SOA-DETR: Small Object Attention with DINOv2 Backbone for Bird vs. Drone Detection Across Varying Dataset Scales

Abstract

Reliable discrimination between birds and unmanned aerial vehicles (UAVs) is critical for airspace safety, wildlife conservation, and airport security, yet existing studies typically evaluate only one or two model families on a single dataset, making it difficult to disentangle the contributions of architecture and data scale. This study proposes a novel transformer-attention hybrid (SOA-DETR) network for bird and UAV detections. The proposed SOA-DETR combines a DINOv2 self-supervised backbone with a novel Small Object Attention-enhanced Feature Pyramid Network. Additionally, this study proposes a novel lightweight edge-deployment detector (SSD-MobileNetV3+SOA) and presents a controlled benchmark of SOA-DETR network against commonly used ten object detection architectures spanning five paradigm families, namely one-stage (YOLOv7, YOLOv8, YOLOv10, YOLOv12, YOLO11), transformer based (RT-DETR, RF-DETR), two-stage (Faster R-CNN), zero-shot (GroundingDINO), lightweight (SSD-MobileNetV3+SOA), evaluated on three bird-vs-drone (BVD) datasets of increasing scale under identical training protocols. The proposed SOA-DETR framework helps to achieve the best accuracy-efficiency trade-off (mAP50, mean Average Precision at IoU 0.5: 0.799/0.988/0.999 at 108 framesper second (FPS) with 28.99 M parameters) while RF-DETR leads in raw accuracy (0.835/0.994/0.997) and SSD MobileNetV3+SOA provides the fastest edge-deployable solution (5.7 ms, 174 FPS, 2.34 M parameters). Moreover, eight ablation studies provide practical deployment guidelines covering augmentation strategy, backbone scaling, input resolution, cross-dataset generalization, and SOA-DETR component analysis.

Keywords: object detection, bird-drone classification, Attention mechanism, Vision Transformer, Small object detection, benchmark