Learning Method to Detect Anomalies in Sewer Pipe Images Using Object Detection

Abstract

Object detection has have been used in a wide range of fields, e.g., manufacturing, construction, and medicine. In this study, we investigated the possibility of applying an object detection method to the sewer pipe inspection task. Currently, many sewer pipes are aging, and there is a shortage of personnel and funding for maintenance work. To address this problem, a system is required to automatically detect sewer pipe anomalies. In a previous study, we classified 15 types of sewer pipe anomalies from a total of 58 types, and we trained using the YOLOv5 object detection method, resulting in an overall accuracy rate of 60%. However, higher accuracy is required for practical application. Thus, in this study, we optimized the YOLOv5 learning process by dividing the sewer pipe images into direct-view and side-view images and by the types of pipes, and we achieved an overall accuracy rate of 70.2% on images categorized according to the types of pipes.