Cross-Attention-Guided Feature Alignment Network for Road Crack Detection

Road crack detection is one of the important issues in field traffic safety and urban planning. Currently, road damage varies type scale, often has different sizes depths, making task more challenging. To address this problem, we propose a Cross-Attention-guided Feature Alignment Network (CAFANet) for extracting integrating multi-scale features damage. Firstly, use dual-branch visual encoder model with same structure but patch (one large small patch) to extract multi-level features. We utilize Cross-Layer Interaction (CLI) module establish interaction between corresponding layers two branches, combining their unique feature extraction capability contextual understanding. Secondly, employ Block (FAB) align from levels or branches terms semantics spatial aspects, which significantly improves CAFANet’s perception regions, reduces background interference, achieves precise segmentation Finally, adopt multi-layer convolutional heads obtain high-resolution maps. validate effectiveness our approach, conduct experiments on public CRACK500 dataset compare it other mainstream methods. Experimental results demonstrate that CAFANet excellent performance tasks, exhibits significant improvements F1 score accuracy, an 73.22% accuracy 96.78%.