Space-Filling Curve-Based Fractional Fourier Feature Transformation for Oriented Object Detection in Optical and SAR Remote Sensing Imagery

Abstract

In recent years, oriented object detection in remote sensing images has received increasing attention due to its wide range of application scenarios. However, in complex scenes, object features are often hard to distinguish, making accurate object detection in this case still a challenging task. In this paper, we propose a space-filling curve-based fractional fourier feature transformation detector (SF³Det) to achieve high-precision oriented object detection in remote sensing images. First, we propose a fractional fourier transform based multi-domain features extraction module. By utilizing amplitude and phase information, the module can provide discriminative and enhanced features. Then, we adopt a locality-preserving feature serialization method for regional proposals in two-stage detectors. Specifically, object features are unfolded along a space-filling curve to obtain feature sequences with local correlation, which helps optimize model convergence and provides efficient local semantic cues for subsequent prediction. Finally, we apply a joint consistency-aware loss to dynamically adjust the optimization of classification and regression losses, consistently improving the performance of both tasks. Extensive experiments on optical remote sensing datasets and synthetic aperture radar image datasets demonstrate the superiority of SF3Det. Our method achieves mAPs of 90.6%, 79.92%, 89.90%, and 80.40% on the HRSC2016, DOTA, SSDD, and HRSID datasets, respectively.