摘要
针对铁路隧道复杂背景下细小裂缝存在图像特征难以提取的技术问题,提出一种基于改进残差网络(Residual Network,ResNet)的铁路隧道裂缝检测算法.该算法采用ResNet对裂缝进行检测,并在此基础上对网络进行了改进:首先将具有不同扩张率的空洞卷积块与传统的卷积块组合形成金字塔空洞卷积模块;其次将该模块放在ResNet的底部以增强其局部感受野,实现多尺度裂缝信息的提取;最后利用基于度量学习的组合损失函数来解决相似类别之间出现的裂缝漏检和误检的问题,进一步提升铁路隧道裂缝检测的效果.在公共裂缝数据集SDNET2018和自制裂缝数据集CRACK上,将其与文献方法、ResNet等其他方法进行了实验对比验证,结果表明:该算法与原始残差网络等方法相比,检测结果有明显的提升(精准率(R)为91.32%,召回率(P)为92.14%,F1值为91.73%),该方法也可为其他混凝土裂缝的检测提供参考.
In view of the technical problem that the image features of small cracks in the complex background of railway tunnels are difficult to extract,a crack detection algorithm for railway tunnels based on improved residual network(Residual Network,ResNet)is proposed.The algorithm uses ResNet to detect cracks and improve the network on this basis.First,the dilated convolution blocks with different expansion ratios are combined with the traditional convolution blocks to form a pyramid dilated convolution module.Secondly,the module is placed at the bottom of ResNet to enhance its local receptive field,which realizes the extraction of multi-scale crack information.Finally,the combined loss function based on metric learning is used to solve the problems of crack detection and false detection between similar categories,and further improve the effect of crack detection in railway tunnels.On the public crack data set SDNET2018 and the self-made crack data set crack,it is experimentally verified with the other three methods in the literature and ResNet.The results showed that the detection result of this algorithm was significantly improved compared with the original residual network method(precision rate 91.32%,recall rate 92.14%,F1 value 91.73%).This method can also provide a reference for the detection of other concrete cracks.
作者
常惠
饶志强
李益晨
赵玉林
CHANG Hui;RAO Zhi-qiang;LI Yi-chen;ZHAO Yu-lin(Beijing Key Laboratory of Information Service Engineering,Beijing Union University,Beijing 100101,China;College of Urban Rail Transit and Logistics,Beijing Union University,Beijing 100101,China)
出处
《东北师大学报(自然科学版)》
北大核心
2021年第3期56-63,共8页
Journal of Northeast Normal University(Natural Science Edition)
基金
国家级轨道交通通信与控制虚拟仿真实验室开放课题资助项目(2019RTCC04)
高铁安全大数据理论与关键技术研究项目(K2019Z006)
北京联合大学科研项目(ZK30202001).
关键词
铁路隧道
裂缝检测
金字塔空洞卷积
残差网络
度量学习
railway tunnel
crack detection
pyramid dilated convolution
residual network
metric learning