Self-training method for structural crack detection using image blending-based domain mixing and mutual learning

Abstract

Deep learning-based structural crack detection utilizing fully supervised methods requires laborious labeling of training data. Moreover, models trained on one dataset often experience significant performance drops when applied to others due to domain shifts prompted by diverse structures, materials, and environmental conditions. This paper addresses the issues by introducing a robust self-training domain adaptive segmentation (STDASeg) pipeline. STDASeg incorporates an image blending-based domain mixing module to minimize domain discrepancies. Additionally, STDASeg involves a two-stage self-training framework characterized by the mutual learning scheme between Convolutional Neural Networks and Transformers, effectively learning domain invariant features from the two domains. Comprehensive evaluations across three challenging cross-dataset crack detection scenarios highlight the superiority of STDASeg over traditional supervised training approaches and current state-of-the-art methods. These results confirm the stability of STDASeg, thus supporting more efficient infrastructure assessments.