Reproduce “Few-Shot Class-Incremental Learning For Efficient SAR Automatic Target Recognition” on OpenResearch

About this paper

George Karantaidis, Athanasios Pantsios, Ioannis Kompatsiaris, Symeon Papadopoulos

A framework named DILHyFS was developed by researchers at the Information Technologies Institute, Centre for Research & Technology Hellas, for Few-Shot Class-Incremental Learning in SAR Automatic Target Recognition, integrating a dual-network architecture, cross-attention, and a hybrid loss function. This framework achieved an 84.06% average incremental accuracy on the MSTAR dataset and demonstrated a 68.64% reduction in catastrophic forgetting compared to prior methods.

Key insights

• The combination of domain-specific feature extraction (ResNet-18) and frequency-domain global context capture (GFNet) yields more robust and discriminative features for SAR-ATR.
• A hybrid loss function incorporating both focal loss and center loss effectively improves both inter-class distinction and intra-class compactness, which are crucial for few-shot and incremental learning.
• Even a lightweight cross-attention mechanism can efficiently integrate diverse feature streams, enhancing overall representation with minimal additional computational overhead.

Summary by alphaXiv.