Self-Supervised Learning for Bridge Crack Classification

This technical report investigates the application of self-supervised learning (SSL) for bridge crack classification within the context of structural health monitoring (SHM). It addresses the challenges associated with traditional supervised learning methods that require extensive labeled datasets for effective crack detection. The report presents a systematic empirical study comparing various SSL methods, including contrastive learning, self-distillation, and masked image modeling. The study evaluates the effectiveness of these methods in improving label efficiency while maintaining classification accuracy. Experimental results indicate that SSL-pretrained models can enhance label efficiency, achieving improved accuracy even with limited labeled data. The findings suggest that SSL can significantly reduce the cost and scalability barriers associated with automated bridge inspection systems. The report also establishes a benchmark for future research in data-efficient crack classification, providing insights into the practical applicability of modern SSL techniques in the field of infrastructure inspection.