Semi-Supervised Generative Adversarial Nets with Multiple Generators for SAR Image Recognition.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Semi-supervised learning models, often based on Generative Adversarial Nets (GANs), excel with optical images.
• GAN training instability hinders performance on Synthetic Aperture Radar (SAR) images, limiting feature extraction.
• Existing methods struggle with the unique challenges of SAR image analysis.

Purpose of the Study:

• To develop a stable and effective semi-supervised GAN for SAR image recognition.
• To improve feature extraction capabilities in GANs applied to SAR data.
• To enhance the performance of SAR image recognition systems.

Main Methods:

• Proposed a novel semi-supervised GAN with Multiple generators and a Classifier (MCGAN).
• Employed multiple generators to stabilize GAN training for SAR images.
• Introduced a multi-classifier sharing low-level layers with the discriminator to leverage labeled data.
• Constructed a recognition network by fine-tuning the trained discriminator and classifier.

Main Results:

• The MCGAN model demonstrated improved training stability for SAR images.
• The developed recognition network achieved superior and more stable performance on MSTAR databases.
• Outperformed traditional semi-supervised methods and other GAN-based approaches in SAR image recognition.

Conclusions:

• The proposed MCGAN framework offers a robust solution for semi-supervised SAR image recognition.
• MCGAN effectively addresses GAN training instability in SAR image analysis.
• The fine-tuned recognition network provides a significant advancement in SAR target recognition accuracy and stability.