文献类型：期刊文献

英文题名：Low-light image enhancement via multi-stream vision state space module

作者：Liu, Shuai[1];Wang, Jinhua[1];Zhang, Sen[1];Yu, Pengcheng[1];Ma, Xiaoyue[1]

通讯作者：Wang, JH[1]

机构：[1]Beijing Union Univ, Smart City Coll, Beijing 100101, Peoples R China

第一机构：北京联合大学继续教育学院

通讯机构：[1]corresponding author), Beijing Union Univ, Smart City Coll, Beijing 100101, Peoples R China.|[1141733]北京联合大学继续教育学院;[11417]北京联合大学;

年份：2025

卷号：19

期号：3

外文期刊名：SIGNAL IMAGE AND VIDEO PROCESSING

收录：;EI(收录号：20250817915279);WOS:【SCI-EXPANDED(收录号:WOS:001408178300016)】；

基金：This work was supported by the Beijing Natural Science Foundation (No. 4252036), the National Natural Science Foundation of China (Nos. 62172045 and 62272049), and the Academic Research Projects of Beijing Union University (No. ZKZD202301).

语种：英文

外文关键词：Low-light image enhancement; Multi-stream state space model; Retinex theory

摘要：Low-light image enhancement (LLIE) aims to improve the brightness of images under low illumination while preserving details and structural information. Although methods based on Retinex theory and Transformer architectures have made significant progress, their dependence on large-scale training data and high computational costs limit practical applications. To address these challenges, researchers have proposed enhancement algorithms based on state-space models (SSM), which significantly reduce computational complexity while maintaining global modeling capabilities. However, existing methods still face issues such as color distortion and noise interference. To tackle these problems, we propose an innovative multi-stream SSM-based low-light image enhancement algorithm, MMamba-LLIE. This algorithm integrates three key modules: (1) a Color Correction Module (CCM) to effectively mitigate color distortion caused by Retinex theory; (2) a Multi-scale Feature Extraction Module (DFM) to capture both global and local structural information; and (3) a Noise Removal Module (UnoisyM) to suppress low-light noise. Experimental results demonstrate that MMamba-LLIE achieves significant improvements on the LOLv2-real dataset, with a 0.302 dB increase in PSNR, a 0.004 increase in SSIM, and a 0.65 reduction in RMSE. On the unparameterized DICM dataset, NIQE and PI are reduced by 0.026 and 0.051, respectively. Extensive experiments validate the superiority of the proposed method in both performance metrics and visual quality, providing a promising solution for low-light image enhancement. For details, please visit: https://github.com/lsaixuexi/MMamba-LLIE