Pavement Distress Segmentation via Multi-scale Edge Attention Enhanced Network from Line-Scan Camera Images

To address challenges posed by complex pavement backgrounds, diverse distress morphologies, and blurred boundaries, this paper presents the Edge Attention Enhancement with Multi-Scale Feature Extraction Network (EAE-MFE-Net) for pixel-level segmentation of pavement distress in high-resolution line-scan images acquired by pavement imaging sensors. The network integrates multi-scale feature extraction as a foundation to capture hierarchical representations and support an edge attention enhancement module, which leverages high-frequency edge information to improve boundary delineation and segmentation accuracy. A decoder incorporating depthwise separable convolutions and multilayer perceptrons refines spatial details, while a deep supervision mechanism ensures consistent performance across scales. Experiments conducted on the Southeast University Pavement Distress (SEU-PD) dataset, comprising multi-scene and multi-category data, as well as on publicly available datasets, demonstrate that EAE-MFE-Net achieves 98.78% accuracy, 77.18% weighted F-measure, 67.28% mean Intersection over Union, and 78.97% Dice coefficient, outperforming existing methods. These results confirm the method’s effectiveness and robustness in complex pavement environments and its capability for precise segmentation of diverse pavement distress types using sensor-based measurements.