基于YOLOX-MobileNetV3模型的路面病害智能识别研究

Automatic Pavement Disease Identification Research Based on YOLOX-MobileNetV3 Model

目前处理探地雷达(GroundPenetrating Radar,GPR)数据主要依赖于人工,容易造成病害识别误判、漏判率大、速度慢等问题,因此GPR智能目标识别已成为近几年的研究热点。本文提出基于卷积神经网络中的YOLOX-MobileNetV3模型来实现路面病害自动识别,利用三维数据的高信息量和深度学习智能提取特征的优势,实现路面病害的智能化识别。首先对三维探地雷达得到的GPR图片进行预处理,然后以3∶1的训练集和测试集数量比例对数据进行3轮训练和测试,并利用平均精确度、全类平均精确度、精确度、召回率、F1值、平均漏检率等指标来评价3次训练和测试的结果。结果表明:YOLOX-MobileNetV3模型的训练损失权重平均为5.014,测试准确率平均为61.35%。该模型识别路面结构病害尤其是裂缝、层间黏结不良的准确率较高。同时随着训练与测试轮数的增加,其精确度也会随之增加,召回率会随之减小。由此可见,YOLOX-MobileNetV3模型能够实现路面病害自动识别。

At present,the processing of GPR data mainly relies on manual processing,and disease identification is easy to cause problems such as false positive,large missed detection rate and slow speed,so GPR intelligent target recognition has become a research hotspot in recent years.In this paper,it is proposed to realize the automatic identification of pavement diseases based on the YOLOX-MobileNetV3 model in convolutional neural networks,and use the high information of three-dimensional data and the advantages of deep learning to extract features intelligently,in order to realize the intelligent identification of pavement diseases.Firstly,the GPR pictures obtained by the 3D GPR are preprocessed,and then the data are trained and tested for three rounds with a ratio of 3∶1 in the number of training sets and test sets,and the results of the three training and tests are evaluated by using the average accuracy,the average accuracy of the whole class,the precision,the recall rate,the F1 value,and the average missed detection rate.The results show that the average training loss weight of the YOLOX-MobileNetV3 model is 5.014,and the average test accuracy is 61.35%.The model has high accuracy in identifying pavement structural diseases,especially cracks and poor adhesion between layers.At the same time,as the number of training and testing rounds increases,its accuracy will also increase,and the recall rate will decrease.It can be seen that the YOLOX-MobileNetV3 model can realize automatic identification of road surface diseases.