基于复合图双卷积神经网络的路面裂缝识别方法  

Pavement Crack Detection Method Based on Composite Image Double CNN Network

作　　者：王浩仰[1,2,3,4] 潘宗俊 曹建坤 张洁[2,3,4] 郭宝栋 WANG Hao-yang;PAN Zong-jun;CAO Jian-kun;ZHANG Jie;GUO Bao-dong(School of Transportation,Southeast University,Nanjing,Jiangsu 211189,China;RoadMainT Co.,Ltd.,Beijing 100095,China;Research Institute of Highway,Ministry of Transport,Beijing 100088,China;National Engineering Research Center for Efficient Maintenance,Safety and Durability of Highways and Bridges,Beijing 100088,China;School of Cyberspace Security,Beijing University of Posts and Telecommunications,Beijing 100876,China)

机构地区：[1]东南大学交通学院,江苏南京211189 [2]中公高科养护科技股份有限公司,北京100095 [3]交通运输部公路科学研究院,北京100088 [4]公路与桥梁高效养护及安全耐久国家工程研究中心,北京100088 [5]北京邮电大学网络空间安全学院,北京100876

出　　处：《公路交通科技》2024年第9期1-9,共9页Journal of Highway and Transportation Research and Development

基　　金：中路高科交通科技集团有限公司交通强国试点项目(JTQG2022-1-3-1)。

摘　　要：为了建立一种基于深度学习卷积神经网络的路面检测模型,提高特殊路面裂缝,如白裂缝、浅裂缝、潮湿裂缝、修补开裂等的识别准确率,在单卷积神经网络结构(单网络)上,提出了基于复合图双卷积神经网络的路面裂缝识别方法。首先,该方法在输入灰度图基础上,考虑裂缝病害图像特征,增加对应二值图组成复合图通道;其次,在单网络结构基础上增加一个针对特殊裂缝识别的单网络,非特殊裂缝网络训练使用全部数据,特殊裂缝网络训练使用特殊裂缝数据,两个网络参数分别独立更新,从而形成复合图双网络结构;然后两个网络分别对同一测试数据进行判定,得出各自的概率矩阵,最后再根据概率单侧抑制的原理将两个单网络输出结果进行叠加,得出最终识别结果。组织了70万张检测车采集图片对复合图双网络方法进行训练和测试。结果表明,复合图双网络识别重叠度、精确度、召回率显著优于灰度图单网络,证明了提出的两处优化,即将单通道灰度图改造为双通道复合输入图和增加一个特殊裂缝识别网络,提升了非特殊裂缝与特殊裂缝区域识别能力。此外,复合图双网络的重叠度、精确度、召回率指标比其他深度学习路面裂缝识别算法方法高。To establish the pavement crack detection model based on deep learning convolutional neural network(CNN)and to improve the identification accuracy of special pavement cracks(e.g.,white crack,shallow crack,wet crack,and repair crack),based on the single CNN structure(single network),the pavement crack detection method based on composite image double CNN network was proposed.First,based on the input gray image,the corresponding binary image was added to form the composite image channel considering the crack disease image characteristics.Second,on the basis of single network structure of CNN,the single network for special crack detection was added,so as to form the composite image double network structure.The non-special crack network training used all data;the special crack network training used special crack data;and the two network parameters were updated independently;thus the composite image double CNN network was formed.Then,two networks judged the same test data respectively to obtain their own probability matrix.Finally,according to the principle of probabilistic one-sided suppression,the output results of two single networks were superposed to get the final recognition results.There were 700000 images collected by detection vehicles for training and detecting the composite image double network method.The result indicates that the intersection over union,precision and recall of the composite image double network are significantly better than those of gray image single network.Two proposed optimizations(i.e.,transforming the single-channel grayscale map into a two-channel composite input map,and adding a special crack identification network)are proved to improve the detection ability for non-special crack and special crack areas.In addition,the intersection over union,precision and recall of the composite image double network are higher than those of other deep learning pavement crack detection algorithms.

关 键 词：智能交通 裂缝识别方法 复合图双卷积神经网络 路面裂缝 二值图 

分 类 号：U418.1[交通运输工程—道路与铁道工程]