Enhanced balancing GAN: minority-class image generation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Generative adversarial networks (GANs) excel at image generation but require large, balanced datasets.
• Traditional GANs struggle with imbalanced datasets, failing to generate minority-class images.
• Existing solutions like Balancing GAN (BAGAN) face instability issues with visually similar classes.

Purpose of the Study:

• To address the instability and limitations of GANs in generating minority-class images from imbalanced datasets.
• To propose a novel GAN architecture that improves stability and generation quality for datasets with class imbalance.
• To enhance the training process of GANs for improved performance on challenging datasets.

Main Methods:

• Developed a supervised autoencoder with an intermediate embedding model to disperse labeled latent vectors.
• Integrated this enhanced autoencoder initialization into a Balancing GAN with Gradient Penalty (BAGAN-GP) architecture.
• Employed gradient penalty to stabilize the training of the GAN model.

Main Results:

• The proposed BAGAN-GP model demonstrates improved stability compared to the original BAGAN, particularly for datasets with similar-looking classes.
• Achieved faster convergence and higher-quality image generations.
• Showcased high performance on imbalanced versions of MNIST Fashion, CIFAR-10, and a medical imaging dataset.

Conclusions:

• The novel BAGAN-GP architecture effectively overcomes the instability issues of previous GAN models in imbalanced scenarios.
• The enhanced autoencoder initialization and gradient penalty contribute to more robust and efficient GAN training.
• The model shows significant potential for applications requiring high-quality image generation from imbalanced datasets, including medical imaging.