RT-DETR-AMRS: Lightweight Real-Time Transformer Detection Framework for Small Remote Sensing Targets

Accurately detecting tiny, densely packed objects in remote-sensing imagery remains a formidablechallenge due to severe feature vanishing and complex background clutter. In this work, wepresent RT-DETR-AMRS, a lightweight re-design of the Real-Time Detection Transformer (RTDETR) speciffcally tailored for aerial small-object scenarios. Our approach integrates three remotesensing–oriented components: C3k2-MCAttn, a dual-path residual block equipped with Monte-Carloattention to enrich ffne-scale cues; C2f-CAMixer, a lightweight channel–spatial mixer that strengthens local feature fusion without increasing architectural complexity; and DySample, a learnable,resolution-aware up-sampler that preserves spatial ffdelity far better than naïve interpolation. Torigorously benchmark these upgrades, we introduce TinyRS, a challenging dataset of 20,000 imagescomprising four representative aerial classes captured under densely cluttered conditions. Extensiveexperiments demonstrate that RT-DETR-AMRS achieves 82.5% mAP50 on TinyRS, outperformingthe RT-DETR-R18 baseline by 2.5 pp and surpassing state-of-the-art YOLO variants. Crucially, ourmodel reduces parameters by 32% and FLOPs by 21%, while exhibiting superior generalization onpublic benchmarks such as DOTA-v2.0 and VisDrone2019, striking an optimal balance betweendetection accuracy and real-time efffciency.