RSO-DETR: Multiscale Feature Integration and Spatial Pixel Calibration for Robust Small Object Detection in UAV Images

Abstract

Small object detection in uncrewed aerial vehicle (UAV) imagery remains challenging because weak target evidence is easily degraded by repeated downsampling and semantic-spatial misalignment during multiscale fusion. This article presents RSO-DETR, an end-to-end detector built on RT-DETR for UAV small-object detection. The efficient multiscale feature integration backbone combines channel splitting with progressive depthwise-separable receptive-field expansion to enhance feature reuse while reducing parameter redundancy. The spatial pixel interaction neck introduces local cross-level correspondence modeling and deformable geometric rectification to align semantic and spatial features before hierarchical fusion. In addition, HLFA-AIFI adapts the HiLo attention principle to RT-DETR’s AIFI stage, using nonoverlapping local-window attention to preserve fine structures and average-pooled global attention to capture coarse context. On VisDrone2019, RSO-DETR achieves 41.4% mAP50 and 24.6% mAP50:95, outperforming RT-DETR-R18 by 3.5 and 2.7 percentage points while reducing parameters by 47.5% to 10.44 M. Evaluations on HIT-UAV and AI-TOD further demonstrate applicability across thermal-infrared and small object aerial settings.