Skip to content
Open access

S$^{2}$Det: A Dual Scale-Aware Detector for Small UAV Targets

2026 · IEEE Transactions on Aerospace and Electronic Systems · Vol 62, pp. 19443-19461 · 0 citations · 46 references

Abstract

Reliable detection of small unmanned aerial vehicles (UAVs) is fundamental to vision-based counter-UAV systems, yet remains challenging because targets at operational distances often occupy only tens of pixels. We identify two structural bottlenecks in modern detection pipelines that disproportionately degrade small-object performance: semantic dilution, where uniform feature fusion overwhelms fine-grained backbone details with semantically rich neck features, and score bias, where classification-based candidate selection systematically underrepresents small objects due to their inherently weaker responses. We propose S$^{2}$Det, a scale-aware detection framework comprising two complementary modules. Scale-adaptive feature fusion generates pixelwise level-conditioned fusion weights through dual-branch attention and learnable scale priors, preserving discriminative details during the fusion process before irreversible loss occurs. Scale-aware query detection head incorporates predicted object scale and pyramid level information into a learned score calibration function, ensuring equitable candidate selection regardless of object size. Experiments on four visible-spectrum UAV detection benchmarks demonstrate consistent improvements across all evaluation metrics over 18 compared methods, and a further evaluation on a thermal infrared benchmark confirms that the gains persist under a change of imaging modality. The framework runs at 46 frames/s on a desktop GPU and 21 frames/s on a Jetson Orin Nano under INT8 quantization.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.