@Article{cmc.2026.083042,
AUTHOR = {Linyu Dong, Tao Li, Hao Li},
TITLE = {HiFreq-DETR: A Hierarchical Framework Synergizing High-Resolution Injection and Frequency-Aware Multi-Scale Interaction for Tiny Object Detection},
JOURNAL = {Computers, Materials \& Continua},
VOLUME = {},
YEAR = {},
NUMBER = {},
PAGES = {{pages}},
URL = {http://www.techscience.com/cmc/online/detail/27121},
ISSN = {1546-2226},
ABSTRACT = {While Transformer-based detectors excel in global modeling, their efficacy in unmanned aerial vehicle (UAV)-based tiny object detection is limited by information loss during aggressive downsampling and the lack of high-frequency structural cues. To bridge this gap, we propose HiFreq-DETR, a dedicated framework that optimizes the synergy between spatial fidelity and semantic discriminability. The core innovation lies in its hierarchical information preservation strategy, which employs a ResNeSt14d backbone coupled with an <mml:math id="mml-ieqn-1"><mml:msub><mml:mi>S</mml:mi><mml:mn>2</mml:mn></mml:msub></mml:math> spatial injection path to recover critical high-resolution structural anchors, and introduces a frequency-selective interaction module to decouple target saliency from background noise. Experimental results demonstrate the substantial value of our approach. On the VisDrone dataset, HiFreq-DETR significantly outperforms the baseline RT-DETR, achieving improvements of 3.9% in AP and 4.4% in <mml:math id="mml-ieqn-2"><mml:msub><mml:mtext>AP</mml:mtext><mml:mi>S</mml:mi></mml:msub></mml:math>, confirming its effectiveness for tiny object detection. Furthermore, an optimized lite variant is evaluated to challenge the limits of high-efficiency processing for resource-constrained scenarios, while superior gains on the HazyDet dataset validate the model’s structural robustness in adverse aerial environments. These findings establish HiFreq-DETR as a high-fidelity and versatile solution for complex remote sensing applications.},
DOI = {10.32604/cmc.2026.083042}
}