Fine-Grained Pavement Crack Segmentation Using Dual-Attention Residual U-Net with ASPP and Boundary Supervision

Abstract

Purpose: Fine-grained pavement crack segmentation remains challenging because cracks generally occupy a small proportion of pavement images and exhibit thin, irregular, and low-contrast structures. These characteristics create severe pixel-level class imbalance and may reduce the stability of conventional U-Net predictions. This study aims to improve crack localization and segmentation consistency through an enhanced U-Net architecture. Methods: This study proposes a pavement-specific Dual-Attention Residual U-Net integrating compact ASPP, strip pooling, and auxiliary boundary-aware supervision for thin and elongated crack structures. The model is evaluated using Crack500, CrackForest, and DeepCrack images through comparison with a U-Net baseline, repeated evaluation runs of the proposed model, and threshold-sensitivity analysis. A supporting LOCO ablation study examines the effects of component removal using seed 42. Results: The proposed model achieved a Dice coefficient of 0.7967 and an IoU of 0.6621 in the seed-42 evaluation. Across seeds 42, 123, and 2024, the reported mean Dice was 0.7963 ± 0.0004 and the mean IoU was 0.6616 ± 0.0006, indicating low observed variation across the evaluated runs. The component ablation provides supporting evidence about the design choices within the complete architecture. Novelty: The architectural contribution is the pavement-specific integration of compact ASPP with strip pooling and auxiliary boundary-aware supervision within a residual dual-attention U-Net. This design combines multiscale context aggregation and elongated spatial context with feature refinement and supervision intended to preserve fine crack structures.