Title: Leveraging Optical Guidance for Neural Collapse in SAR Few-Shot Class-Incremental Learning

Abstract

Synthetic aperture radar (SAR) imagery poses distinct difficulties for few-shot class-incremental learning (FSCIL), primarily driven by significant data shortages and inherent SAR variability. The pronounced azimuth sensitivity of SAR systems amplifies intra-class diversity while blurring inter-class boundaries. Furthermore, the sequential nature of FSCIL updates exacerbates the issue of catastrophic forgetting, causing models to lose knowledge of previously acquired classes. Drawing inspiration from the phenomenon of neural collapse, we introduce a novel framework for SAR FSCIL that incorporates optical guidance. This approach extracts orthogonal feature subspaces from a data-abundant optical Automatic Target Recognition (ATR) dataset, utilizing them as geometric priors to steer SAR feature extraction. By projecting SAR features onto these orthogonal subspaces through principal angle constraints, the model effectively transfers discriminative structural information from the optical domain to the SAR domain.

Our methodology employs a projection loss alongside a classifier loss, both optimized within a frozen simplex-Equal Norm Tight Frame (ETF) geometry. This joint optimization fosters neural collapse by clustering features tightly around class centroids while maximizing the angles between different classes. We validated our technique using a benchmark that integrates an optical ATR dataset with a SAR ATR dataset containing 24 target classes. The experimental setup involved one base training session followed by seven incremental sessions. Our results demonstrate superior final accuracy compared to contemporary FSCIL techniques, such as NCFSCIL, offering a better balance between peak performance and resistance to performance degradation. Additionally, neural collapse metrics reveal enhanced intra-class compactness and inter-class separability, suggesting that the learned features align more closely with the ideal simplex-ETF geometry.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC