基于可变形卷积网络和YOLOv8的衬砌裂缝检测模型研究

Research on detection model of lining crack based on deformable convolutional network and YOLOv8

为解决裂缝性状发育随机度高、标注框分辨率低、分布密集易重叠、目标相对小等因素引起的智能检测精度及效率差等问题,基于改进可变形卷积神经网络对YOLOv8骨干网络进行融合,提出1种能够适应隧道复杂场景的裂缝检测模型D-YOLO。模型首先对第3版可变形卷积网络(DCNv3)的空间聚合权重softmax归一化步骤进行去除以增强网络卷积效率,再利用新DCNv4对骨干网络C2f卷积模块进行融合以提升对网络图像中不同尺度裂缝性状及空间位置变化的细节感知能力,并采用自建裂缝数据集对SSD,Faster-RCNN,YOLOv5和YOLOv84种检测模型进行对比验证。研究结果表明:D-YOLO的F_(1)分数为80.82%,mAP@0.5为86.90%,相较于SSD、Faster-RCNN、YOLOv5和YOLOv8都有所提升;D-YOLO的单张图像检测速度为20.36 ms,相较于各种对比模型分别加快37.06%、65.33%、45.22%和28.39%;同时,D-YOLO对衬砌裂缝图像特征关注范围有所增加。研究结果可为隧道运营期内衬砌安全检测提供新思路。

In order to solve the problems of poor intelligent detection accuracy and efficiency caused by the factors such as high randomness of crack characteristic development,low resolution of annotation box,dense distribution and easy overlap,and relatively small target,the YOLOv8 backbone network was fused based on the improved deformable convolutional neural network,and a crack detection model D-YOLO that can adapt to complex tunnel scenes was proposed.The normalization step of spatial aggregation weight softmax in the deformable convolutional network v3(DCNv3)was removed to enhance the convolutional efficiency of network,and the new DCNv4 was used to fuse the C2f convolution module of backbone network to enhance the detail perception ability of different scale crack characteristics and spatial position change in the network images.The self-built crack dataset was used to compare and verify four detection models including SSD,Faster-RCNN,YOLOv5,and YOLOv8.The results show that the F_(1) score of D-YOLO is 80.82%,mAP@0.5 is 86.90%,and both of them are improved than those of SSD,Faster-RCNN,YOLOv5,and YOLOv8.The single image detection speed of D-YOLO is 20.36 ms,which is 37.06%,65.33%,45.22%,and 28.39%faster than those of various comparison models,respectively.Meanwhile,the attention range of image features of lining crack is increased through D-YOLO.The research results can provide new ideas for the safety detection of lining during tunnel operation.