Pavement anomaly detection based on transformer and self-supervised learning

Pavement anomaly detection can help reduce the pressure of data storage, transmission, labelling and processing. This paper describes a novel method based on transformer and self-supervised learning that assists in locating anomaly sections. Experimental results reveal that self-supervised learning can improve performance on a small dataset with unlabeled images. Transformer is proven to be applicable in the pavement distress detection field. The facial recognition-like framework we built can enhance the performance without training by putting new patches into the gallery. Removing similar patches does not affect the recognition results. The method is sufficiently efficient and miniaturized to support real-time work and can be applied directly to edge detection.