Accurate and efficient detection of road pavement distress is essential for AI-driven maintenance systems in modern infrastructure management. Pavement distress exhibits significant size variability while maintaining similar local characteristics, necessitating detection models that can effectively capture both local and global features. However, existing approaches relying solely on Convolutional Neural Networks (CNNs) or Transformers often result in either global feature loss or inefficient feature extraction processes. To address these limitations, we propose PD-DETR, a novel pavement distress detection model that integrates the advantages of real-time object detection Transformer (RT-DETR) architecture. Our method incorporates an optimized NextViT backbone enhanced with Depthwise Separable Convolutions (DSConv) and a fine-tuned MLP ratio, enabling superior capture of both global and local characteristics of pavement distress while maintaining model efficiency. Furthermore, we introduce an innovative AttentionGate mechanism within the hybrid encoder, which facilitates more effective multi-scale feature fusion through channel weighting after feature concatenation. Extensive experiments on the challenging RDD2022 dataset demonstrate the superior performance of PD-DETR, achieving a mean Average Precision (mAP) of 64.27\% with a compact parameter size of 30.0M. The proposed model outperforms state-of-the-art baseline models, including YOLOv5s (57.85\%) and YOLOv9s (60.12\%). This work represents a significant advancement in Transformer-based architectures for pavement distress detection, particularly in complex real-world scenarios, offering both improved accuracy and computational efficiency for practical road maintenance applications.

Exception Detection,Transformer,Object detection