Answers are grounded in this paper and cite their sources.
Oriented object detection in synthetic aperture radar (SAR) imagery remains challenging because speckle noise, blurred target boundaries, strong clutter, and large-aspect-ratio objects often degrade localization accuracy and increase false alarms. Existing detectors frequently struggle to balance noise suppression, boundary preservation, and context modeling for reliable oriented localization. To address these issues, this article proposes an oriented SAR object detector based on spectral–spatial synergistic learning and context-aware edge enhancement. Specifically, a smoothed gradient enhancement module integrates differentiable Gaussian smoothing and Scharr gradient extraction to enhance blurred target boundaries while reducing spurious edge responses. A dual-domain denoising block jointly models spatial- and frequency-domain features through reparameterized spatial convolutions and spectral feature modulation, thereby suppressing clutter while preserving structural information. In addition, an existing large selective kernel block is incorporated to adaptively model long-range contextual dependencies for multiscale and large-aspect-ratio targets. Extensive experiments are conducted on RSAR, SSDD, and HRSID. On RSAR, the proposed method achieves 92.88% mAP<tex-math notation="LaTeX">₅₀</tex-math> and 64.70% mAP<tex-math notation="LaTeX">50:95</tex-math> with 32.94M parameters and 14.08 frames/s on the test set. On SSDD and HRSID, it further obtains 99.1%/81.0% and 92.0%/66.6% in terms of mAP<tex-math notation="LaTeX">₅₀</tex-math>/mAP<tex-math notation="LaTeX">50:95</tex-math>, respectively. Additional per-class analysis, ablation studies, hyperparameter sensitivity experiments, and failure-case visualizations demonstrate that the proposed design improves oriented localization robustness while revealing remaining limitations under extremely strong clutter, very small targets, and dense target distributions.
No takes yet. Share an insight, caveat, or question.
Zhan et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: