Detection of Pavement Cracks by Deep Learning Models of Transformer and UNet

Surface fractures are a significant problem in engineering structures like buildings and roads. Therefore, detecting such cracks is essential for assessing damage and maintaining these structures. The emergence of deep learning techniques has significantly enhanced the capability to detect surface cracks. Convolutional Neural Networks (CNNs) are predominantly used for this task, but recently introduced transformer architectures could offer improvements. In this research, we developed software that integrates nine advanced models and various activation functions to evaluate their effectiveness in detecting pavement cracks. The evaluation is based on the models' accuracy, complexity, and stability. We generated 711 images, each $224\times 224$ pixels, with crack labels, selected the most effective loss function, and compared the performance metrics of both validation and test datasets. We also examined the data details and evaluated the segmentation results for each model. Our results show that transformer-based models are more likely to converge during training and achieve higher accuracy, albeit with increased memory usage and reduced processing speed. Considering both accuracy and efficiency, SwinUNet outperforms the other two transformers and emerges as the superior choice among the nine evaluated models. We further confirm our conclusions with two public crack datasets. These findings shed light on the capabilities of various deep-learning models for surface crack detection and offer guidance for future applications in the field.