基于多尺度网格的隧道衬砌裂缝目标检测方法  

Tunnel lining crack target detection method based on multi-scale grid

作　　者：李少华 沈翔 李树忱 江玉生[4] Li Shaohua;Shen Xiang;Li Shuchen;Jiang Yusheng(China Railway 15th Bureau Group Co,Ltd.,Shanghai 200070,China;Shenzhen University,Shenzhen 518060,China;China Universityof Mining and Technology,Xuzhou 221116,China;China University of Mining and Technology-Beijing,Bejing 100083,China)

机构地区：[1]中铁十五局集团有限公司,上海200070 [2]深圳大学,广东深圳518060 [3]中国矿业大学,徐州221116 [4]中国矿业大学(北京),北京100083

出　　处：《土木工程学报》2024年第S1期95-101,共7页China Civil Engineering Journal

基　　金：国家重点研发计划(2023YFC3806705);中国施工企业管理协会青年创新项目(2023-B-028);中国铁建地下空间利用领域科研计划项目(2024-W25);中铁十五局集团有限公司科研项目(2022B4,2021C61)。

摘　　要：为提高盾构隧道衬砌裂缝检测方法效率与精度,使Faster R-CNN目标检测模型能够更为精确地在多尺度上完成对裂缝的识别,增加网格检测器作为小尺度的目标检测手段,同时采用Swin Transformer作为特征提取骨干网络来获取盾构隧道衬砌表观缺陷的特征,优化大尺度目标监测模型,以达到多尺度网格裂缝识别的效果。收集现场管片裂缝图像作为数据集,验证了所提出方法的有效性。研究结果表明:(1)改进后的大尺度目标检测的mAP50-95精度达到了79.1%,小尺度目标检测的分类精确度达到了95.3%。(2)Swin Transformer网络具有更好的性能,在大尺度目标检测中,精度提升了4.0~7.0%。(3)通过多尺度目标检测分支的精度指标绘制的最终目标检测结果能够更好地对混凝土表面的裂缝进行预测。本文改进后的裂缝检测模型将更加准确高效地反映裂缝问题,可为现场隧道衬砌裂缝识别提供参考依据。In order to improve the efficiency and accuracy of the shield tunnel lining crack detection method,the Faster R-CNN target detection model accurately complete the identification of cracks at multiple scales.The grid detector is added as a small-scale target detection method.At the same time,Swin Transformer is used as a feature extraction backbone network to obtain the characteristics of the apparent defects of the shield tunnel lining,and the large-scale target monitoring model is optimized to achieve the effect of multi-scale grid crack identification.Field segment crack images are collected as a dataset to verify the effectiveness of the proposed method.The results show that:(1)The mAP50-95 accuracy of the improved large-scale target detection reaches 79.1%,and the classification accuracy of the small-scale target detection reaches 95.3%.(2)The Swin Transformer network has better performance,and the accuracy is improved by 4.0%-7.0%in large-scale target detection.(3)The final target detection results drawn by the accuracy index of the multi-scale target detection branch can better predict the cracks on the concrete surface.The improved crack detection model in this paper will reflect the crack problem more accurately and efficiently,which can provide reference for the identification of lining cracks in tunnel.

关 键 词：隧道衬砌 裂缝 多尺度网格 目标检测 目标标记方法 

分 类 号：U45[建筑科学—桥梁与隧道工程]