摘要
为了克服传统人工目测法的缺陷,提出针对隧道衬砌裂缝特征的改进YOLOv7算法模型TLC-YOLO (Tunnel Lining Crack-YOLO).对比4类骨干网络对隧道衬砌裂缝的检测效果,认为复杂环境下的裂缝检测存在强背景干扰、训练样本质量不平衡等问题,使用轻量级卷积GSConv和Slim-neck架构,嵌入动态稀疏注意力模块BiFormer加强通道信息传输,提高TLC-YOLO模型的实时反应速度和检测精度,实现更灵活的计算分配和内容感知.为较好地训练样本分配梯度和抑制较差的训练实例,采用Wise-IoU v3作为坐标回归损失函数来提高模型的泛化能力.结果表明:通过自建隧道衬砌裂缝数据集训练之后,与YOLOv7相比,在多组试验中TLC-YOLO模型能同时提高隧道裂缝检测结果的准确率、召回率、F1值和mAP@0.5值,证明TLC-YOLO模型在隧道衬砌裂缝智能识别中具有更好的检测和分类能力.
In order to overcome the inconvenience of the traditional manual visual inspection method,this paper proposes Tunnel Lining Crack-YOLO(TLC-YOLO),an improved YOLOv7 algorithm model for tunnel lining crack characteristics.This paper compares the detection effect of four types of backbone networks on tunnel lining cracks,and concludes that crack detection in complex environments has problems such as strong background interference and imbalance in the quality of training samples.By using lightweight convolutional GSConv and Slim-neck architectures in the TLC-YOLO model,and by embedding a dynamic sparse attention module,BiFormer,and by enhancing the transmission of channel information,we can improve the real-time response speed and detection accuracy of the model,enabling more flexible computation allocation and content awareness.
作者
贺泳超
陈秋南
程家杰
HE Yongchao;CHEN Qiunan;CHENG Jiajie(School of Resources,Environment and Safety Engineering,Hunan University of Science and Technology,Xiangtan 411201,China;Hunan Provincial Key Laboratory of Geotechnical Engineering Stability Control and Health Monitoring,Hunan University of Science and Technology,Xiangtan 411201,China)
出处
《湖南科技大学学报（自然科学版）》
CAS
北大核心
2024年第2期35-43,共9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(52078211)
湖南省交通运输厅科技进步与创新项目(202308)。
