Unsupervised deep learning method for concrete and asphalt crack segmentation using vision transformer and probability thresholding

Crack detection is vital for maintenance of civil structures. Recently, deep learning-based semantic segmentation models have shown promise in accurately identifying cracks. However, these methods require laborious manual data labeling. To address this, an unsupervised learning-based crack segmentation method was proposed, using a self-supervised Vision Transformer (ViT) as a backbone network to learn crack patterns from unlabeled images. A diverse crack image dataset with various crack types and backgrounds was used to train the network without time-consuming labeling, and to test the model after constructing ground-truths. The model was optimized with unsupervised contrastive loss function parameters, and probability thresholding was applied to enhance detectability by eliminating low confidence pixels, reducing false positives. On 1,399 test images, the unsupervised model achieved a mean F1-score of 75.02% and a mean Intersection over Union (mIoU) of 63.01%, with mIoU improving to 66.14% after thresholding, which shows great detection performance of unsupervised model. The model's application to high-resolution real crack images using a sliding window technique further demonstrated its suitability for field use, offering an efficient solution for real-time structural monitoring. These findings highlight the potential of unsupervised deep learning for crack detection, significantly reducing the need for manual labeling while delivering strong performance.