SC-RT-DETR: a real-time pavement distress detection model based on multi-scale feature enhancement

Abstract

With the continuous expansion of road networks and increasing traffic loads, efficient automated detection and localization of pavement distresses are essential for timely road maintenance and traffic safety. However, accurate real-time detection remains challenging because pavement distresses often exhibit weak structural cues and substantial variations in scale and morphology, while elongated cracks show pronounced directional continuity. To address these challenges, this study proposes a Star-Chain-based Real-Time Detection Transformer (SC-RT-DETR) for pavement distress detection. First, a Star-Chain backbone network integrates lightweight star operations with spatial calibration at the backbone stage outputs to strengthen crack responses, suppress background interference, and reduce computational overhead. Second, Smoothing Learned Two-Dimensional Positional Encoding applies axial linear-smoothing initialization to horizontally and vertically decoupled learnable positional embeddings, introducing an ordered spatial prior for representing elongated and direction-dependent pavement cracks. Finally, a Hierarchical Content-Aware Feature Fusion Module combines hierarchy-specific heterogeneous kernel configurations with content-aware feature reassembly in a bidirectional refinement pathway to improve the joint representation of local details and global semantic context across scales. On the Road Damage Detection 2022 dataset, SC-RT-DETR achieves 83.21% mean average precision (mAP) at an intersection over union (IoU) threshold of 0.5 (mAP@0.5) and 54.09% mAP averaged over IoU thresholds from 0.5 to 0.95 (mAP@0.5–0.95), outperforming RT-DETR-R18 while using 49.4% fewer parameters and 59.5% lower computational cost. The source code is publicly available at https://github.com/Xtbreak/SC-RT-DETR.