Performance Study of Image Data Augmentation by Generative Adversarial Networks

Abstract

Convolutional Neural Networks (CNN) are becoming increasingly popular among machine learning techniques for image classification problems. But these networks are highly dependent on large set of data for proper learning to avoid overfitting in the context of supervised classification. Contrary to numerical data, image data poses difficulty for data augmentation by artificial increase of training data. Recently Generative Adversarial Networks (GAN) have been developed and are increasingly used for generation of image data. It has been proved that GAN has higher potential in generating image data for training purpose compared to traditional methods of data augmentation in the area of image processing. In this work we have studied the performance of different GAN developed recently in terms of their effectiveness in classification problems compared to baseline CNN without GAN. We have also done the comparative study on the qualities of fake image generation by different GAN in terms of IS (Inception Score) and FID (Fréchet Inception Distance). In this work we have studied the performance of Lightweight GAN, Data-efficient GAN and StyleGAN-ADA, in addition to the most conventional Deep Convolutional Generative Adversarial Networks (DCGAN). Simulation experiments have been done using bench mark data set Food-101. The results show that Lightweight GAN is the most efficient for data augmentation.