Real-Time Pavement Crack Segmentation Network with Multi-Depthwise Attention

Pavement crack images often contain complex backgrounds and noise interference, presenting significant challenges for real-time segmentation. This study proposes a real-time pavement crack segmentation network with multidepthwise attention. The encoder-decoder architecture integrates two key innovations: a multi-scale depthwise attention block (MDAB) that combines depthwise separable convolutions (DSCs) with varying receptive fields and dual attention mechanisms for precise feature extraction, and a depthwise skip chain block (DSCB) that employs densely connected DSCs to recover crack edge details. Experimental results on the DeepCrack dataset demonstrate that the proposed model achieves superior intersection over union (IoU) and Dice coefficient compared to most existing methods, with only 0.77 million parameters and 1.36 G GFLOPs while maintaining real-time performance. The lightweight architecture is deployable on compact terminal devices, providing an effective solution for real-time pavement crack detection.