Small Target Detection Algorithm Based on Improved RT-DETR Model

Aiming to address the limitations of existing real-time detectors in handling low-resolution small objects , this paper proposes an enhanced RT-DETR model by integrating StarBlock modules and dynamic upsampling strategy. First, the StarBlock is introduced into the backbone network to replace conventional BasicBlocks. By combining reparameterized convolution and element-wise star-shaped multiplication operations, StarBlock achieves implicit high-dimensional feature mapping with reduced computational costs, enhancing fine-grained feature extraction for small targets. Second, the DySample dynamic upsampling technique is adopted to replace static interpolation in the feature fusion stage. DySample employs learnable offset prediction with spatial constraints, enabling content-aware feature alignment and improving boundary clarity for small objects. Experiments on the VisDrone dataset demonstrate that the improved model reduces parameters by 4.85% compared to the baseline RT-DETR-R18 while achieving [email protected] and [email protected]:0.95 improvements of 0.7% and 1.1%, respectively. Further evaluations on YOLOv5m and YOLOv8m reveal that the improved model outperforms these CNN-based detectors by 4.3–4.8% in [email protected] under complex scenarios. The proposed model effectively balances accuracy, lightweight design, and real-time inference capabilities, providing a practical solution for small object detection tasks with stringent computational constraints.