Reliable recommendations for CCTV sewer inspections through multi-label image classification

Abstract

Sewer infrastructure is crucial for public health and environmental protection. The maintenance of these sewerage networks, with millions of kilometers of pipe, heavily relies on efficient Closed-Circuit Television (CCTV) inspections for identifying and addressing these issues promptly. This paper investigates the potential of hierarchical multi-label image classification to provide reliable recommendations for sewer pipe defects to assist CCTV inspectors. It focuses on both primary defect types and their specific subcategories, as defined by the European standard EN 13508-2. Experiments were conducted on a dataset of 1.2 million annotated sewer inspection images. Surprisingly, the simplest approach of directly predicting the final defect categories outperformed more complex hierarchical methods. When compared against expert human annotators, the multi-label classification methods provided substantially more reliable recommendations. While opportunities remain to further improve performance, these results underscore the promising potential of these methods to assist human inspectors in the maintenance of wastewater infrastructures.