SC-DeformableDETR: An End-to-End Model for Small Target Detection in Drones

The significant distance between drones and targets during flight operations results in insufficient resolution of captured images from drone perspectives to discern discriminative target features, leading to compromised identification accuracy. To address this limitation, we propose SC-DeformableDETR, an enhanced algorithm specifically designed for small target detection in drone-based scenarios. The proposed framework achieves performance improvements through three core innovations:UDAFFPN integration to strengthen multi-scale feature representation. SCBAM (Spatial-Channel Bottleneck Attention Module) implementation for refined feature extraction, enabling the model to capture subtle image characteristics. A novel composite loss function that addresses localization inaccuracies in bounding box regression, thereby improving detection precision. Experimental validation demonstrates that SC-DeformableDETR achieves 2.5% and 2.7% improvements in mean Average Precision AP50 over DeformableDETR on the VisDrone2019 and RSOD datasets, respectively. Comparative analyses further confirm its superior performance in drone-view small target detection against existing state-of-the-art models.