基于深度学习的预制梁表面气泡缺陷检测  

Deep Learning-based Detection of Bubble Defects on the Surface of Precast Beams

作　　者：陈烨[1] 夏文俊 钱家辉 赵章焰[1] CHEN Ye;XIA Wenjun;QIAN Jiahui;ZHAO Zhangyan(School of Transportation and Logistics Engineering,Wuhan University of Technology,Wuhan 430063,China;China Railway Siyuan Survey and Design Group Co.Ltd.,Wuhan 430063,China)

机构地区：[1]武汉理工大学交通与物流工程学院,武汉430063 [2]中铁第四勘察设计院集团有限公司,武汉430063

出　　处：《武汉理工大学学报（交通科学与工程版）》2024年第2期380-384,391,共6页Journal of Wuhan University of Technology（Transportation Science & Engineering）

基　　金：中交集团首个揭榜挂帅科技攻关项目(2021-ZJKJ-JBGS01)。

摘　　要：文中提出一种基于YOLOv5s的预制梁表面气泡缺陷检测算法.该算法在原模型的基础上引入CBAM注意力模块,增强通道间信息的关联性及兴趣特征的关注度;在颈部网络中用BiFPN加权双向金字塔结构,改进网络特征融合模块,实现快速的多尺度特征融合.在检出气泡缺陷后,提出基于面积和直径的两个评价指标对气泡进行分类.结果表明:改进模型具有更强的特征提取能力,平均检测精度(mAP)为95.8%,相对于原模型提高了2.3%,准确率提高了6.5%,召回率提高了3.5%,在气泡缺陷检测任务中有效减少了漏检和误检,具备更好的检测性能.An algorithm for detecting bubble defects on the surface of precast beams based on YOLOv5s was proposed.Based on the original model,the algorithm introduced the CBAM attention module to enhance the relevance of information between channels and the attention of interest features.In the neck network,BiFPN weighted bidirectional pyramid structure was used to improve the network feature fusion module and realize fast multi-scale feature fusion.After detecting bubble defects,two evaluation indexes based on area and diameter were proposed to classify bubbles.The results show that the improved model has stronger feature extraction ability,the average detection accuracy(mAP)is 95.8%,which is 2.3%higher than the original model,the accuracy is 6.5%higher,and the recall is 3.5%higher.In the task of bubble defect detection,the missed detection and false detection are effectively reduced,and the detection performance is better.

关 键 词：预制梁 气泡缺陷 YOLOv5s 注意力机制 BiFPN 

分 类 号：U446[建筑科学—桥梁与隧道工程]