Small object detection in UAV imagery holds significant application value in military reconnaissance

disaster relief

and other fields. However

existing algorithms commonly face challenges such as difficulties in lightweight deployment

insufficient multi-scale feature fusion

and low target matching accuracy in dynamic scenes. To address these challenges

this study proposes UAV-DETR

a lightweight UAV object detection model based on the DETR architecture. Its core innovations include: employing a lightweight feature extraction network through deep separable convolutions and channel attention mechanisms

reducing parameter count by 42% while preserving feature representation capabilities; designing a cross-scale feature pyramid fusion module that achieves effective multi-resolution feature aggregation via adaptive weight allocation. Experiments will be conducted on the VisDrone and UAVDT public datasets. By comparing against mainstream algorithms such as Faster R-CNN and YOLOv5

verifying the proposed model's comprehensive advantages in detection accuracy (mAP)

inference speed (FPS)

and model size (MB). It is anticipated to achieve an 8%–12% improvement in mAP for small object detection and a model compression ratio exceeding 3.5 times

providing an efficient solution for real-time UAV detection tasks.