Multi-Center Prototype Feature Distribution Reconstruction for Class-Incremental SAR Target Recognition

Area of Science:

• Artificial Intelligence
• Computer Vision
• Remote Sensing

Background:

• Deep learning-based Synthetic Aperture Radar (SAR) Automatic Target Recognition (ATR) systems face challenges with continuously emerging target categories.
• Catastrophic forgetting, where models lose previously learned information when acquiring new knowledge, hinders practical CIL applications.

Purpose of the Study:

• To propose a novel Class-Incremental Learning (CIL) method for SAR ATR systems to address catastrophic forgetting.
• To enhance the adaptability and performance of SAR ATR systems in dynamic environments with evolving target sets.

Main Methods:

• Developed Multi-center Prototype Feature Distribution Reconstruction (MPFR), a CIL method for SAR ATR.
• Introduced a Multi-scale Hybrid Attention feature extractor fusing SAR amplitude and Attribute Scattering Center data.
• Utilized multiple prototypes per class and parameterized Gaussian diffusion to model feature distributions and retain old knowledge.

Main Results:

• MPFR demonstrated superior performance compared to existing CIL approaches, including SAR-specific methods, on public SAR datasets.
• Ablation studies confirmed the effectiveness of individual components of the MPFR method.
• The proposed method successfully addresses CIL for SAR ATR without requiring storage of historical raw data.

Conclusions:

• MPFR effectively mitigates catastrophic forgetting in SAR ATR by preserving feature space capacity and modeling feature distributions.
• The method offers a robust solution for incremental learning in SAR ATR, enabling systems to adapt to new targets without performance degradation.
• MPFR represents a significant advancement in developing resilient and continuously learning SAR ATR systems.