摘要
列车高速运行易导致列车表面部件出现机械损伤,影响列车的运行安全。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,且因列车运动、拍摄设备的震动等带来的拍摄图片不同程度的模糊,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,故提出一种基于轻量级注意力生成对抗网络的TEDS图像盲去模糊算法。第1步,采用改进的带通道注意力和空间注意力机制的线性倒残差瓶颈模块构建轻量级特征提取网络,将其提取的5种尺度的特征送入特征金字塔网络(FPN)构建生成器,使生成器能有效地关注重点信息、综合底层和高层信息、从多尺度提取特征;第2步,采用谱归一化U_Net作为判别器网络,对局部信息产生更精确的梯度反馈,并在局部,全局双判别器的基础上加入逐像素判别,增强对图像纹理和细节上的对抗学习。研究表明,算法处理后TEDS图片较其他算法对不同尺度的目标均有更好的去模糊效果,图像有更高的清晰度;评价指标PSNR和SSIM达到28.6和91.2%,较其他算法分别提升了0.7和3.8个百分点;轻量级网络参数文件只有13.6 M,与其他算法相比,其速度取得几十倍的提升,在不使用GPU的情况下每分钟可对75张TEDS图片进行去模糊处理,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,提高损伤检测和标注的精准度,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.Moreover,due to train movement and vibration of shooting equipment,the captured images may be blurred to varying degrees,which can interfere with the analyzing and labeling of faults by staff and affect the real-time and accuracy of detection.Therefore,a blind deblurring algorithm for TEDS images based on lightweight attention generative adversarial networks was proposed.Firstly,an improved linear inverse residual bottleneck module with channel attention and spatial attention mechanisms was used to construct a lightweight feature extraction network.Subsequently,the extracted features of five scales were fed into the Feature Pyramid Network(FPN)to construct a generator,enabling the network to effectively focus on key information,integrate low-level and high-level information,and extract features from multiple scales.Secondly,spectral normalization U_Net was used as the discriminator network to generate more accurate gradient feedback for local information.On the basis of local and global dual discriminators,pixel by pixel discrimination was added to enhance adversarial learning on image texture and details.Research results are shown as follows.The TEDS images processed by this algorithm have higher clarity compared to other algorithms,with evaluation indicators PSNR and SSIM reaching 28.6 and 91.2%,which are improved by 0.7 and 3.8%compared to other algorithms.The lightweight network parameter file is only 13.6 M,which can achieve several tens of times the speed improvement of other algorithms,and can deblurring 75 TEDS images per minute without GPU,meeting the real-time requirements of the TEDS system.The results can effectively improve the image quality of the TEDS system,improve the accuracy of damage detection and labelin
作者
王登飞
苏宏升
陈光武
吕晓聪
赵小娟
WANG Dengfei;SU Hongsheng;CHEN Guangwu;LÜXiaocong;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;China Railway Shanghai Group Co.,Ltd.,Nanjing Dongcheduan,Nanjing 210012 China)
出处
《铁道科学与工程学报》
EI
CAS
CSCD
北大核心
2024年第9期3797-3808,共12页
Journal of Railway Science and Engineering
基金
国家自然科学基金资助项目(61867003)
甘肃省科技计划资助项目(21ZD4WA018,23JRRA1693)。
关键词
动车组运行故障图像检测系统
盲去模糊
注意力机制
生成对抗网络
MobileNet
trouble of moving electric multiple units detection system(TEDS)
blind deblurring
attention mechanism
generative adversarial network
MobileNet
