基于改进Cascade R-CNN算法的道路表面缺陷检测  

Road Defect Detection Based on Improved Cascade R-CNN Algorithm

作　　者：刘旭 陈里里[1] 范国栋 李博涵 LIU Xu;CHEN Lii;FAN Guodon;LI Bohan(School of Electromechanical and Vehicle Engineering,Chongqing Jiaotong University,Chongqing 400074,China)

机构地区：[1]重庆交通大学机电与车辆工程学院,重庆400074

出　　处：《自动化与仪器仪表》2023年第10期10-14,共5页Automation & Instrumentation

基　　金：中国博士后科学基金面上资助(2020M 683256);重庆市技术创新与应用发展专项重点项目(cstc2020jscx-gksbX0010);交通工程应用机器人重庆市工程实验室2020年度开放课题,CELTEAR-KFKT-202003。

摘　　要：作为最基础的交通基础设施之一,道路对交通运输以及城市发展有着不可替代的作用。道路表面缺陷是道路状态的真实反映,准确地进行道路缺陷检测对道路状态检测和维护具有重要意义。道路缺陷分布具有不确定性,现阶段采用的人工检测方法存在检测效率低、检测周期长等缺点。针对目前道路表面缺陷检测中存在的问题,提出一种改进的Cas-cade R-CNN道路缺陷检测算法,引入了递归特征金字塔结构,使融合特征获得更多的语义信息细节,更利于小目标检测;选用ResNet50作为主干网络并进行改进,使其能够接受来自递归特征金字塔的特征输入。实验结果表明,改进后的算法在测试数据集上的表现优于Faster R-CNN、Grid R-CNN,检测精度和小目标检测能力均得到了提升。As one of the most basic transport infrastructures,roads play an irreplaceable role in transport as well as in urban de-velopment.Road defects are a true reflection of the state of the road,so accurate road defect detection is of great significance to road condition detection and maintenance.Road defects have the characteristics of random distribution and uncertainty,the current stage of field detection methods based on manual vision,low detection efficiency,detection cycle plant,and the existence of a certain degree of danger.To address the current problems in road detection,this paper proposes an improved Cascade R-CNN road defect detection algorithm,the new algorithm introduces a recursive feature pyramid structure,which takes into account more feature information in feature fusion,increases feature utilization,and is more conducive to feature fusion capability and small target detection;ResNet50 is selected as the feature extraction network and improved so that it can ResNet50 was chosen as the feature extraction network and im-proved so that it could accept feature inputs from the recursive feature pyramid to further improve the feature detection capability.The experimental results show that the improved algorithm performs better than Fast R-CNN and Grid R-CNN in the test data set,and the detection accuracy and small target detection ability are improved.

关 键 词：道路表面缺陷 缺陷检测 Cascade R-CNN 目标检测 深度学习 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]