Automated Sewer Pipe Defect Classification Using Inspection Robot Imagery and Deep Neural Network

Sewer systems play a critical role in transporting wastewater. Efficient maintenance requires accurate identification of sewer pipe defects, traditionally performed through manual interpretation of Closed-Circuit Television (CCTV) images from robot inspections. This manual process is time-consuming, labor-intensive, and susceptible to human error. While conventional computer vision methods (machine learning methods) have been explored, they often involve extensive image pre-processing and intricate feature extraction. In this study, we propose an automated approach for classifying sewer pipe defects in Iran using deep learning, specifically leveraging the ResNet50 architecture for defect classification. We applied this deep convolutional neural network (CNN) model to automatically detect three common defect types as root intrusions, deposits, and open joints within a dataset 4630 images extracted from CCTV inspection videos of sewer pipes. Our model achieved an average testing accuracy of 96%, highlighting its potential for automating the interpretation of sewer CCTV footage.