ENHF-YOLO: Enhanced High-Frequency Domain Feature Extraction of Small Targets in Remote Sensing

Remote sensing object detection is critical for disaster monitoring and related applications, but edge devices impose stringent requirements on model lightweighting. Existing lightweight detectors, however, often show limited sensitivity to small objects and rarely exploit frequency-domain information that is crucial for preserving high-frequency spatial cues. To address this gap, we propose EnHF-YOLO, a lightweight framework built on the YOLOv11 architecture that enhances high-frequency feature learning. Frequency Domain Dynamic Convolution (FDConv) is introduced into the backbone to construct frequency-aware kernels, reinforcing fine-grained feature representation. In addition, we design the lightweight CKS module with a dual-unit configuration that adaptively modulates receptive fields through spatial–spectral weighting while suppressing background interference. We further develop a bidirectional four-layer path aggregation network that strengthens shallow feature fusion and multi-scale information flow. Extensive experiments on VisDrone2019 and five additional benchmarks validate the effectiveness of EnHF-YOLO, achieving 93.9% mAP@0.5 (+8.8%) on RSOD.