基于改进YOLOv8的沥青路面深度图像病害检测算法  

Asphalt Pavement Depth Image Disease Detection Algorithm Based on Improved YOLOv8

作　　者：英红 覃云涛 刘曦民 朱嘉丽 陈蔚铭 YING Hong;QIN Yuntao;LIU Ximin;ZHU Jiali;CHEN Weiming(School of Architecture and Transportation Engineering,Guilin University of Electronic Technology,Guilin 541004,China)

机构地区：[1]桂林电子科技大学建筑与交通工程学院,广西桂林541004

出　　处：《湖南科技大学学报(自然科学版)》2025年第1期88-101,共14页Journal of Hunan University of Science And Technology:Natural Science Edition

基　　金：国家自然科学基金资助项目(51668012,51968011)。

摘　　要：为解决沥青路面病害检测人工成本高、检测精确度与效率低下的问题,提出一种面向多尺度目标沥青路面深度图像的轻量化道路病害检测模型RBDN-YOLO-E.通过无人机快速扫描沥青路面并建立三维点云模型得到路面深度图像,将2 496张深度图像建立沥青路面病害检测数据集,并按7∶2∶1的比例划分为训练集、验证集以及测试集.在网络主干的C2f中引入重塑空间注意力卷积模块RFAConv得到C2f_RFAConv模块以关注感受野空间特征,为图像每个区域的特征提取提供了定制化的关注度;在网络颈部的C2f中引入可变形卷积DCNv3得到C2f_DCNv3模块以提升模型对目标形变的建模能力,降低输出通道数和模型计算成本,提高识别的精确度;将网络中SPPF模块替换为SPPELAN模块的空间金字塔池化结构以生成更多尺度,提高特征表示能力.试验结果表明:RBDN-YOLO-E模型相对于YOLOv8n模型的mAP50、F1值、精确率、召回率和推理时间分别提升了1.70%,2.00%,4.29%,2.00%,3.40%,而模型计算量、参数量和模型大小分别降低了0.9 GFLOPs, 0.73 M,1.4 M.通过结合无人机与改进模型RBDN-YOLO-E,可以更安全、快速和准确地提取沥青路面病害,有效解决传统沥青路面病害的检测精度与成本问题,提高对沥青路面病害的检测效率.In order to solve the problems of high labor cost,low detection accuracy and low efficiency of asphalt pavement disease detection,a lightweight road disease detection model RBDN-YOLO-E for multi-scale target asphalt pavement depth image is proposed.The pavement depth image is obtained by scanning the asphalt pavement quickly and establishing a three-dimensional point cloud model.The 2496 depth images are established into a data set of asphalt pavement disease detection,and divided into training set,verification set and test set according to the ratio of 7:2:1.The receptive field attention convolution module RFAConv is introduced into the C2f of the network backbone to obtain the C2f_RFAConv module to focus on the spatial features of the receptive field,which provides customized attention for the feature extraction of each region of the image.The deformable convolution DCNv3 is introduced into the C2f of the network neck to obtain the C2f_DCNv3 module to improve the modeling ability of the model to the target deformation,reduce the number of output channels and the computational cost of the model,and improve the accuracy of recognition.The SPPF module in the network is replaced by the spatial pyramid pooling structure of the SPPELAN module to generate more scales and improve the feature representation ability.The experimental results show that the mAP50,F1 value,precision rate,recall rate and inference time of the RBDN-YOLO-E model are increased by 1.70%,2.00%,4.29%,2.00%and 3.40%respectively compared with the Y0LOv8n model,while the model calculation amount,parameter quantity and model size are reduced by 0.9 GFLOPs,0.73 M and 1.4 M respectively.By combining the UAV and the improved model RBDN-YOLO-E,the asphalt pavement disease can be extracted more safely,quickly and accurately,which can effectively solve the problem of traditional asphalt pavement disease detection accuracy and cost,and improve the efficiency of asphalt pavement disease detection.

关 键 词：病害检测 轻量化 无人机 深度图像 YOLOv8n 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] U418.6[自动化与计算机技术—计算机科学与技术]