NSFA AQE CenterOBB detects oriented SAR ships using frequency domain noise suppression and angle quality estimation

Ship detection in synthetic aperture radar (SAR) imagery underpins maritime surveillance, traffic monitoring, and safety-of-navigation applications, yet high-quality detection in real-world SAR remains difficult. Multiplicative speckle and sensor-dependent noise degrade contrast, sea clutter and strong land or shore backscatter confuse inshore scenes, and elongated ships need reliable orientation estimates that a horizontal bounding box simply cannot give. We present NSFA–AQE CenterOBB, an end-to-end, anchor-free, one-stage detector built to address these three problems within a single oriented-detection pipeline. Its backbone is lightweight and includes Noise Suppression and Feature Amplification (NSFA) blocks that modulate the spectrum of intermediate feature maps, designed to attenuate clutter while preserving ship-discriminative structure; a Sobel-derived edge enhancement module adds boundary cues intended to support localization under weak contrast. Building on a CenterNet-style objects-as-points formulation, a probabilistic angle head predicts both an orientation mean and an uncertainty estimate, trained with a periodic negative log-likelihood loss that respects the π-periodic symmetry of ship orientation. At inference, this uncertainty drives an Angle Quality Estimation (AQE) mechanism that recalibrates detection scores during decoding and oriented non-maximum suppression. A symmetric Kullback–Leibler divergence regularizer maps predicted and ground-truth oriented boxes onto two-dimensional Gaussians and enforces geometric consistency across center, size, and orientation. On SSDD + , the rotated-box relabeling of the SAR Ship Detection Dataset, the detector performs competitively across both offshore and challenging inshore scenes, and our cumulative ablation shows that adding a geometry-consistency term to the training objective gives the numerically largest single step in mean average precision (part of which offsets a preceding drop at the probabilistic-angle stage); a controlled comparison across four geometry-consistency loss formulations shows the symmetric KLD variant we adopt achieves the highest mAP50, although the differences among the geometry losses are within run-to-run variation, and it is not the best in mAP75 or angular accuracy. The predicted angle uncertainty is, however, only weakly informative at the evaluated checkpoint, so AQE does not improve AP in our controlled comparison.