PNet:融合注意力机制的多级低照度图像增强网络  

PNet:multi-level low-illumination image enhancement network based on attention mechanism

作　　者：杨微 张志威 成海秀 Yang Wei;Zhang Zhiwei;Cheng Haixiu(Dept.of Software Engineering,Software Engineering Institute of Guangzhou,Guangzhou 510990,China;School of Computer Science&Engineering,South China University of Technology,Guangzhou 510641,China;Machine Learning&Data Mining Team,South China University of Technology,Guangzhou 510641,China;Guangdong Province Computer Network Key Laboratory,South China University of Technology,Guangzhou 510641,China)

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 [3]华南理工大学机器学习与数据挖掘团队,广州510641 [4]华南理工大学广东省计算机网络重点实验室,广州510641

出　　处：《计算机应用研究》2022年第5期1579-1585,共7页Application Research of Computers

基　　金：2018年度广东省普通高校重点科研平台和科研项目(2018KQNCX395,2018KQNCX394);2021年度广东省普通高校特色创新(自然科学)项目(2021KTSCX160,2021KTSCX161)。

摘　　要：低照度图像存在亮度低、噪声伪影、细节丢失、颜色失真等退化问题,使得低照度图像增强成为一个多目标增强任务。现有多数增强算法不能很好地在多个增强目标上取得综合的性能,对此,提出PNet——融合注意力机制的多级低照度图像增强网络模型,通过构建多级串联增强任务子网,结合注意力机制设计多通道信息融合模块进行有效特征筛选及记忆,网络以序列方式处理图像流,协同渐进式完成图像全局自适应亮度提升、噪声伪影抑制、细节恢复、颜色矫正等多任务。此外,通过与现有主流算法进行定量及定性分析对比,结果显示该方法能实现自适应图像亮度增强、细节对比度提升,增强后图像整体亮度自然,没有明显光晕及伪影且色彩较丰富真实,在PSNR、SSIM、RMSE指标中较次优算法分别提升0.229、0.112、0.335。实验结果表明,该方法在低照度图像增强的多目标任务上取得了综合较优秀的表现,具有一定的应用价值。Low-illumination image has degradation problems such as low brightness,noise artifact,detail loss and color distortion,which makes it a multiobjective task of the low-illumination image enhancement.As most existing enhancement algorithms fail to provide comprehensive performance in enhancing multiple targets,this paper proposed a model——PNet:multi-level low-illumination image enhancement network based on attention mechanism,which built a multi-stage tandem enhancement task subnet,and designed a multi-channel information fusion module for effective feature selection and memory with attention mechanism.With it,the network could process the image stream in a sequential manner,and collaboratively and incrementally completed multi-tasks such as the global brightness adaptive image enhancement,noise and artifact suppression,detail restoration,and color correction.In addition,through quantitative and qualitative comparison with existing mainstream algorithms,it showed that the proposed method could achieve brightness adaptive image enhancement and detail contrast enhancement.The enhanced image had an overall natural brightness with no obvious halo and artifacts on the one hand,and the color of which was rich and true on the other.Besides,compared with the sub-optimal algorithm,the index of the PSNR,SSIM and RMSE of the images processed by the proposed model was increased by 0.229,0.112,and 0.335.Experiment results show that the proposed method achieves excellent performance in multi-objective task of the low-illumination image enhancement,which has certain value in practice.

关 键 词：低照度图像增强 注意力机制 长短记忆 监督学习 多级子网 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]