基于改进Swin-Transformer的农村路面裂缝检测算法  

Rural pavement crack detection algorithm based on improved Swin-Transformer

作　　者：李禹萱 宋伟东[1,2] 孙尚宇[1,2] 张晋赫 LI Yuxuan;SONG Weidong;SUN Shangyu;ZHANG Jinhe(School of Mapping and Geographical Science,Liaoning Technical University,Fuxin Liaoning 123000,China;Collaborative Innovation institute of Geospatial Information Service,Liaoning Technical University,Fuxin Liaoning 123000,China)

机构地区：[1]辽宁工程技术大学测绘与地理科学学院,辽宁阜新123000 [2]辽宁工程技术大学地理空间信息服务协同创新研究院,辽宁阜新123000

出　　处：《北京交通大学学报》2024年第5期88-97,共10页JOURNAL OF BEIJING JIAOTONG UNIVERSITY

基　　金：国家自然科学基金(42071343)。

摘　　要：裂缝作为农村道路病害的主要组成部分,在检测过程中易受到路面阴影、杂草、泥土等干扰因素的影响,导致基于路面图像的自动化检测变得更加困难.为解决这一问题,提出一种基于Swin-Transformer主干网络的农村道路裂缝检测(Swin-Transformer Rural road Crack Detection,S-TRCD)模型.针对模型在裂缝特征提取过程中受到周围干扰物影响导致识别精度降低的问题,设计一种自适应的混合注意力机制模块CAS(Channel and Spatial),该模块能够在空间和通道两个维度上调整裂缝的权重,提高检测模型的抗干扰能力;针对多个裂缝在同一图像上尺寸差异较大导致识别困难的问题,改进了一种带注意力机制的多尺度目标检测头AHead(Attention Head),该检测头可以自适应调整网络感受野,实现多尺度的裂缝检测.为验证S-TRCD检测模型的检测性能,制作农村路病害基准数据集LNTU_RDD_NC,并对S-TRCD检测模型以及路面裂缝检测领域常用的改进YOLOv5、Faster R-CNN、YOLOv8检测模型进行训练.实验结果表明:S-TRCD检测模型在农村路面裂缝检测中较改进YOLOv5、Faster R-CNN、YOLOv8检测模型平均识别精度分别高4.06%、12.12%、2.84%,证明在农村路面裂缝检测领域中,S-TRCD检测模型具有较好的检测性能.Cracks are a primary form of rural pavement distress,and their detection is often hindered by interference factors such as road shadows,weeds,and soil,complicating automated detection based on road images.To address this issue,this study proposes the Swin-Transformer Rural Road Crack Detection(S-TRCD)model,which leverages the Swin-Transformer backbone network.To mitigate the reduced recognition accuracy caused by surrounding interference during feature extraction,an adaptive hybrid attention mechanism module,CAS(Channel and Spatial),is designed.This module adjusts the crack weights in both spatial and channel dimensions,enhancing the model’s resistance to interference.To address the challenge of identifying cracks of varying sizes within the same image,a multi-scale object detection head with an attention mechanism,AHead(Attention Head),is developed.This detection head adaptively adjusts the network’s receptive field,enabling effective multi-scale crack detection.A rural pavement distress benchmark dataset,LNTU_RDD_NC,is created to evaluate the performance of the S-TRCD model.The study also trains and compares the S-TRCD model with commonly used detection models in the field,including improved YOLOv5,Faster R-CNN,and YOLOv8.Experimental results demonstrate that the S-TRCD model achieves mean average precision 4.06%,12.12%,and 2.84%higher than the improved YOLOv5,Faster R-CNN,and YOLOv8 models,respectively,highlighting its superior detection performance for rural pavement crack detection.

关 键 词：裂缝检测 深度学习 农村道路 多尺度特征 混合注意力机制 

分 类 号：U418.67[交通运输工程—道路与铁道工程]