A transformer cascaded model for defect detection of sewer pipes based on confusion matrix

• Published in: Measurement science & technology Vol. 35; no. 11; p. 115410
• Main Authors: Yu, Zifeng, Li, Xianfeng, Sun, Lianpeng
• Format: Journal Article
• Language: English
• Published: 01.11.2024
• ISSN: 0957-0233; 1361-6501
• DOI: 10.1088/1361-6501/ad6f35

Urban sewerage systems are critical to urban infrastructure. However, they are often subject to defects that threaten their operational reliability and efficiency. Some different types of sewer defects often have similar features. Recently, a number of deep learning models have emerged to automatically identify sewer failures, but these models often fail to accurately distinguish between them. In this paper, we propose a cascaded model to address this problem. Our work is based on the confusion matrix, which is obtained from a baseline model. With the confusion matrix, we can group confusable defects for better processing. We then design a Transformer cascaded model consisting of two steps. The first step performs coarse-grained defect detection to predict either a specific type of defect, or a group of confusable defects. In the later case, we perform a fine-grained defect detection to further distinguish the specific type of defect with the sub-model specifically trained for that group. The experimental results show that this cascaded model achieves a significant performance improvement by improving the mean average precision from 0.767 to 0.818 with our sewer object detection dataset. This method paves the way for accurate detection of defects in sewer systems.