Symmetry Frequency-Aware Fourier Series Network for Aerial Small Object Detection

Aerial tiny objects naturally possess conjugate symmetry in the frequency domain, yet complex scenarios, heavy clutter, and frequent rotation/occlusion lead to severe detail loss, inaccurate contour modeling, and high false/miss rates in existing detectors. Current Fourier-based methods neither leverage object symmetry nor coordinate with Fourier analysis to jointly enhance contour representation and spatial-frequency feature learning, suffering from weak fusion, phase-sensitive coefficient regression, and poor discriminability. To fill this gap, we propose the Frequency-Aware Fourier Series Detection Network (FAFSDet), which explicitly exploits the inherent symmetry of tiny objects and their frequency-domain representations. Specifically, FAFC (Frequency-Aware Feature Fusion) employs conjugate-symmetry-guided dynamic low-pass filtering, similarity-based rearrangement, and adaptive high-frequency enhancement to recover degraded symmetric patterns. FSPRM (Fourier Series Profile Representation) utilizes the symmetric positive–negative frequency distribution to achieve compact parametric contour encoding and normalized centroid-shape description. FSDIM (Fourier Series Detection Inference) incorporates symmetric multi-scale branches, a rolling-optimization loss that eliminates phase interference while preserving coefficient-regression symmetry, and inverse Fourier transform for precise contour reconstruction and end-to-end detection. Extensive experiments on DOTA, AI-TOD, and UCAS-AOD demonstrate that our method achieves superior performance (mAP 82.18%, 51.2%, and 90.70%, respectively) and strong generalization, particularly in scenarios where symmetry is most severely compromised, confirming that exploiting these symmetry properties substantially boosts detection accuracy.