BibTeX
@inproceedings{icassp2025_collaborativedua,
title = {Collaborative Dual-Branch Spatial-Frequency Enhancement Network for Low-Light Images},
author = {Tao He and Tiecheng Song and Yin Liu and Feng Yang and Ruiyuan Chen and Zhixin Li},
booktitle = {ICASSP 2025},
year = {2025}
}