Dual Framework for Fine Pavement Crack Detection Using Hybrid Attention

Traditional semantic segmentation methods for pavement crack detection face significant challenges in handling irregular and thin cracks, particularly when distinguishing them from similar background features. To address these issues, this study propose a novel dual framework that leverages hybrid attention mechanisms by integrating convolutional neural networks (CNNs) and Transformer modules through high-resolution and low-resolution branches. The high-resolution branch incorporates a Detail Perception Module (DPM) to preserve fine-grained details of cracks, while the low-resolution branch employs a Global Context Module (GCM) to enhance the differentiation of cracks from the background. Additionally, a Multi-Scale Fusion Module (MSFM) is introduced to effectively integrate local and global information from both branches. Furthermore, a skeleton-guided task is utilized to augment the morphological representation of cracks. Experimental results demonstrate that our proposed framework achieves more continuous and accurate segmentation of pavement cracks, with significantly improved robustness in complex backgrounds compared to existing methods.