基于自注意力机制与卷积神经网络的隧道衬砌裂缝智能检测  

Intelligent Detection of Tunnel Lining Cracks Based on Self⁃attention Mechanism and Convolution Neural Network

作　　者：周中[1,2] 闫龙宾 张俊杰 张东明 龚琛杰[1,2] ZHOU Zhong;YAN Longbin;ZHANG Junjie;ZHANG Dongming;GONG Chenjie(School of Civil Engineering,Central South University,Changsha 410075,China;Hunan Tieyuan Civil Engineering Testing Co.,Ltd.,Changsha 410075,China;College of Civil Engineering,Tongji University,Shanghai 200092,China)

机构地区：[1]中南大学土木工程学院,湖南长沙410075 [2]湖南铁院土木工程检测有限公司,湖南长沙410075 [3]同济大学土木工程学院,上海200092

出　　处：《铁道学报》2024年第9期182-192,共11页Journal of the China Railway Society

基　　金：国家自然科学基金(51908557);长沙理工大学公路养护技术国家工程研究中心开放基金(kfj220101);湖南省自然科学基金(2024JJ5428)。

摘　　要：卷积操作固有的局部性,导致难以捕获裂缝的全局特征表示。针对这一问题,以YOLOv5网络框架为基础,通过融合自注意力机制和卷积神经网络提出一种高精度隧道衬砌裂缝智能检测算法ST-YOLO。构建一种双分支特征提取模块,分支一采用基于自注意力机制的Swin Transformer网络提取裂缝的全局特征,分支二采用基于卷积神经网络的CSPDarkent提取裂缝的局部细节特征;采用卷积注意力增强的特征融合模块对两个分支所提取的裂缝特征进行多尺度融合,以实现裂缝全局特征和局部细节特征互补;借助YOLOv5的解耦头模块输出得到裂缝的分类置信度和位置结果。为验证本文提出算法的有效性,选用SSD、YOLOv4、YOLOv5、EfficientDet、Faster-RCNN五种基于卷积神经网络的目标检测算法进行对比分析。结果表明:在本文构建的隧道衬砌裂缝数据集上,ST-YOLO模型的F1分数和AP分别为8395%和8526%,相较于五种对比算法具有更高的识别精度及更好的环境适应性,适用于实际隧道工程裂缝病害检测任务。The inherent localization of the convolutional operation makes it difficult to capture the global feature repre-sentation of the cracks.To address this problem,proposes a high-precision tunnel lining crack intelligent detection algo-rithm,ST-YOLO,based on the YOLOv5 network framework by fusing the self-attention mechanism and convolutional neural networks(CNN).In ST-YOLO,a double-branched feature extraction module was designed.In the first branch,Swin Transformer based on self-attention mechanism was used to extract the global features of cracks,while in the second branch,CSPDarkent based on convolution neural network was used to extract the local details of cracks.Furthermore,the convolution attention enhanced feature fusion module was used to fuse the crack features extracted from the two bran-ches on multiple scales,so as to achieve the complementarity of the global features and local features.Finally,the clas-sification confidence and location results of cracks were obtained by using the decoupled head of YOLOv5.To verify the effectiveness of the ST-YOLO,five target detection algorithms based on CNN,including SSD,YOLOv4,YOLOv5,Effi-cientDet and Faster-RCNN,were selected for comparative analysis.The results show that based on the tunnel lining crack data set constructed in this paper,the F1 score and AP of ST-YOLO model are 8395%and 8526%,respectively.Compared with the five comparison algorithms,ST-YOLO has higher recognition accuracy and better environmental adaptability,which is more suitable for the crack detection task for tunnels.

关 键 词：隧道工程 结构裂缝 目标检测 深度学习 神经网络 

分 类 号：U45[建筑科学—桥梁与隧道工程]