暗区特征引导的多分支低照度图像增强网络  

Dark-region Features Guided Multi-branch Network for Low-light Image Enhancement

作　　者：杜晓刚 曾杰鹏 雷涛[1,2] 张学军[3] 王营博 DU Xiaogang;ZENG Jiepeng;LEI Tao;ZHANG Xuejun;WANG Yingbo(Shaanxi Joint Laboratory of Artificial Intelligence,Shaanxi University of Science and Technology,Xi′an 710021,China;The School of Electronic Information and Artificial Intelligence,Shaanxi University of Science and Technology,Xi′an 710021,China;The School of Electronic and Information Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China)

机构地区：[1]陕西科技大学人工智能联合实验室,西安710021 [2]陕西科技大学电子信息与人工智能学院,西安710021 [3]兰州交通大学电子与信息工程学院,兰州730070

出　　处：《小型微型计算机系统》2024年第10期2465-2472,共8页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61861024,62271296,62201334)资助;陕西省教育厅科学研究计划项目(23JP022,23JP014)资助;陕西省重点研发计划项目(2021ZDLGY08-07)资助.

摘　　要：针对低照度图像增强时存在的局部增强过度或不足以及细节丢失的问题,提出了基于暗区特征引导的多分支低照度图像增强网络.该网络有3个优势:首先,引入暗区特征提取子网络,提取低照度图像的亮度分布作为后续输入,以提升对局部暗区的曝光处理能力;其次,设计了基于逐层特征校正的分解子网络,通过使用坐标与通道注意力来精准定位和抑制噪声,并进行多阶段信息融合来输出具有丰富细节的低噪声反射分量;最后,设计了基于多分支特征补全的增强子网络,通过提取纹理和色差特征来补充细节,并使用暗区特征引导机制强调图像的亮度分布,从而增强模型的曝光处理能力和细节保留能力.在主流的公开数据集上进行实验,结果表明:与流行的低照度图像增强网络相比,该网络在主客观评价上均取得了更好的增强效果.Aiming at the problems of local over-enhancement or under-enhancement and detail loss in low-light image enhancement,a dark-region features guided multi-branch network for low-light image enhancement is proposed in this paper.The network has three advantages:firstly,a dark-region feature extraction sub-network is introduced to extract the brightness distribution information of low-light images as the input for subsequent processing,thereby improving exposure processing ability of the model for local dark regions.Secondly,a decomposition sub-network based on layer-wise feature correction is designed to accurately locate and suppress noise using coordinate and channel attention mechanisms.It performs multi-stage information fusion to output a low-noise reflection component with rich detail information.Finally,an enhancement sub-network based on multi-branch feature completion is designed to supplement details by extracting texture and color difference features,while using the dark-region feature guided mechanism to emphasize the brightness distribution of the image,thereby enhancing the exposure handling capability and detail preservation ability of the model.Experimental results conducted on mainstream public datasets demonstrate that compared to popular low-light image enhancement method,the proposed model achieves better enhancement results in both subjective and objective evaluations.

关 键 词：低照度图像 深度学习 图像增强 注意力机制 RETINEX理论 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]