Self-supervised learning approach for automatic sewer defect detection

Abstract

Automated sewer defect detection has advanced through deep learning, particularly supervised methods using CCTV images, but based on large annotated datasets. This study proposes a semi- supervised learning (SSL) approach to reduce the dependency on annotations. The method includes two stages: self-supervised pre-training on unlabelled images using SwAV (Swapping Assignments between multiple Views of the same Image), followed by fine-tuning on labelled images for multi-label image classification. Experiments on the Sewer-ML dataset show that both ImageNet-pre-trained models -supervised and SwAV- outperform models trained from scratch on 1.04 million images, achieving higher F1-scores with just 13k labelled samples. The proposed SSL approach achieves 64.22% precision, 66.06% recall, and a 65.13% F1 score, surpassing the fully supervised baseline. Additionally, scaling up the pre-training dataset further enhances performance. These findings underscore the importance of ImageNet initialization and highlight self-supervised learning as an accurate, scalable, and cost-effective alternative to supervised methods, particularly in data-scarce scenarios.