A lightweight frequency-domain enhanced framework for small object detection in remote sensing

Abstract

Remote sensing small-object detection remains challenging: weak target details are attenuated by repeated downsampling, sparse object responses are difficult to separate from background clutter, and objects vary widely in scale and aspect ratio. We propose FD-YOLO, a compact detector with three complementary components. The Feature Enhancement Cascaded Architecture (FECA) reorganizes the neck so that each detection scale is reached through paths of different depth, and implements its fusion nodes with a cascaded unit that keeps a shallow localization anchor together with every refinement state and re-weights them through an input-conditioned depth router. Size-Aware Frequency Convolution (SAFConv) derives a radial energy descriptor from the complex Fourier spectrum, predicts per-image gains over learnable Gaussian frequency bands, and modulates spectral amplitude while preserving phase, so that the emphasized bands follow the scale composition of the current scene instead of a fixed partition; a parallel position-dependent spatial branch retains localization cues. Multi-Scale Expansion Attention (MSEA) mixes multi-dilated local-attention responses along an explicit scale axis, adds horizontal and vertical strip attention for elongated structures, and applies input-adaptive multi-kernel refinement. On the SIMD dataset, FD-YOLO achieved an mAP@50 of 83.97%, which was 8.84% higher than the YOLOv11n baseline, and also achieved the highest APS among all the compared detectors. On the DIOR dataset, its mAP@50 reached 74.6%, which was 1.6% higher than the baseline.