Research on Automatic Identification of Drainage Network Defects based on Transformer and CNN

Abstract In order to achieve the governance of urban water environment systems and ensure the normal operation of urban drainage systems, it is necessary to regularly inspect the drainage pipe network to detect and identify defects. Traditional methods of manually identifying defects in Close Circuit Television (CCTV) data and annotating them are labor-intensive and inefficient. To address this issue, a method for automatic identification of pipe network defects based on convolutional neural network (CNN) and transformer is proposed. The model is named RFCBAM-CGA-RTDETR and has been applied to detect 13 types of pipe network defects. Experimental results indicate an evaluation accuracy of 61.8%, demonstrating its effectiveness and reliability. This model outperforms RT-DETR and YOLOv8 models by 6.6% and 10%, respectively, providing a novel approach for drainage pipe network defect detection.