Auxiliary Conditional Generative Adversarial Networks for Image Data Set Augmentation

Abstract

Adversarial models have been widely used for data generation and classification in the fields of Computer Vision and Artificial Intelligence. These adversarial models are defined over a framework in neural networks called Generative Adversarial Networks. In this paper, we use auxiliary conditional generative models which are special kinds of GANs employing label conditioning that result in newly generated images exhibiting global coherence. This conditional version of generative models is constructed by feeding data that we wish to condition on generator network and discriminator network in a GAN. The analysis has experimented on a high-resolution dataset called FMNIST across 60,000 samples of training images with reshaped image resolution size of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$28^{\ast}28$</tex> . The following procedure is used for image dataset augmentation which improves the accuracy of image classifiers/segmentation techniques.