A Defect Detection Method of Drainage Pipe Based on Improved YOLOv5s

In response to the existing challenges associated with manual interpretation, low efficiency, high leakage, and misdetection rates in detecting defects in urban underground drainage pipes, this study presents a defect detection method of drainage pipe based on improved YOLOv5s. The proposed method improves the detection of large target defects and reduces the leakage detection rate by increasing a deep target detection layer. Additionally, the introduction of deformable convolutional networks (DCN) allows for more accurate feature extraction from targets with complex shapes. Furthermore, the loss function is improved by employing MPDIoU as the bounding box loss function, which not only accelerates the convergence speed of bounding boxes but also enhances target recognition accuracy. Experimental results demonstrate that the improved model surpasses the performance of the original YOLOv5s, exhibiting an improvement of 3.8% in accuracy, 1.9% in recall, and 2.1% in average precision. Additionally, the proposed method achieves an impressive inspection speed of up to 54.64 FPS (frames per second), enabling real-time and efficient drain defect detection. This method is highly practical as it provides technical support for the future deployment of CCTV pipeline robots.