基于光照不变正则约束的稳健性桥梁病害识别  

Illumination-invariance approach based on invariant regularization for bridge disease identification

作　　者：刘新龙 邓磊 杨建喜[1] 田丽萍 LIU Xinlong;DENG Lei;YANG Jianxi;TIAN Liping(School of Information Science and Engineering,Chongqing Jiaotong University,Chongqing 400074,China;China Communication Information Technology Group Co.,Ltd.,Beijing 100088,China)

机构地区：[1]重庆交通大学信息科学与工程学院,重庆400074 [2]中国交通信息科技集团有限公司,北京100088

出　　处：《铁道科学与工程学报》2023年第7期2728-2739,共12页Journal of Railway Science and Engineering

基　　金：重庆市教委科学技术研究资助项目(KJQN202100748,KJZD-M202000702);重庆交通大学校内科学基金资助项目(20JDKC-B038)。

摘　　要：桥梁表观病害识别是桥梁运营养护的关键技术之一。近年来,卷积神经网络(Convolutional Neural Networks,CNN)被广泛应用于桥梁表观病害识别。然而在光照不足条件下,卷积神经网络对于桥梁表观病害识别的稳健性通常不足。针对上述问题,提出一种基于光照不变正则约束的稳健性桥梁病害识别方法。该方法利用不变正则约束,同时约束CNN的特征提取模块和分类器模块,实现光照不足条件下的不变特征学习和分类器学习,进而增强CNN模型对于桥梁病害识别的光照稳健性。运用该方法基于桥梁表观病害数据集进行实验验证。实验结果表明:所提方法在光照不足条件下的桥梁病害的平均识别准确率高于对比方法;该方法在正常光照图像和光照变化图像上提取的特征矢量的欧氏距离为7.57,较对比方法提取的特征矢量具有更高的相似度。基于光照不变正则约束的稳健性桥梁病害识别方法在光照不足条件下提取的特征具有较强的不变特性,使得模型具有较强的病害识别稳健性,具有良好的工程应用价值,能够为桥梁的运营养护提供更准确的决策支持。Bridge surface disease identification is one of the key technologies of bridge operation and maintenance.In recent years,Convolutional Neural Networks(CNN)has been widely used in the identification of bridge surface disease.However,the robustness of CNN model for bridge surface disease identification is usually insufficient under the condition of insufficient illumination.To address this problem,an illuminationinvariance approach based on invariant regularization for robust bridge disease identification was proposed in this paper.In the proposed method,to achieve robust bridge disease identification to illumination,the invariant regularity constraint was used to constrain both the feature extraction module and classification module of CNN for invariant feature learning and classifier learning,and then enhance the illumination robustness of CNN model for bridge disease recognition,respectively.The proposed method was applied to the experimental verification based on the bridge surface diseases dataset.The results show that the average identification accuracy of the proposed method for bridge diseases under insufficient illumination is higher than that of the comparison method.The Euclidean distance of the feature vector extracted by this method on the normal illumination image and the illumination change image is 7.57,which has a higher similarity than the feature vector extracted by the contrast method.The features extracted by an illumination-invariance approach based on invariant regularization for robust bridge disease identification under the condition of insufficient illumination have strong invariance,which makes the proposed model have strong robustness of disease identification,has good engineering application value,and can provide more accurate decision support for the operation and maintenance of bridges.

关 键 词：桥梁表观病害识别 卷积神经网络 光照稳健性 正则约束 特征学习 

分 类 号：U446[建筑科学—桥梁与隧道工程] TP391.41[交通运输工程—道路与铁道工程]