基于One-Stage目标检测的路面损害检测算法对比分析

Comparative analysis of road damage detection algorithms based on One-Stage object detection

针对如何方便、快速、准确地检测常见的路面损害问题,提高路面损害的检测效率,选取三种常用的OneStage目标检测算法(SSD、YOLOv3、RetinaNet),以智能手机拍照的方式收集路面损害数据,利用LabelImg工具制作图像标签数据集,通过替换其主干网络的方式训练了6种检测模型(SSD-MobileNetv1、YOLOv3-MobileNetv1、YOLOv3-DarkNet53、YOLOv3-ShuffleNetv2、YOLOv3-ResNet50、RetinaNet-ResNet50),并采用两组不同类型的数据集(路面背景较为干净的数据集A和路面背景含有大量树枝阴影和水渍的数据集B)对上述6种模型的检测性能进行对比分析。实验结果表明:在检测精度方面,YOLOv3-ResNet50模型的检测精度高于另外5种算法模型,比YOLOv3-MobileNetv1模型提高1.6个百分点,比RetinaNet-ResNet50模型提高3.7个百分点,比YOLOv3-DarkNet53模型提高4.5个百分点;在模型参数规模方面,SSD-MobileNetv1模型最轻且参数规模最小,比YOLOv3-MobileNetv1模型减少76.9%的参数量,比YOLOv3-ShuffleNetv2模型减少21.4%的参数量;在模型漏检和误检率方面,YOLOv3-DarkNet53模型的漏检率最低(7/403),YOLOv3-MobileNetv1模型的误检率最低(8/403)、鲁棒性最好。因此YOLOv3-ResNet50算法模型适合用于路表面较为干净且精度要求较高的路面损害检测;SSD-MobileNetv1和YOLOv3-ShuffleNetv2的参数规模较小,可应用于嵌入式设备的检测;YOLOv3-MobileNetv1不易受路面条件的干扰,能够满足正常路面检测要求。

To detect common road damage problems conveniently,quickly and accurately and improve the detection efficiency of road damage,three common One-Stage object detection algorithms(SSD(Single Shot multibox Detector),YOLOv3 and RetinaNet)were selected,the road damage data were collected by taking photos with smart phones,and the image label dataset was produced by using LabelImg tool. Six detection models(SSD-MobileNetv1,YOLOv3-MobileNetv1,YOLOv3-DarkNet53,YOLOv3-ShuffleNetv2,YOLOv3-ResNet50,RetinaNet-ResNet50)were trained by replacing their backbone networks. Two different types of datasets(dataset A with a clean road background and dataset B with a large number of tree branches and water stains)were used to compare the detection performance of the above six models. The experimental results demonstrated that the detection accuracy of the YOLOv3-ResNet50 model was higher than those of the other five algorithms. It showed 1. 6 percentage points improvement over YOLOv3-MobileNetv1 model,3. 7 percentage points improvement over RetinaNet-ResNet50 model,and 4. 5 percentage points improvement over YOLOv3-DarkNet53 model. In terms of model parameter size,the SSD-MobileNetv1 model was the lightest and had the smallest parameter size,reducing the number of parameters by 76. 9% compared to the YOLOv3-MobileNetv1 model and 21. 4% compared to the YOLOv3-ShuffleNetv2 model. In terms of model missed and false detection rates,the YOLOv3-DarkNet53 model showed the lowest missed detection rate(7/403)and the YOLOv3-MobileNetv1 model showed the lowest false detection rate(8/403)and the best robustness. As a result,the YOLOv3-ResNet50 algorithm model is found to be suitable for road damage detection where the road surface is relatively clean and the accuracy requirement is high,SSD-MobileNetv1 and YOLOv3-ShuffleNetv2 with small parameter size are suitable for partial detection by embedded equipment. YOLOv3-MobileNetv1 is less susceptible to interference from road conditions and is able to meet normal road detection requirements.