An enhanced RT-DETR with frequency decoupling and orthogonal regularization for UAV infrared small target detection
TL;DR
This work proposes an enhanced real-time detection transformer tailored for drone-based infrared scenarios that leverages orthogonal regularization to eliminate redundancy and purify target representations from background clutter, and introduces a dynamic scaling factor to provide smoother gradients and accelerate localization convergence.
Abstract
Summary Infrared small target detection in unmanned aerial vehicle (UAV) imagery is pivotal for low-light surveillance. However, existing frameworks suffer from feature redundancy, suboptimal context modeling, and gradient vanishing during tiny target localization. To address these bottlenecks, we propose an enhanced real-time detection transformer tailored for drone-based infrared scenarios. First, the Ortho-Block module leverages orthogonal regularization to eliminate redundancy and purify target representations from background clutter. Second, the AIFI-HiLo module decouples scene semantics via high- and low-frequency attention within intra-scale interactions to capture dense targets. Furthermore, a dual-stream GLSA mechanism with lightweight deformable convolutions adapts to extreme scale variations. Finally, the NWD-SIoU loss introduces a dynamic scaling factor to provide smoother gradients and accelerate localization convergence. Experimental results on HIT-UAV (Harbin Institute of Technology Unmanned Aerial Vehicle dataset) show a 4.8% mAP50 improvement over the RT-DETR baseline, while evaluations on VisDrone2019 confirm robust cross-scene adaptability.