FEMT-YOLO: Frequency-Enhanced Multi-Scale Network for Small Object Detection in Aerial Images

• First to identify and address three fundamental limitations in aerial small object detection through targeted innovations. • Novel frequency-enhanced pyramid network combining spatial and frequency domain processing for superior small object feature preservation. • Innovative task-aligned dynamic head with feature decomposition and geometric adaptation for optimized classification localization balance. Small object detection in aerial images faces significant challenges: objects typically occupy <0.3% of image area, are embedded in complex backgrounds, and exhibit extreme scale variations. Existing methods have three key limitations: fixed receptive field constraints, feature degradation during propagation, and optimization conflicts in coupled detection heads. We propose FEMT-YOLO (Frequency-Enhanced Multi-Scale Network You Only Look Once), a novel architecture that systematically addresses these challenges through progressive feature refinement across three synergistic stages: extraction, aggregation, and prediction. The Multi-Scale Edge Information Enhancement Block (MSEIBlock) constructs adaptive feature pyramids with explicit boundary perception, addressing fixed receptive field limitations in the extraction stage. The Small Object Detection with Frequency-Enhanced Pyramid Network (SOFE-PNet) performs spatial-frequency dual-path processing, capturing long-range dependencies via frequency domain global modeling while suppressing background noise in the aggregation stage. The Task-Aligned Dynamic Detection Head (TAD-Head) resolves optimization conflicts through task decomposition and dynamic geometric alignment in the prediction stage. Extensive experiments on three challenging datasets validate our approach. On VisDrone-2019, the method achieves 38.20% mAP@50 and 22.90% mAP@50:95, improving 7.20% and 4.90% over YOLOv11 baseline. On RSOD and NWPU VHR-10, mAP@50 reaches 93.60% and 88.50%, respectively, with consistent improvements across all metrics. Cross-dataset evaluation demonstrates an average mAP@50:95 improvement of 3.97%, validating strong generalization capability and establishing competitive performance against state-of-the-art methods for UAV-based vision applications.