基于改进YOLOv4的隧道衬砌裂缝检测算法  

Tunnel Lining Crack Detection Algorithm Based on Improved YOLOv4

作　　者：周中[1] 张俊杰 鲁四平[1] ZHOU Zhong;ZHANG Junjie;LU Siping(School of Civil Engineering,Central South University,Changsha 410075,China)

机构地区：[1]中南大学土木工程学院,湖南长沙410075

出　　处：《铁道学报》2023年第10期162-170,共9页Journal of the China Railway Society

基　　金：国家自然科学基金(50908234);湖南省自然科学基金(2020JJ4743)。

摘　　要：针对传统隧道衬砌裂缝检测手段中存在的检测精度低、泛化能力差、检测速度慢的问题,对YOLOv4目标检测算法进行改进:引入Mosaic数据增强技术对输入图片数据进行预处理;采用轻量级网络MobilenetV3取代CSPDarknet53作为YOLOv4神经网络的主干特征提取网络;将YOLOv4网络中卷积核大小为3×3的标准卷积替换为深度可分离卷积。为验证改进后算法的有效性和可靠性,采用Faster-RCNN、SSD、YOLOv3、YOLOv4四种算法进行对比验证分析,结果表明:该算法在检测性能方面表现优异,测试集的平均精度为78.05%,精确率以及召回率的加权调和平均值为84.44%,均高于其余四种算法。在模型大小方面,该算法模型的大小仅为55.1 MB,相对于原始YOLOv4模型压缩了78.0%,且远小于Faster-RCNN、SSD、YOLOv3模型。此外,该算法的单张图片的检测时间为23.75 ms,每秒帧数为42.1帧/s,很好地满足了隧道衬砌裂缝进行实时检测时移动设备对帧率的要求。且算法泛化能力良好,能够较为准确的对不同光照和复杂背景条件下的裂缝进行检测并标记。基于提出的改进YOLOv4算法构建隧道衬砌裂缝检测平台,实现对实际隧道工程中衬砌裂缝的准确、快速、智能化识别。In response to the problems of low detection accuracy,poor generalization ability and slow detection speed in the traditional tunnel lining crack detection methods,this paper improved the YOLOv4 target detection algorithm by introducing the mosaic data enhancement technology to preprocess the input image data,using lightweight network Mobilenetv3 to replace CSPdarknet53 as the backbone feature extraction network of YOLOv4 neural network,and replacing the standard convolution with 3×3 size of convolution kernel in YOLOv4 network with deep separable convolution.In order to verify the effectiveness and reliability of the improved algorithm,Faster RCNN,SSD,YOLOv3 and YOLOv4 algorithms were used for comparative verification and analysis.The results show that the algorithm in this paper has excellent detection performance,with an average accuracy of the test set of 78.05%,and weighted harmonic average of accuracy and recall of 84.44%,which are higher than the other four algorithms.In terms of model size,the size of the algorithm model in this paper is only 55.1MB,which is 78.0%compressed compared with the original YOLOv4 model,and is much smaller than Faster RCNN,SSD and YOLOv3 models.In addition,the detection time of a single picture of this algorithm is 23.75 ms with frames per second of 42.1 frames/s,which well meets the requirements of mobile equipment for frame rate in real-time detection of tunnel lining cracks.The algorithm,with good generalization ability,can accurately detect and mark cracks under different lighting and complex background conditions.Based on the improved YOLOv4 algorithm proposed in this paper,a tunnel lining crack detection platform was constructed,realizing accurate,fast and intelligent identification of lining cracks in actual tunnel works.

关 键 词：隧道工程 衬砌裂缝 目标检测 YOLOv4 深度学习 神经网络 

分 类 号：U455[建筑科学—桥梁与隧道工程]