Enhanced Balancing GAN: Minority-class Image Generation

TL;DR

This paper introduces an improved GAN architecture that stabilizes minority-class image generation in imbalanced datasets by using a supervised autoencoder and gradient penalty, achieving faster convergence and higher quality results.

Contribution

It proposes a novel autoencoder-based initialization and an enhanced BAGAN with gradient penalty to improve stability and performance in minority-class image generation.

Findings

Achieves high-quality minority-class images on imbalanced datasets.

Converges faster than original BAGAN.

Effective on medical and standard image datasets.

Abstract

Generative adversarial networks (GANs) are one of the most powerful generative models, but always require a large and balanced dataset to train. Traditional GANs are not applicable to generate minority-class images in a highly imbalanced dataset. Balancing GAN (BAGAN) is proposed to mitigate this problem, but it is unstable when images in different classes look similar, e.g. flowers and cells. In this work, we propose a supervised autoencoder with an intermediate embedding model to disperse the labeled latent vectors. With the improved autoencoder initialization, we also build an architecture of BAGAN with gradient penalty (BAGAN-GP). Our proposed model overcomes the unstable issue in original BAGAN and converges faster to high quality generations. Our model achieves high performance on the imbalanced scale-down version of MNIST Fashion, CIFAR-10, and one small-scale medical image dataset.