DERNet detects small objects 6x more efficiently by shifting to spectral processing

Small object detection has long been bottlenecked by feature scarcity—tiny targets lose critical high-frequency details during spatial-domain processing. A new paper from Yuhan Rui and ten co-authors (arXiv:2606.23825) tackles this by shifting from spatial to spectral feature processing. Their proposed DERNet framework introduces a unified Decompose–Enhance–Reconstruct (DER) operator, implemented as three lightweight, plug-and-play modules: Wavelet-Difference Gate (WDG) for frequency decomposition, Log-Gabor Enhancer (LGE) for amplifying discriminative components, and Frequency-Driven Head (FDHead) for final localization. These modules systematically inject frequency-aware modulation into the backbone, neck, and head of any detector architecture (CNN or Transformer).

Extensive validation on four multi-domain benchmarks—VisDrone2019, UAVDT, TinyPerson, and DOTAv1—shows consistent gains. Most notably, the DERNet series outperforms YOLOv11 models at the same scale while using only 1/6 the parameters. This parameter efficiency comes from decoupling feature modeling from resolution reduction, enabling accurate localization without expensive upscaling or noise amplification. The work opens a new direction for efficient, frequency-guided small object detection in drone surveillance, autonomous navigation, and remote sensing.