SFF

Real-time object detection in remote sensing (RS) images presents significant challenges due to the small pixel size of objects, inconsistent contrast, and highly cluttered backgrounds. Recent methods relied mainly on convolutional neural networks (CNN) or multihead self-attention (MHSA), but often struggled to balance detection speed and performance. FNet introduced an innovative mechanism that replaces multi-head self-attention with Fast Fourier Transform (FFT) to accelerate Transformer structures. Based on this, we propose Selective Fourier Fusion (SFF) for remote sensing object detection. SFF can serve as an alternative to MHSA in object detection models. Experiments conducted on models like YOLOv10 and RT-DETR with public datasets VisDrone DET-2019 demonstrate that SFF significantly improves inference speed while enhancing the accuracy of the original models.