Abstract: A general segmentation model for dam concrete surface cracks in images often faces a shortage of training data due to the high cost of manual annotation, resulting in insufficient accuracy in its results. This paper presents an automatic annotation and segmentation algorithm that integrates image feature extraction and deep learning techniques. The algorithm first adopts a strategy for combining binarization and edge detection to annotate unlabeled crack defects automatically, and constructs a large-scale dataset of 19,101 crack masks. Then, a hybrid model for combining Swin-Transformer and Unet (Swin-Unet) is designed by introducing the hierarchical attention mechanism of Swin-Transformer into the Unet architecture. Finally, the model is validated through experiments and result analyses on the self-constructed datasets. The results show this Swin-Unet model achieves the highest crack classification accuracy (100%) and a segmentation IoU of 93.1% or 7.5% improvement over the Unet segmentation model (85.6%). This indicates the introduction of the Swin-Transformer architecture enhances the model's capability of associating global and local features, significantly improving the crack defect segmentation accuracy. Besides, an analysis of the minimum enclosing rectangle of cracks reveals significant clustering in both the direction and shape distribution of cracks, deepening our understanding of the mechanisms of crack formation and useful for predicting crack propagation direction.

Key words: deep learning, crack segmentation, automatic annotation, Otsu binarization, edge detection, attention mechanism