SPG-IAD: Scattering-Point-Guided and IoU-Aware Dynamic Pseudolabel Selection for Semi-Supervised SAR Ship Detection

This is the authors' abstract. We don't add key points for this paper.

Deep learning-based synthetic aperture radar (SAR) ship detection has achieved remarkable progress in recent years. However, the high cost of annotating SAR images severely limits the practical deployment of fully supervised detection methods. Pseudolabel-based semisupervised learning is a mainstream approach for alleviating the reliance on large-scale annotations, and it has achieved superior performance in general object detection. Nevertheless, when directly transferred to SAR ship detection, these methods face two critical challenges due to the unique imaging characteristics of SAR data: first, they struggle to balance pseudolabel precision and recall, as ship diversity and speckle noise make confidence scores unreliable; second, they cannot effectively filter out low-quality pseudolabels during student training, which limits further performance improvement. To address the issues above, this article proposes a novel semisupervised object detection framework, named SPG-IAD, specifically for SAR ship detection. The framework consists of two core components: the scattering-point-guided dual-criterion coarse filtering (SPG-DC) strategy and the Intersection over Union (IoU)-aware teacher–student dynamic fine selection (IA-TSD) strategy. Specifically, SPG-DC combines classification scores with the physical prior of ship scattering points to collaboratively filter detection outputs, effectively recalling low-confidence yet well-localized candidate predictions. IA-TSD introduces an IoU prediction branch to quantify localization quality and compares teacher–student localization quality using a scale-adaptive threshold. This strategy enables dynamic pseudolabel selection and suppresses the misleading effects of low-quality pseudolabels. Experiments on the HRSID and SSDD datasets demonstrate that under the semisupervised setting with 10% labeled data, the proposed method achieves a significant improvement in average precision over the fully supervised baseline, and outperforms existing state-of-the-art methods.