基于聚焦线性注意力Retinexformer的TEDS图像实时暗光增强方法研究  

Real time low light enhancement method of TEDS images based on focused linear attention Retinexformer

作　　者：王登飞 苏宏升[1] 陈光武 陈登科 赵小娟 WANG Dengfei;SU Hongsheng;CHEN Guangwu;CHEN Dengke;ZHAO Xiaojuan(School of Automation and Electrical Engineering,Lanzhou Jiaotong University,Lanzhou 730070 China;Key Laboratory of Plateau Traffic Information Engineering and Control of Gansu Province,Lanzhou 730070 China;Baotou Vehicle Depot,China Railway Hohhot Group Co.,Ltd.,Hohhot 010000,China)

机构地区：[1]兰州交通大学自动化与电气工程学院,甘肃兰州730070 [2]甘肃省高原交通信息工程及控制重点实验室,甘肃兰州730070 [3]中国铁路呼和浩特局集团公司包头车辆段,内蒙古呼和浩特010000

出　　处：《铁道科学与工程学报》2024年第11期4840-4850,共11页Journal of Railway Science and Engineering

基　　金：国铁集团科技计划项目(N2023G064);甘肃省科技计划资助项目(21ZD4WA018,23JRRA1693)。

摘　　要：列车高速运行下,表面部件易产生机械损伤,影响列车的安全运行。用于损伤检测的动车组运行故障图像检测系统(TEDS)需进行检测的部件形态多样、体积大小不一,且因对列车底部、夜晚进行图像采集时的暗光环境导致图像大部分区域偏暗,对比度低,给工作人员对故障的分析和标注带来干扰,影响检测的实时性和准确率,提出一种基于线性聚焦注意力的Retinexformer(RetinexFLAformer)网络对TEDS图像进行暗光增强。首先分析Retinexformer中进行自注意力计算的相似矩阵存在低秩的问题,采用线性聚焦注意力对网络进行改进,在保证计算复杂度不变的情况下,提高相似矩阵的秩以增加网络的特征多样性;其次增加空间一致性损失、曝光控制损失和颜色恒定损失,来抑制由于曝光不均引起的局部区域对比度下降和颜色畸变;最后在以上改进的基础上进一步调整网络结构构建FastRetinexFLAformer,以达到更快的暗光图像处理速度。研究表明,改进后的RetinexFLAformer能有效提高TEDS图片的暗光增强效果,和其他算法对比,评价指标PSNR和SSIM分别提高0.55和0.023;FastRetinexFLAformer网络参数文件只有3.34 M,可达到当前主流方法相当的处理效果,且能有效提升暗光增强速度,达到TEDS系统的实时性需求。研究成果可有效提高TEDS系统的图片质量,提高损伤识别和标注的精准度,提升工作人员的效率,更好地保障铁路的安全运行。High speed operation of trains can easily cause mechanical damage to surface components,affecting the safety of train operation.The trouble of moving electric multiple units detection system(TEDS)used for damage detection needs to detect components with diverse shapes and various volume sizes.And due to the dark light environment during image acquisition,most areas of the image are dark and the contrast is low,which interferes with the analysis and annotation of faults by the staff,affecting the real-time and accuracy of detection.In this paper,a Retinexformer network based on linear focused attention(RetinexFLAformer)was proposed for dark light enhancement of TEDS images.Firstly,this research analyzed the low rank issue of the similarity matrix used for self-attention calculation in Retinexformer and adopts linear focused attention to improve the network,so as to increase the rank of the similarity matrix and the feature diversity of the network while ensuring the same computational complexity.Secondly,add spatial consistency loss,exposure control loss,and color constant loss to suppress local contrast reduction and color distortion caused by uneven exposure.Finally,based on the above improvements,FastRetinexFLAformer was contrusted by further improving the network structure to achieve faster dark light processing speed.Research has shown that the improved RetinexFLAformer can effectively improve the dimming effect of TEDS images.Compared with other algorithms,the evaluation metrics PSNR and SSIM have increased by 0.55 and 0.023 respectively.The FastRetinexFLAformer network parameter file is only 3.34 M,which can achieve the same processing effect as current mainstream methods and effectively improve the speed of dark light enhancement,meeting the real-time requirements of TEDS systems.The research results can effectively improve the image quality of the TEDS system,enhance the accuracy of damage diagnosis and labeling,improve the efficiency of staff,and better ensure the safe operation of railways.

关 键 词：动车组运行故障图像检测系统 暗光增强 Retinexformer 线性聚焦多头自注意力 空间一致性损失 

分 类 号：U279.3[机械工程—车辆工程]