基于迁移学习的桥梁表观病害检测技术研究  

A bridge surface distress detection technology based on transfer learning

作　　者：王桂平 陈旺桥 杨建喜[1] 唐于凌 吴波 WANG Guiping;CHEN Wangqiao;YANG Jianxi;TANG Yuling;WU Bo(School of Civil Engineering,Chongqing Jiaotong University,Chongqing 400074,China;School of Information Science and Engineering,Chongqing Jiaotong University,Chongqing 400074,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]重庆交通大学土木工程学院,重庆400074

出　　处：《铁道科学与工程学报》2022年第6期1638-1646,共9页Journal of Railway Science and Engineering

基　　金：国家自然科学基金资助项目(62073051);重庆市教委科学技术研究项目青年项目(KJQN201900726);重庆市教委科学技术研究项目重大项目(KJZD-M202000702)。

摘　　要：桥梁表观病害检测是保证桥梁设施安全的关键技术之一。深度卷积网络因其强大的特征提取能力,被广泛应用于土木工程领域的结构病害识别与检测,然而在土木工程领域中往往缺乏用于训练深度学习网络的高质量大规模病害图像数据集。针对上述问题,提出一种基于迁移学习的桥梁表观病害检测方法。该方法运用迁移学习技术,通过迁移VGG16网络模型结构及全部卷积层参数,并在迁移后的模型结构上添加新的全连接层,以此来解决训练数据集不足的问题。运用动态学习率调整策略,以不同的学习率对卷积层和全连接层参数分别进行微调,用于提高模型的识别准确率。实验对比ResNet18,ResNet50,VGG19,VGG16等主流深度学习网络模型,该方法在验证集上取得了最高准确率,为98.86%。用实拍的未经过处理的桥梁表观病害图像数据集进行测试,该方法的整体结构表观病害识别准确率达到88.33%,其中泛碱、露筋和裂缝3类病害的测试准确率分别达到96.25%,80.00%和88.75%,具有较高的病害识别准确率,可以用于在役桥梁表观病害识别。Bridge surface distress detection is one of the key technologies to ensure the safety of bridge facilities.Deep convolution network is widely used in the identification and detection of structural damages in civil engineering due to its powerful feature extraction capability.However,in the field of civil engineering,there is often a lack of high-quality large-scale distress image datasets for training deep learning network.To solve the above problems,a bridge surface distress detection method based on transfer learning was proposed.The method used the transfer learning technology to solve the problem of insufficient training dataset by migrating the VGG16 network model structure and all convolution layer parameters and adding a new full connection layer to the migrated model structure.The dynamic learning rate adjustment strategy was used to fine-tune the parameters of the convolution layer and the full connection layer with different learning rates to improve the recognition accuracy of the model.Compared with the mainstream deep learning network models such as ResNet18,ResNet50,VGG19 and VGG16,this method achieved the highest accuracy of 98.86% in the validation set.As tested by the unprocessed bridge surface distress image dataset,the overall structural surface distress identification accuracy of this method reached 88.33%.Among them,the test accuracy of corrosion,exposed rebar and cracks reached 96.25%,80.00% and 88.75%,respectively,which has a high distress identification accuracy and can be used to identify surface distress of in-service bridges.

关 键 词：桥梁表观病害检测 迁移学习 深度卷积网络 动态学习率调整 微调 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]