面向小样本地下排水管道缺陷的识别方法  

Few-shot defect recognition method of underground drainage pipelines

作　　者：琚锋 钱强强 杨珍 尤加俊 JU Feng;QIAN Qiangqiang;YANG Zhen;YOU Jiajun(Huzhou Institute of Surveying and Mapping,Huzhou 313000,China;School of Electronic Information,Huzhou College,Huzhou 313000,China)

机构地区：[1]湖州市测绘院,浙江湖州313000 [2]湖州学院电子信息学院,浙江湖州313000

出　　处：《测绘通报》2025年第3期156-160,共5页Bulletin of Surveying and Mapping

基　　金：湖州市公益性应用研究项目(2024GZ69);浙江省自然资源厅科技项目(2024ZJCH025);浙江省自然科学基金(LY24F030012)。

摘　　要：城市地下排水管道检测、评估和维修是保障排水管道系统安全运行的必要手段。深度学习为排水管道缺陷检测识别的自动化和智能化提供了新的方法。然而,部分管道缺陷样本集匮乏和各缺陷类型间样本不均衡极大地影响了管道缺陷识别模型的泛化能力和稳健性,导致现有排水管道缺陷识别模型易出现错检、漏检、识别准确度较低等问题。针对上述问题,本文基于度量学习提出了一种面向小样本管道缺陷的识别方法,采用深度细部特征表征检测图像,依据支持集中不同缺陷检测图像特征向量与查询集图像之间的度量值识别其缺陷的类型。试验结果表明,该方法识别非常见管道缺陷类型的准确率为65%左右,可为排水管道缺陷样本不充足不均衡情况下智能识别提供参考。The inspection,evaluation and maintenance of urban underground drainage pipelines are the necessary means to ensure the safe of drainage pipeline system.Deep learning technology provides a new method for the automation and intelligence of underground drainage pipelines defects detection and recognition.However,the lack of sample set for some pipeline defect and the imbalance of samples among different defect types greatly affect the generalization ability and robustness of defect recognition models,resulting in the problems of false detection,missed detection,and low recognition accuracy of existing defect recognition models.To solve the above problems,this paper proposes a defect recognition method for few-shot defect based on metric learning.The method uses detailed features to represent the test images,and identifies the types of defects according to the similarity between the feature vectors of different defect detection images in the support set and the images in the query set.The experimental results show that the accuracy of the proposed method in identifying scarce pipeline defect types is about 65%,and can be used as a solution for intelligent recognition in the case of insufficient and unbalanced pipeline defect samples.

关 键 词：排水管道缺陷 小样本学习 CCTV检测 度量学习 智能识别 

分 类 号：P237[天文地球—摄影测量与遥感]