Seeing the Unseen: Overcoming Representational Fragility in UAV Detection with Star-Enhanced Cross-Scale Alignment

Abstract With the global rise of the low-altitude economy, Unmanned Aerial Vehicle (UAV) applications are experiencing explosive growth across various sectors. However, small object detection in UAV imagery remains a critical bottleneck and urgent challenge. This is primarily revdue to the extreme representational fragility of targets (typically smaller than 32 × 32 pixels) and complex background interference. To address these challenges, this paper proposes DMS-DETR, a robust small-object detection method tailored for UAV scenarios, built upon the RT-DETR framework.Specifically, an SDNet backbone is designed to mitigate the irreversible loss of fine-grained details during downsampling. Within this network, the Space-to-Depth (S2D) rearrangement module achieves lossless spatial transformation to preserve critical geometric cues. Simultaneously, Star Operation is introduced at deep layers to explicitly encode high-order channel interactions, significantly enhancing semantic expressiveness without increasing channel overhead. Furthermore, the Directional Region Attention Module (DRAM) is proposed to combine directional modeling with regional statistics, thereby alleviating background interference and highlighting orientation-specific texture contours. To bridge the pronounced semantic gap between low-level spatial details and high-level semantic information , the Multi-Directional Feature Enhancement Module (MFEM) is designed to expand receptive fields and suppress environmental noise, ensuring robust multi-scale feature alignment.Extensive experiments on two typical UAV datasets verify the superiority of the proposed method. On the VisDrone dataset, DMS-DETR achieves a 3.6% increase in mAP while reducing the parameter count by 17%. On the AI-TOD dataset, mAP increases by 4.8% while maintaining excellent inference performance. These results demonstrate that DMS-DETR provides an effective accuracy-efficiency trade-off for UAV small-object detection, offering a promising dataset-validated algorithmic approach for UAV small-object detection.