Memory Bank Assisted Pseudo Label Refinement for Semi-Supervised Object Detection in SAR Images

Synthetic Aperture Radar (SAR) object detection plays an important role in military reconnaissance, disaster assessment, and environmental monitoring. Conventional methods rely heavily on fully supervised learning, which requires costly annotations. Semi-supervised approaches can reduce annotation costs, but their performance in SAR imagery is often limited by speckle noise and class imbalance, which degrade pseudolabel quality, particularly in multi-class scenarios. To address these challenges, we propose MB-SSOD, a semi-supervised object detection framework with two key components. First, a Memory Bank aggregates historical features to stabilize pseudo-label confidence estimation and mitigate noise effects. Second, a Class-wise Adaptive Local Threshold (CALT) dynamically adjusts selection thresholds according to each class's learning dynamics, lowering thresholds for minority classes while maintaining stricter criteria for majority classes. Extensive experiments on SARDet-100K, SAR-AIRcraft-1.0, and MSAR-1.0 demonstrate that MB-SSOD achieves state-of-the-art performance with minimal annotations and provides significant improvements under class-imbalanced conditions.