Spatial–Frequency Domain Joint Learning With Shape Constraints for Fine-Grained Aircraft Detection in SAR Imagery

Fine-grained aircraft detection aims to detect aircraft and identify its subcategory, which is important for military reconnaissance and airport management. Compared with optical imagery, aircraft in Synthetic Aperture Radar (SAR) images exhibit high azimuth sensitivity and discrete scattering characteristics, leading to significant intra-class variance and topological fragmentation, which make fine-grained aircraft detection very challenging. Existing methods mainly rely on spatial domain feature processing and scattering keypoint supervision, which do not fully utilize frequency domain features that are particularly important for fine-grained detection. This paper proposes a novel dual domain feature learning architecture with shape constraints, SAR-SFNet, to enhance the fine-grained aircraft detection performance in SAR imagery. First, a Spatial-Frequency Domain Joint Learning (SFDJL) is proposed via integrating Fractional Gabor Transform (FrGT)'s localized, orientation-tuned responses with the Fourier's global contextual cues to enhance the saliency of aircraft under varied aspect angles. Second, a Class-Aware Shape Constraint (CASC) is designed by leveraging class-specific shape priors to mitigate intra-class variance and topological fragmentation. Extensive experiments on SAR-RADD and FAIR-CSAR datasets demonstrate that SAR-SFNet achieves a mean Average Precision (mAP) of 79.3% and 50.6%, outperforming state-of-the-art methods by 3.7% and 5.2%, respectively, while maintaining a competitive inference speed of 39.5 Frames Per Second (FPS). Furthermore, with a lightweight architecture of 7.8 M parameters and 15.3 G Floating Point Operations (FLOPs), the proposed method exhibits its potential for resource-constrained, real-time applications.