Global Receptive-Based Neural Network for Target Recognition in SAR Images

Area of Science:

• Artificial Intelligence
• Machine Learning
• Signal Processing

Background:

• Neural networks require large labeled datasets for parameter estimation, which are difficult to obtain for radar target recognition.
• High computational costs and slow convergence limit practical applications of current neural network models.

Purpose of the Study:

• To develop a novel receptive neural network strategy for radar target recognition.
• To address challenges of limited labeled samples and high computational costs.
• To improve efficiency and performance in realistic environments.

Main Methods:

• A two-module strategy: feature generation and feature refinement.
• Feature generation uses Riesz transform and bandpass filter banks to create receptive signals for global feature extraction.
• Receptive signals are predefined, not learned, to handle small sample sizes; refinement uses randomly generated weights and Hilbert space casting for kernel refinement.

Main Results:

• The proposed receptive neural network effectively extracts global features from SAR images.
• Predefined receptive signals and a specialized refinement unit enable learning with limited data.
• The strategy aims for comparable or superior performance with reduced training resources.

Conclusions:

• The novel receptive neural network offers a promising solution for radar target recognition with limited training data.
• This approach mitigates the need for extensive data collection and reduces computational burden.
• Expert knowledge integration and kernel refinement enhance model robustness and efficiency.