A deep learning benchmark analysis of the publicly available WRc dataset for sewer defect classification

Deep learning has the potential to transform sewer pipe inspection by automating the process, which could improve efficiency and consistency. However, progress has been hampered by limited publicly available, wellannotated benchmark datasets for defect classification. To address this gap, we present a comprehensive analysis using the publicly available Water Research Centre (WRc) sewer image dataset. We evaluated several deep learning architectures (MobileNet-v2, Inception-ResNet-v2 and ResNet-18) across key performance metrics such as accuracy and F1-score, with Top-1 accuracies ranging from 61.54% to 71.61% and Top-3 accuracies ranging from 86.88% to 92.61%. This research contributes to a reproducible performance baseline, enabling rigorous comparison of different models and serves as a foundation for future research in developing AI-assisted inspection systems.