基于VGG16-UNet语义分割模型的路面龟裂形态提取与量化  

Extraction and quantification of pavement alligator crack morphology based on VGG16-UNet semantic segmentation model

作　　者：张伟光[1] 钟靖涛 呼延菊 马涛[1] 朱俊清 何亮[2] ZHANG Wei-guang;ZHONG Jing-tao;HUYAN Ju;MA Tao;ZHU Jun-qing;HE Liang(School of Transportation,Southeast University,Nanjing 210096,Jiangsu,China;School of Civil Engineering,Chongqing Jiaotong University,Chongqing 400074,China)

机构地区：[1]东南大学交通学院,江苏南京210096 [2]重庆交通大学土木工程学院,重庆400074

出　　处：《交通运输工程学报》2023年第2期166-182,共17页Journal of Traffic and Transportation Engineering

基　　金：国家重点研发计划(2020YFB1600102);国家自然科学基金项目(52278443)。

摘　　要：提出了基于VGG16-UNet语义分割模型的路面龟裂分割技术,开发了该模型提取龟裂形态的量化评价方法;通过移动拍照设备采集路面干燥、浸水与存在道路标线影响等不同条件下的龟裂图像,构建了涵盖轻度、中度、重度3种破损程度的多场景龟裂图像数据集;分析了VGG16-UNet、VGG19-UNet、PSPNet、SegNet与DeepLab v3+语义分割模型的龟裂图像分割效果与相应分割指标,明确了最优分割模型并提取了龟裂形态特征;对比了外接正矩形、最小外接矩形和凸包3种轮廓形状拟合边界方法,确定了最优拟合边界,并据此计算了龟裂面积;提出了基于轮廓边界极点坐标的龟裂块度计算方法,分析了路面龟裂面积和块度特征。研究结果表明:VGG16-UNet模型与VGG19-UNet、PSPNet、SegNet、DeepLab v3+这4种语义分割模型相比,VGG16-UNet具有数量为14710464的最小总参数量,模型训练速度可达每个世代训练时间为118 s,并能达到每个图像0.021 s的最高预测速度,同时在公开数据集Crack500、CrackTree与EdmCrack600中对复杂背景的裂缝分割准确率为81%,召回率为82%,调和均值为0.81,平均交并比为0.73,优于4种对比模型,模型泛化能力更高;凸包边界拟合方法计算龟裂面积最小,相较于外接正矩形与最小外接矩形边界拟合方法,拟合率分别提高了14.47%、9.30%,能得到最优的路面龟裂轮廓边界;基于轮廓边界极点计算的龟裂块度能有效评价龟裂破损程度,弥补了现有路面龟裂块度的计算难题。A pavement alligator crack segmentation technology was proposed based on the VGG16-UNet semantic segmentation model,and a quantitative evaluation method was developed for extracting alligator crack morphology via the proposed model.Alligator crack images were collected by a mobile camera under the conditions of dry pavement,waterlogged pavement,and road marking.A multi-scene alligator crack image dataset was established including light,medium,and heavy damages.The alligator crack image segmentation effects and corresponding segmentation indexes of VGG16-UNet,VGG19-UNet,PSPNet,SegNet,and DeepLab v3+were analyzed to select the optimal segmentation model,and the alligator crack morphology features were extracted.Three contour fitting boundary methods of the circumscribed regular rectangle,minimum circumscribed rectangle,and convex hull were compared.The optimal fitting boundary was determined,and the area of alligator cracks was calculated.A calculation method of alligator crack fragmentation based on the coordinates of the extreme points of the contour boundary was proposed.The pavement alligator crack area and fragmentation features were analyzed accordingly.Research results show that compared with four semantic segmentation models of VGG19-UNet,PSPNet,SegNet,and DeepLab v3+,VGG16-UNet has the smallest model parameter quantity of 14 710 464,and the training speed can reach 118 s for each training epoch.It possesses the highest prediction speed of 0.021 s for each image.Additionally,VGG16-UNet shows a precision of 81%,recall of 82%,harmonic mean of 0.81,and mean intersection over union(MIoU) of 0.73 on open-source Crack500 dataset,CrackTree dataset,and EdmCrack600 dataset.It outperforms the other four models and shows higher generalization ability.The proposed convex hull boundary fitting method gets the smallest alligator crack area.Compared with the boundary fitting methods of circumscribed regular rectangle and minimum circumscribed rectangle,the proposed method increases the fitting rate by 14.47% and 9.30%,res

关 键 词：路面工程 语义分割 龟裂边界拟合 龟裂形态 龟裂面积 龟裂块度 

分 类 号：U418.66[交通运输工程—道路与铁道工程]