Improved RT-DETR model for pavement distress detection

Accurate and efficient pavement distress detection is essential for intelligent highway maintenance. To address the current challenges of insufficient accuracy and low efficiency in distress detection, this paper proposes an improved Real-Time Detection Transformer (RT-DETR) model. In the backbone, a RepFEMA module is designed by integrating RepConv-enhanced FasterNet blocks with an efficient multi-scale attention (EMA) module, improving spatial feature extraction while reducing computational redundancy. In the encoder, the DySSTF module is introduced, consisting of dynamic scale sequence feature fusion (DySSFF) and triple feature encoder (TFE), to alleviate small-target feature loss and improve multi-scale perception. Moreover, Conv3XCC3 adopts dual-path dynamic fusion to strengthen detection accuracy and detail representation. Experiments show that the improved RT-DETR model achieves 67.2% mAP@0.5:0.95 with only 17.2M Params, outperforming conventional RT-DETR and YOLO series models in multi-scale feature modelling and robustness under complex noisy environments.