Generating and compressing images from a large volume of discrete datasets using GANs along with different compression techniques and studying the results

Abstract

Image is among the most common and important factors in modern day research. From Image processing to Image synthesis all the aspects of image are necessary and have always been prioritized. For this we tried to incorporate a comparatively new process of image generation in our research, that is GAN. In this new era of technology GAN has gained a lot of popularity for generating new images and synthesizing old images. Our thesis is a study of two popular GANs that is CGAN and DCGAN where we came up with the working ability of both the GANs by analyzing its training and testing with the help of a large volume of discrete datasets. One of the datasets consists of almost 16000 cars images and the other dataset is of dogs images which contains almost 5000 dogs images. We have run both the DCGAN and CGAN for both the datasets with 50 epochs in training and testing. Moreover besides the use of GAN and comparing it we compared three different techniques of image compression which are Discrete Cosine Transform that is DCT, K-Means Clustering and the Pillow Library of Python. With the use of image compression tools, we can compress images fast and efficiently, resulting in a reduction in storage space while maintaining a minimal influence on picture quality. We compressed both the real images from our dataset and the fake generated images. After that we studied the results by comparing the compression percentage and differentiating the images quality. We believe that our research will provide an excellent comparison of the GANs and compression techniques which will help future researchers to understand which technique to use for optimum result. We hope to improve our models in the future and also incorporate both the image generation and compression to come up with better quality images using less memory space. It means that we will be able to achieve the greatest amount of clarity while taking up the least amount of space.