针对地铁隧道内部光线昏暗,诸多附属设备与衬砌表观病害纹理及灰度近似,病害检测精度低的问题,提出了一种基于改进YOLOv7(You Only Look Once v7)的地铁隧道衬砌表观病害与附属设施的检测方法.在原YOLOv7 模型的骨干部分引入混合卷积模块(Mixed Convolutional Module,ConvMixer)并增加了微小物体检测头,降低网络复杂度并提高对微小病害的敏感度;在原YOLOv7模型的颈部将路径聚合网络(Path Aggregation Network,PANet)替换为双向特征金字塔网络(Bidirectional Feature Pyramid Network,BiFPN),用于捕获全局特征信息,在隧道复杂环境下精确定位;在最大池化卷积(MaxPooling Convolution,MPC)模块中引入无参注意力机制(Simple Parameter Free Attention Mechanism,SimAM),给检测目标的特征矢量赋予三维权重,以增加检测精度.检测结果表明,本文提出的改进模型的检测准确率和召回率分别达到89.1%、87.8%,且权重文件大小仅为59.6 MB,检测速率为89帧/s.该模型在保证较高检测精度的同时降低了权重文件大小,提高了检测速率,适用于隧道巡检系统.
Detection Method for Apparent Defects of Subway Tunnel Lining Based on Improved YOLOv7 Model
A detection method for subway tunnel lining defects and ancillary facilities based on improved YOLOv7(You Only Look Once v7)model was proposed to address the issues of dim lighting inside subway tunnels,similar texture and grayscale of many ancillary equipment and lining defects,and low inspection accuracy of defects.The Mixed Convolutional Module(ConvMixer)was introduced into and a detection head for small objects was added to the backbone of the original YOLOv7 model,which reduced network complexity and increased sensitivity to small defects.The Path Aggregation Network(PANet)was replaced by a Bidirectional Feature Pyramid Network(BiFPN)at the neck of the original YOLOv7 model,which was used to capture global feature information and accurately determine the position in complex tunnel environment.A Simple Parameter Free Attention Mechanism(SimAM)was introduced into the MaxPooling Convolution(MPC)module to assign three-dimensional weights to the feature vectors of the detection target to increase inspection accuracy.The detection results show that the inspection accuracy and recall rate of the improved model proposed in this paper reach 89.1%and 87.8%,respectively,and the weight file size is only 59.6 MB,with a detection rate of 89 frames per second.This model not only ensures high inspection accuracy,but also reduces the weight file size and increases the detection rate,which is suitable for tunnel inspection system.
万方数据
维普
