Decoupled Self-Supervised Subspace Classifier for Few-Shot Class-Incremental SAR Target Recognition

Synthetic aperture radar automatic target recognition (SAR ATR) has ushered in a new era dominated by deep-learning (DL) techniques. However, the DL-based recognition systems inevitably confront catastrophic forgetting for learned knowledge and overfitting for the new, once deployed in openly dynamic scenarios where targets of new classes continually appear with few-shot instances. For practical applications, a decoupled self-supervised subspace classifier with few-shot class-incremental learning (FSCIL) ability is proposed for prompt knowledge transferring and stable discrimination, w.r.t., intrinsic and domain-specific challenges of the FSCIL of SAR ATR. Specifically, observing the significant componentity and azimuth sensitivity of targets in SAR imagery, two self-supervised tasks powered by a scattering mixup module and a rotation-aware transformation module are designed to synthesize virtual samples and related labels to unleash the classifier's transferability to future categories while enhancing its discriminability to fine-grained scattering patterns. Once deployed, the model's parameters are frozen to decoupled with dynamic worlds for general knowledge extraction. At inference, a subspace classifier with class-aware target priors proposed by a max-coverage feature selection mechanism is formed for stable point-to-space discrimination. Extensive experiments on three FSCIL datasets built from SAR-AIRcraft-1.0, Self-owned, and MSTAR datasets, which cover various categories captured by airborne and spaceborne SAR payloads, show the state-of-the-art performance achieved by our method compared to numerous latest benchmarks.