基于改进YOLOv5的红外沥青路面裂缝检测方法

Infrared Asphalt Pavement Crack Detection Method Based on Improved YOLOv5

针对沥青路面可见光图像在某些光照条件下检测效果不佳的问题,提出一种基于改进YOLOv5的红外图像沥青路面裂缝检测方法。以YOLOv5s为基础,融合Ghost模块优化主干网络和颈部网络结构,提高模型信息感知的全面性,降低模型复杂程度;引入坐标注意力(Coordinate Attention,CA)模块,增强对裂缝关键信息的提取,提高红外裂缝检测模型的准确度。采用红外热成像技术,结合数据增强构建红外沥青路面裂缝数据集,在自建的数据集上进行对比实验。实验结果表明,该方法的准确率(Precision,P)、召回率(Recall,R)、平均精度均值(mean Average Precision,mAP)可达到87.4%,82.5%,86.7%,相较于YOLOv5s检测方法分别提升3.4%,2.8%,5.0%,且模型计算量节省了48.73%,在嵌入式设备Jetson Xavier NX上可以实现31.25 f·s^(-1)的检测速度,为后续移动机器人端实时路面裂缝检测提供了一种解决方案。

Aiming at the problem that visible images of asphalt pavement have poor detection effect under certain illumination conditions,an infrared image crack detection method of asphalt pavement based on improved YOLOv5 is proposed.Based on YOLOv5s,Ghost module is integrated to optimize the backbone network and neck network structure to improve the comprehensiveness of the model information perception and reduce model complexity;Coordinate Attention(CA)module is introduced to enhance the extraction of key fracture information and improve infrared crack detection model accuracy.Infrared thermal imaging technology is used to construct infrared asphalt pavement crack data set combined with data enhancement,and comparative experiments are carried out on the selfbuilt data set.The experimental results show that Precision(P),Recall(R)and mean Average Precision(mAP)of this method can reach 87.4%,82.5%and 86.7%,which is 3.4%,2.8%and 5.0%higher than Y0L0v5s method,and the model calculation amount saves 48.73%.The embedded device Jetson Xavier NX can achieve a detection speed of 31.25 f·s^(-1),providing a solution for the follow-up real-time pavement crack detection by mobile robots.