Stagewise Feature-Guided Imbalance Diffusion Model for Imbalanced Semi-Supervised SAR Target Recognition

Automatic Target Recognition (ATR) in Synthetic Aperture Radar (SAR) imagery plays a key role in various defense and civilian applications. While recent semi-supervised learning methods have advanced SAR ATR, they often struggle when training data is class-imbalanced-a common scenario in practice. To address this challenge, we propose a method that combines class-rebalancing with self-training. Our approach consists of two main parts: a sample synthesis module and a semi-supervised learning module. The synthesis module uses a specially designed diffusion model to generate realistic samples for minority classes, guided by both global features and scattering center features in a stage-wise manner. The semi-supervised module introduces a Bayesian-rebalancing energy score to improve pseudo-label screening, reducing bias toward majority classes. Together, these components help the model learn a more balanced and accurate classifier. Experiments on highly imbalanced SAR datasets show that our method significantly improves recognition performance and reduces class imbalance effects.