Efficient UAV Detection via Coordinate Pyramid Recalibration and Feature Selection-Driven Pyramid Fusion

Abstract

UAV detection in surveillance is difficult due to small target size, cluttered backgrounds, and inter-class confusion with birds. We propose <bold>CPCR-DEIM</bold>, built on the DEIM framework, with two lightweight modules. First, a Coordinate Pyramid Channel-guided Recalibration (CPCR) block is inserted at an intermediate backbone stage to improve target-background separability via multi-scale directional reweighting before further spatial downsampling. Second, a Feature Fusion Enhancement Pyramid Network (FFEPN) replaces standard neck fusion nodes with FFE operators that use deep-layer features to compute per-channel weights, reducing the influence of clutter-dominated channels during top-down aggregation. Experiments on Air-UAV (visible light) and Anti-UAV (infrared) show that CPCR-DEIM achieves <inline-formula><tex-math notation="LaTeX">$80.4\pm 0.2\%$</tex-math></inline-formula> <inline-formula><tex-math notation="LaTeX">$\text{mAP}_{50}$</tex-math></inline-formula> and <inline-formula><tex-math notation="LaTeX">$44.9\pm 0.2\%$</tex-math></inline-formula> <inline-formula><tex-math notation="LaTeX">$\text{mAP}_{50:95}$</tex-math></inline-formula> at 25.1 GFLOPs and 10.2 M parameters, with 5.99 ms latency (167.07 FPS) on the measured GPU platform. Among the compared YOLO-family and transformer-based detectors, CPCR-DEIM gives the best strict-IoU localization accuracy on Air-UAV and the best accuracy on Anti-UAV.