融合深层语义与空间精确度的低照度图像增强方法  

A Low-Light Image Enhancement Method Combining Deep Semantic and Spatial-accuracy Features

作　　者：杨微 陈孝如 张志威 陈立军 YANG Wei;CHEN Xiaoru;ZHANG Zhiwei;CHEN Lijun(Department of Software Engineering,Software Engineering Institute of Guangzhou,Guangzhou 510990;School of Computer Science&Engineering,South China University of Technology,Guangzhou 510641)

机构地区：[1]广州软件学院软件工程系,广州510990 [2]华南理工大学计算机科学与工程学院,广州510641

出　　处：《计算机与数字工程》2023年第6期1276-1284,1322,共10页Computer & Digital Engineering

基　　金：国家自然科学基金项目(编号:61761049);2019年度广东省普通高校重点科研平台和科研项目(编号:2019KQNCX222);2021年度广东省普通高校特色创新(自然科学)项目(编号:2021KTSCX160,2021KTSCX161)资助。

摘　　要：全局深层上下文语义特征与空间精确度特征对深度学习低照度图像增强都非常重要,论文针对主流增强网络架构中深层上下文语义获取与保留空间精确度间的矛盾,提出融合深层语义及空间准确度的低光照图像增强方法。首先,不同深度的双编码结构并行提取全局深层上下文语义特征及空间精确度特征;其次,采用注意力及特征融合机制筛选聚合上下文语义特征与空间精确度特征;最后由单解码结构完成图像增强重建。实验结果,定性分析该方法能适应不同退化光照场景且增强结果纹理细节丰富、全局亮度提升自然、噪声伪影不明显、色彩真实;定量分析在PSNR、SSIM、RMSE、NIQE中较次优算法分别提升0.259、0.132、0.446、0.083。结果表明,论文方法有较好泛化性,结果综合质量较优,具有一定应用价值。Both global semantic context-deep and spatial-accuracy features are very important for deep learning low-light im-age enhancement.However,there is a contradiction between deep context semantic feature acquisition and spatial accuracy feature retention in the mainstream network architecture.Aiming at the contradiction,a novel low-light image enhancement method is de-signed,which integrates deep semantic and spatial-accuracy features.Firstly,this paper parallels dual coding structure with differ-ent depths,extracts respectively spatial accuracy and rich semantic characteristics.Secondly,the attention and feature fusion mech-anism is used to screen the semantic features and spatial accuracy features of aggregate context.Finally,image enhancement and re-construction are completed by single decoding structure.The result shows that with quantitative analysis,the proposed method can adapt to different degraded lighting scenes,and has the advantages of natural global brightness enhancement,indistinct noise arti-facts,rich and real color.With quantitative analysis and compared with the sub-optimal algorithm,the PSNR,SSIM,RMSE and NIQE indexes are improved by 0.259,0.132,0.446 and 0.083 respectively.The experimental results show that the method has good generalization and better comprehensive quality of the enhanced results,which has certain application value.

关 键 词：图像增强 上下文语义 注意力机制 特征融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]