DPLR-DETR: Small Object Detection Model for UAV Aerial Imagery

This is the authors' abstract. We don't add key points for this paper.

Small object detection in UAV aerial imagery presents significant challenges due to limited pixel coverage and complex backgrounds. This paper introduces DPLR-DETR (Dynamic Position Large Receptive Field DETR), a novel architecture that enhances the RT-DETR framework to improve detection accuracy and computational efficiency in UAV applications. DPLR-DETR incorporates three key innovations: (1) an Enhanced Shifted Residual Block (ESRB) backbone for efficient feature extraction, (2) a Gated Channel Attention Multiscale Fusion module that implemented a “deep processing for small targets—lightweight preservation for backgrounds” strategy, and (3) a Dynamic Position Bias Attention (DPBA) module that enhances spatial modeling capabilities. Experimental results on the VisDrone2019 dataset demonstrate that DPLR-DETR significantly improves small object detection performance while maintaining computational efficiency suitable for resource-constrained scenarios.