Wasserstein GANs for Generation of Variated Image Dataset Synthesis

Abstract

Deep learning networks required a training lot of data to get to better accuracy. Given the limited amount of data for many problems, we understand the requirement for creating the image data with the existing sample space. For many years the different technique was used to develop data fixtures to improve modelling and training efficiently with the advent of GAN we were now able to get close to real data. However, the standard GANs require a lot of effort in training and not cost-efficient. A more practical way of training GANs is Wasserstein GAN. They can be used for efficient generating for data taking the training sample space. Better representation GANs with WGANS solved the problem of learning a probability density. In this paper, we now use WGANs for classification training data given the sample data. We intend to deal with the following objectives. (a)To consider the sample space for the training data to mock with WGAN. (b)To build WGAN in combination with the network for classification and evaluating the models' performance. (c)To compare WGAN and standard GAN for knowing the increased accuracy of the classifier.