[1]殷浤益,孙全,熊友亮,等.基于Transformer和CNN的排水管网缺陷识别方法[J].中国给水排水,2025,41(5):123-129.
　YINHong-yi,SUNQuan,XIONGYou-liang,et al.A Method for Identifying Defects in Drainage Networks Based on Transformer and CNN[J].China Water & Wastewater,2025,41(5):123-129.

点击复制

基于Transformer和CNN的排水管网缺陷识别方法

中国给水排水[ISSN:1000-4062/CN:12-1073/TU] 卷: 第41卷期数: 2025年第5期页码: 123-129 栏目: 出版日期: 2025-03-01

Title:: A Method for Identifying Defects in Drainage Networks Based on Transformer and CNN

作者:: 殷浤益^1,2，孙全³，熊友亮^1,2，郭林松³，曾招财³，郭清露¹; （1.长江勘测规划设计研究有限责任公司，湖北武汉 430010；2.水利部长江勘测技术研究所，湖北武汉 430010；3.长江生态环保集团有限公司，湖北武汉 430010）

Author(s):: YIN　Hong-yi^1，2， SUN　Quan³， XIONG　You-liang^1，2， GUO　Lin-song³， ZENG　Zhao-cai³， GUO　Qing-lu¹; （1. Changjiang Survey， Planning， Design and Research Co. Ltd.， Wuhan 430010， China； 2. Changjiang Institute of Survey Technical Research， Ministry of Water Resources， Wuhan 430010， China； 3. Yangtze Ecology and Environment Co. Ltd.， Wuhan 430010， China）

关键词:: 排水管网检测; 缺陷识别; RT-DETR; 注意力机制; inner-MPDIoU

Keywords:: drainage network inspection; defect identification; RT-DETR; attention mechanism; inner-MPDIoU

摘要:: 鉴于传统人工检测并识别排水管网缺陷的方法存在工作量巨大且费时费力的问题，提出了一种基于卷积神经网络（CNN）和Transformer的排水管网缺陷自动识别方法，模型命名为RFCBAM-CGA-RTDETR。该模型在RT-DETR的基础上，首先采用感受野卷积注意力机制强化网络在主干部分的特征学习能力，然后采用级联分组注意力机制减少网络在编码过程中的计算量并保持对深层特征的提取能力，最后采用inner-MPDIoU算法优化预测框回归收敛过程。采用该模型对14种管网缺陷进行检测，并与其他目标检测模型进行对比，结果显示，该模型可以精准识别各类排水管网缺陷，平均精度mAP50值相较于RT-DETR和YOLOv8分别提高8.6%和12%，并且各类评价指标均优于其他模型，证明了该模型在排水管网缺陷识别上的有效性和泛化性。

Abstract:: Given the significant workload and time-consuming nature associated with traditional manual detection and identification of drainage network defects, this paper proposed an automated identification method for drainage network defects utilizing convolutional neural network (CNN) in conjunction with Transformer. The proposed model is designated as RFCBAM-CGA-RTDETR. Based on RT-DETR, the model initially employed the receptive field convolutional block attention module to enhance the feature extraction capability of the network within the backbone. Subsequently, it utilized the cascaded grouped attention module to mitigate computational complexity during the encoding process while preserving the capacity to extract deep features. Finally, the inner-MPDIoU algorithm was applied to optimize the convergence of the predicted bounding box regression. The model was utilized to detect 14 types of drainage network defects. Compared with other target detection models, this model could accurately identify various drainage network defects. Specifically, the average precision (mAP50) was 8.6% and 12% higher than that of RT-DETR and YOLOv8, respectively. All evaluation metrics were superior to those of other models. These findings confirm the effectiveness and generalizability of the model in identifying drainage network defects.

更新日期/Last Update: 2025-03-01