Robust pavement crack segmentation network based on transformer and dual-branch decoder

The application of deep learning techniques for semantic segmentation of crack images has become a significant research direction in road maintenance and safety. Despite the extensive research in recent years on semantic segmentation algorithms based on convolutional neural networks, their relatively small actual receptive fields cannot effectively handle long and fine pavement cracks. In contrast, transformer-based models can effectively utilize contextual semantic information. Therefore, a robust pavement crack segmentation network, CSTF, is proposed based on the Swin Transformer encoder. Within CSTF, a feature pyramid pooling module is introduced to provide global priors, and a dual-branch decoder is designed to preserve and learn semantic information, enabling CSTF to handle large-scale images and wide-spanning cracks. The results demonstrate that CSTF achieved an mIoU of 0.813 and 22.97 FPS on the large-scale dataset constructed in this study, enabling highprecision real-time detection. Moreover, it exhibits robustness against common interfering patterns like striped patches or other disturbances found in pavement crack images.