A comparative study of car image generation quality using DCGAN and VSGAN

Abstract

In today’s modern society, image generation (synthesis) has a great number of uses in various tasks. Image generation is used in crime forensics, improving image quality and generating better images. In 2014, a scientific breakthrough occurred in the machine learning community when Ian Goodfellow and his colleagues introduced the GAN (Generative Adversarial Network). Ever since then, GANs have become a more popular concept in the scientific community. Even today, GANs are being used, utilized and upgraded. This thesis is a comparative study of two GANs used for generating images of cars- DC-GAN (Deep Convolution) and VS-GAN (Vehicle Synthesis). The study will determine which of the two is better suited to generate high quality images of cars. We will train both GANs using the same dataset. The dataset consists of about 16185 Google images of random cars, 8144 for training and another 8041 for testing. The dataset is already preprocessed and split. We will compare the GANs training times, losses, accuracies and pictures generated, showing how well they perform. We will run all the GANs for 40 epochs in both training and testing. We will compare the CGAN, DCGAN, VSGAN, WGAN and WGAN-GP, to see which performs the best. We have used K-Nearest Neighbors, Regression and Random Forest Classifier to calculate the accuracies of all the GANs. We have displayed the results in tabular and graphical formats. We believe this will improve GAN research by providing an excellent comparison between the GANs and determine which is better suited for the given task. We also hope to improve the models further in the future and make an even more in depth comparison between the GAN architectures.