基于迁移学习的公路边坡滑塌识别研究  

Research on Identification of Highway Slope Collapse Diseases Based on Transfer Learning

作　　者：胡良军 吴建良 杨军 赵亚宇 HU Liang-jun;WU Jian-liang;YANG Jun;ZHAO Ya-yu(Guangzhou Municipal Engineering Testing Co.,Ltd.,Guangzhou 510520;Guangzhou Institute of Building Science Company Limited,Guangzhou 510440;Hunan Institute of Science and Technology,College of Civil Engineering and Architecture,Yueyang 414006;Tsinghua University,Beijing 100084;Guangdong Provincial Engineering Research Center for Assembled Underground Structure Detection and Monitoring,Guangzhou 510520)

机构地区：[1]广州市市政工程试验检测有限公司,广州510520 [2]广州市建筑科学研究院集团有限公司,广州510440 [3]湖南理工学院土木建筑工程学院,岳阳414006 [4]清华大学,北京100084 [5]广东省装配式地下结构检测与监测工程技术研究中心,广州510520

出　　处：《广州建筑》2022年第6期57-60,共4页GUANGZHOU ARCHITECTURE

基　　金：广州市建筑集团有限公司科技计划项目([2021]-KJ026,[2022]-KJ022)。

摘　　要：自动化拍摄代替人工进行边坡滑塌病害巡查是近年来的新趋势,本文提出基于迁移学习的边坡病害识别方法。首先采集3000余张边坡病害照片,并将样本量扩充了7倍后划分为70:30的训练集与验证集。然后基于预训练目标识别模型,更新模型目标推荐模块,建立了多个Faster R-CNN框架下的目标识别模型。然后结合模型误差随机梯度下降等训练策略,完成模型训练。分别以目标框内病害的判别概率大于90%、80%、70%作为识别依据,计算准确率、召回率,在验证集上验证模型效果。发现:降低目标识别框的类别概率阈值要求后,边坡病害的漏判率有明显下降,准确率降低。训练效果最好的Faster R-CNN ResNet-50 FPN模型以大于70%概率作为识别目标标准时,对验证集的漏判率低于4%。适当放松准确率要求,模型漏判率随之降低,对缺陷、异常的图像识别具有实用价值。Automatic camera shooting replace manual inspection of slope collapse disease is a new trend in recent years.This paper proposes a slope disease identification method based on migration learning.Firstly,more than 3000 photos of slope diseases were collected to mark landslide diseases.The sample set was expanded by 7 times and was divided into 70:30 training set and verification set.Then,based on the pre training target recognition model,the target recommendation module of the model is updated,and multiple target recognition models under the framework of Faster R-CNN are established.Then,combined with the training strategies such as random gradient descent of model error,the model training is completed.Taking the identification probability of diseases in the target frame greater than 90%,80%and 70%as the identification basis,calculate the accuracy and recall rate,and verify the effect of the model on the verification set.It is found that after reducing the category probability threshold of the target recognition frame,the missed judgment rate and accuracy of slope diseases are significantly reduced.When the Faster R-CNN Resnet-50 FPN model with the best training effect takes the probability greater than 70%as the recognition target standard,the missed judgment rate of the verification set is less than 4%.The appropriate relaxation of the accuracy requirements and the reduction of the missing judgment rate of the model have practical value for the identification of unusual diseases such as defects and abnormalities.

关 键 词：公路边坡 滑塌 目标识别 Faster R-CNN 漏判率 

分 类 号：U416.14[交通运输工程—道路与铁道工程]