A Dual-Branch Network for Pavement Distress with Region-Level Detection and Pixel-Level Segmentation

Accurate perception of pavement distress is essential for intelligent road maintenance. Existing detection and segmentation methods are still constrained by model complexity, feature redundancy and inadequate multi-scale feature modelling, thereby hindering fine-grained pavement distress perception in complex scenarios. To address these issues, this paper proposes PavDS-Net (Pavement Detection and Segmentation Network), a YOLO11-based dual-branch framework for pavement distress detection and crack instance segmentation. In the detection branch, PavD-Net (Pavement Detection Network) introduces an Efficient Channel Attention (ECA) module to recalibrate high-level semantic features. A P2 detection layer is further incorporated to construct a four-scale detection structure, improving the detection of small distress regions. The segmentation branch, PavS-Net (Pavement Segmentation Network), replaces the Spatial Pyramid Pooling-Fast (SPPF) structure with a progressive Squeeze–Spatial–Window (SSW) module for crack instance segmentation. Moreover, Binary Cross-Entropy (BCE) Mask Loss, Dice + BCE joint loss and Boundary Loss are incorporated to optimise supervision constraints. Experiments were conducted on multiple datasets for detection and segmentation tasks. For the detection task, PavD-Net achieves 88.7% Precision and 61.2% mAP50-95, improving over the baseline by 5.2% and 2.2%, respectively. For the segmentation task, PavS-Net achieves 93.4% Precision and 90.9% mAP50, improving over the baseline by 7.9% and 12.0%, respectively. Ablation and comparative experiments demonstrate that PavDS-Net improves region-level localisation and fine-grained crack instance segmentation within the proposed dual-branch framework, achieving a favourable balance between accuracy and deployment efficiency.