文献类型：期刊文献

英文题名：Low-Light Image Enhancement using Retinex-based Network with Attention Mechanism

作者：Ma, Shaojin[1,2]; Pan, Weiguo[1,2]; Li, Nuoya[1,2]; Du, Songjie[1,2]; Liu, Hongzhe[1,2]; Xu, Bingxin[1,2]; Xu, Cheng[1,2]; Li, Xuewei[1,2]

第一作者：Ma, Shaojin

机构：[1] Beijing Key Laboratory of Information Service Engineering, Beijing Union University, China; [2] College of Robotics, Beijing Union University, Beijing Union University, China

第一机构：北京联合大学北京市信息服务工程重点实验室

年份：2024

卷号：15

期号：1

起止页码：489-497

外文期刊名：International Journal of Advanced Computer Science and Applications

收录：EI(收录号：20240815571764);Scopus(收录号：2-s2.0-85185007976)

语种：英文

外文关键词：Finite element method

摘要：Images in low-light conditions typically exhibit significant degradation such as low contrast, color shift, noise and artifacts, which diminish the accuracy of the recognition task in computer vision. To address these challenges, this paper proposes a low-light image enhancement method based on Retinex. Specifically, a decomposition network is designed to acquire high-quality light illumination and reflection maps, complemented by the incorporation of a comprehensive loss function. A denoising network was proposed to mitigate the noise in low-light images with the assistance of images’ spatial information. Notably, the extended convolution layer has been employed to replace the maximum pooling layer and the Basic-Residual-Modules (BRM) module from the decomposition network has integrates into the denoising network. To address challenges related to shadow blocks and halo artifacts, an enhancement module was proposed to be integration into the jump connections of U-Net. This enhancement module leverages the Feature-Extraction- Module (FEM) attention module, a sophisticated mechanism that improves the network’s capacity to learn meaningful features by integrating the image features in both channel dimensions and spatial attention mechanism to receive more detailed illumination information about the object and suppress other useless information. Based on the experiments conducted on public datasets LOL-V1 and LOL-V2, our method demonstrates noteworthy performance improvements. The enhanced results by our method achieve an average of 23.15, 0.88, 0.419 and 0.0040 on four evaluation metrics - PSNR, SSIM, NIQE and GMSD. Those results superior to the mainstream methods. ? (2024), (Science and Information Organization). All Rights Reserved.