Crack45K: Integration of Vision Transformer with Tubularity Flow Field (TuFF) and Sliding-Window Approach for Crack-Segmentation in Pavement Structures

Recently, deep-learning (DL)-based crack-detection systems have proven to be the method of choice for image processing-based inspection systems. However, human-like generalization remains challenging, owing a wide variety factors such as crack type and size. Additionally, because their localized receptive fields, CNNs high false-detection rate perform poorly when attempting capture relevant areas an image. This study aims propose vision-transformer-based framework that treats data succession small patches, retrieve global contextual information (GCI) through self-attention (SA) methods, which addresses CNNs’ problem inductive biases, including locally constrained receptive-fields translation-invariance. The vision-transformer (ViT) classifier was tested enhance classification, localization, segmentation performance by blending with sliding-window tubularity-flow-field (TuFF) algorithm. Firstly, ViT trained on custom dataset consisting 45K images 224 × pixels resolution, achieved accuracy, precision, recall, F1 scores 0.960, 0.971, 0.950, respectively. Secondly, integrated (SW) approach, obtain crack-localization map from large images. SW-based then merged TuFF algorithm, acquire efficient crack-mapping suppressing unwanted regions in last step. robustness adaptability proposed integrated-architecture were new acquired under different conditions not utilized during training validation model. ViT-architecture evaluated compared various state-of-the-art (SOTA) approaches. experimental results show equipped algorithm can real-world performance.