为了保证城市污水管道的安全、正常运行，需要对管道进行缺陷检测和及时维护。传统的检测方法依赖人工判读，过于依赖从业经验，易出现漏判、误判、错判，易导致管道缺陷不能及时修复，造成严重的安全隐患和经济损失。本文以实际城市污水管道缺陷构建管道缺陷数据库，研究建立了基于YOLOv7的城市污水管道缺陷检测算法，缺陷检测精度高且运行速度较快，具有较好的应用前景。首先，利用CCTV管道机器人在不同管况下采集的含有16类管道缺陷的图像，并结合部分企业收集的数据集，共构建城市污水管道缺陷数据库。采用moasic等数据增强方法扩增数据量，避免缺陷检测算法发生过拟合现象。其次，研究基于YOLO和数据增强方法的管道缺陷检测方法，用于检测管道中小样本缺陷。然后比较YOLOv7、YOLOv5s、YOLOv3-spp和Faster R-CNN网络的检测性能。通过YOLOv7网络与各种数据增强方法相结合，建立了一个DA-YOLOv7模型。DA-YOLOv7模型在复杂场景中具有最佳的检测性能和较强的泛化能力，每幅图像的mAP、精度、召回率、F1得分和平均检测时间分别为96.03%、94.76%、95.54%、95.15%和0.025 s。因此，YOLOv7与数据增强相结合，可用于检测城市污水管道缺陷。这项研究为复杂条件下城市污水管道缺陷的检测提供了理论参考。最后，提出了一个改进的YOLOv7网络模型，该模型引入了小型缺陷检测层、轻量级卷积和CBAM注意力机制，以实现多尺度特征提取和融合，并减少模型的参数量。该模型的性能在城市污水管道缺陷的测试集中进行了测试，平均准确率（mAP@0.5）达到97.29%，平均预测时间为69.38 ms，与原YOLOv7相比，参数数量和计算成本分别减少了11.21 M和28.71 G。同时，Pipe-YOLOv7模型的实验结果显示，该模型的性能达到了目前最先进的水平。
To ensure the safety and normal operation of urban sewer pipes, it is necessary to detect defects and perform timely maintenance. Traditional detection methods rely on manual interpretation, which is heavily dependent on practitioner experience and can result in missed, false, or misjudged defects. This may lead to unaddressed pipe defects, causing severe safety hazards and economic losses. In this study, we construct a sewer pipe defect database using actual urban sewer pipe defect data and develop a high-precision, fast-operating urban sewer pipe defect detection algorithm based on YOLOv7, demonstrating promising application prospects.First, we collect images containing 16 classes of pipe defects taken by CCTV pipe robots under different conditions and combine them with datasets gathered from some enterprises to construct an urban sewer pipe defect database. We use data augmentation techniques such as mosaic to increase the data volume and avoid overfitting in the defect detection algorithm.Next, we investigate pipe defect detection methods based on YOLO and data augmentation techniques for detecting small-sample defects in pipes. We then compare the detection performance of YOLOv7, YOLOv5s, YOLOv3-spp, and Faster R-CNN networks. By combining YOLOv7 with various data augmentation techniques, we establish a DA-YOLOv7 model. The DA-YOLOv7 model achieves the best detection performance and strong generalization ability in complex scenarios, with mAP, precision, recall, F1-score, and average detection time per image at 96.03%, 94.76%, 95.54%, 95.15%, and 0.025 seconds, respectively. Therefore, YOLOv7 combined with data augmentation can be used for detecting urban sewer pipe defects, providing a theoretical reference for detecting sewer pipe defects under complex conditions.Finally, we propose an improved YOLOv7 network model that introduces a small defect detection layer, lightweight convolution, and CBAM attention mechanism to achieve multi-scale feature extraction and fusion while reducing the number of model parameters. The model‘s performance is tested on a test set of urban sewer pipe defects, with an average precision (mAP@0.5) of 97.29% and an average prediction time of 69.38 ms, reducing parameter count and computational cost by 11.21 M and 28.71 G compared to the original YOLOv7, respectively. At the same time, experimental results of the Pipe-YOLOv7 model show that the model‘s performance reaches the current state-of-the-art level.
