基于边缘增强感知的混凝土裂缝病害检测方法  

Crack damage detection method based on edge feature reinforcement learning

作　　者：谭兆 王保宪[2] 秦守鹏 赵维刚[2] TAN Zhao;WANG Baoxian;QIN Shoupeng;ZHAO Weigang(China Railway Design Corporation,Tianjin 300251,China;School of Safety Engineering and Emergency Management,Shijiazhuang Tiedao University,Shijiazhuang 050043,China)

机构地区：[1]中国铁路设计集团有限公司,天津300251 [2]石家庄铁道大学安全工程与应急管理学院,河北石家庄050043

出　　处：《铁道科学与工程学报》2023年第8期3172-3180,共9页Journal of Railway Science and Engineering

基　　金：城市轨道交通数字化建设与测评技术国家工程实验室开放课题基金资助项目(2021ZH01);国铁集团实验室基础研究项目(L2021G013);国家自然科学基金资助项目(52178293,51808358,U2034207);河北省高等学校科学技术研究项目(BJ2020057);河北省自然科学基金创新研究群体项目(E2021210099)。

摘　　要：现有基于深度学习的混凝土表面裂缝病害检测与识别模型未关注裂缝目标边缘关键特征学习,易导致在复杂背景下细微裂缝漏检以及裂缝病害宽度感知精度不高。针对此问题,提出将裂缝边缘作为目标关键特征进行感知学习,通过构建一种新的裂缝边缘预测分支网络,并与裂缝主干分割网络联合训练,形成一种基于双任务学习的裂缝检测与识别模型。利用Sobel边缘检测算子提取裂缝真值图中的边缘,作为裂缝边缘预测感知标记。通过通道卷积操作将原始裂缝识别网络的解码层划分为2条独立支路,一条支路用于完成裂缝主干分割任务,另一条支路则利用U-Net++网络解码器具有的聚合上下文功能,建立边缘预测分支网络。整个框架以正则化约束方式,将裂缝主干分割网络与裂缝边缘预测分支网络的损失函数联合训练,由此提高算法对裂缝病害的综合感知能力。通过选取一定数量的混凝土表面细微裂缝图像进行算法测试,实验结果表明本文算法在引入裂缝边缘预测增强感知模块后,在细微裂缝识别方面,取得了比FCN,U-Net及U-Net++网络更优的检测与识别效果。The existing concrete crack detection and recognition models based on deep learning pay little attention to the key feature learning of crack target edges,which leads to the missing detection of fine cracks in complex backgrounds and low detecting accuracy of crack disease width.To solve this problem,this paper proposed to take the crack edge as the target key feature for perceptual learning.By constructing a novel crack edge prediction branch network and co-training it with the crack backbone segmentation network,a crack detection and recognition framework based on two-task learning was established.Specifically,Sobel edge detection operator was first used to extract the edges of crack ground-truth data,and it was utilized as the manual labels of crack edge predicting network.Then,the decoder of original crack recognition network was divided into two branches by channel convolution operation:one branch was used for segmenting the backbone of cracks,and the other one utilized the aggregated context function of U-Net++decoder network for building the crack edge predicting branch.In the whole framework,the loss functions of backbone network for crack region segmentation and edge prediction branch network were jointly trained in the form of regularization constraints,thereby improving the comprehensive perception accuracy of crack damage detector.Finally,the comparative experiments were implemented using a certain number of concrete fine crack images.The experimental results demonstrated that by adding the edge-prediction enhanced perception module,the proposed crack detector achieved better crack detecting results than FCN,U-Net and U-Net++networks in terms of detecting the fine cracks.

关 键 词：裂缝检测 U-Net++网络 边缘检测 双任务学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]