Unsupervised anomaly detection for tile spalling segmentation using synthetic outlier exposure and contrastive learning

Tile spalling poses significant threats to pedestrians on sidewalks. Recently, deep learning-based approaches have been developed for autonomous building assessments. However, training a supervised model typically requires a large labeled dataset, which is often unavailable in new domain tasks. Moreover, data acquisition and ground-truth labeling are costly. This paper presents an unsupervised framework for anomaly detection of tile spalling. The framework incorporates uncertainty estimation and contrastive learning by training a segmentation model on a source dataset containing known classes, excluding spalling. Spalling is subsequently identified as outlier pixels based on elevated uncertainty scores. Additionally, a synthetic pattern, dubbed “Spalling Craft”, is developed for outlier exposure to further enhance model performance. The proposed approach outperforms state-of-the-art baselines by approximately 18.4%, 46.6%, and 31.7% in AUC, AP, and FPR95 scores, respectively. Compared to supervised learning methods, the framework significantly improves data efficiency while achieving strong performance in tile spalling segmentation.