ISTD-CrackNet: Hybrid CNN-transformer models focusing on fine-grained segmentation of multi-scale pavement cracks

Pavement crack detection is essential for the automated evaluation of pavement damage. In this context, we introduce ISTD-CrackNet, a semantic segmentation model based on a hierarchical Transformer architecture. This model features a multi-angle strip convolution module, a dynamic upsampling module, and a multi-scale transposed convolution segmentation head. The multi-angle strip convolution module is designed to establish long-range dependencies in four directions, ensuring the continuity of crack segmentation. An attention-guided dynamic upsampling module is employed to enhance the recognition accuracy of small cracks. Additionally, the multi-scale transposed convolutional segmentation head integrates shallow positional information with deeper categorical details to improve the fine-grained performance of crack edge segmentation. Compared to mainstream segmentation models, ISTD-CrackNet effectively addresses issues of segmentation discontinuities, low multi-scale accuracy, and boundary blurring. Experiments conducted on 5 publicly available datasets demonstrate its excellent generalization ability and robustness, highlighting its significant potential for intelligent pavement evaluation applications.