Crack Detection and Localization in Stone Floor Tiles using Vision Transformer approach

Cracks are the initial indicators of deterioration of any civil infrastructures. Structures are typically monitored manually by inspectors, which is time consuming, laborious, costly, and easily prone to human error. To address these limitations this paper aims to present a vision transformer-based stone floor tiles crack detection and localization approach. The proposed model is trained on a custom dataset acquired from various stone tiles under various illumination condition in United Arab Emirates. The dataset consists of 5800 images having resolution of 224×224 pixels. The evaluation metrics, testing accuracy, precision, recall, F1 score and computational are used to analyze the performance of the proposed model. The input patch size of the ViT model is varied to investigate its effect on the performance of the model. The experimental results shows that input patch size has significant on the performance of the models. The ViT model trained on the lowest patch size of 14×14 pixel achieved the highest the testing accuracy, precision, recall and F1 score of 0.8612, 0.8840, 0.8304 and 0.8564 respectively. The inference time of the ViT model for a single patch is 0.092 sec. The crack localization is performed by combining the proposed trained ViT model with the sliding window approach. Overall, the proposed model showed promising performance and can used in detection and localization of cracks in stone floor tiles structures.