BKD-CL: Balanced Knowledge Distillation-Contrastive Learning for Distribution-Unknown Generalized Category Discovery in SAR ATR

Abstract

Open-environment machine learning is crucial for category discovery in synthetic aperture radar automatic target recognition (SAR ATR). However, SAR ATR toward intelligent applications requires addressing not only open-world distributions but also data imbalance. In this letter, we first propose the distribution-unknown generalized category discovery (DUGCD) problem and introduce the balanced knowledge distillation-contrastive learning (BKD-CL) framework, which includes the frequency attention ViT (FAViT) module and a multilayer perceptron (MLP) projection head. Second, we optimize the loss function using both supervised and self-supervised contrastive learning methods to learn feature representations from labeled and unlabeled data. We also implement self-distillation and entropy regularization to facilitate knowledge training for a parameterized classifier aimed at classification learning. Finally, to tackle the issue of data imbalance, we introduce balanced knowledge distillation, which selectively transfers knowledge using weighted coefficients to address the poor recognition performance caused by imbalanced data distributions. Extensive experiments conducted on the MSTAR dataset demonstrate the superiority of our proposed method.