@MASTERSTHESIS{pgi2020001, author = "T. Y. Chen", supervisor = "D. Roussinov", title = "Synthesising Images by Imagination ", school = "Department of Computer and Information Sciences, University of Strathclyde", year = "2019", abstract = "In this dissertation, four GANs models (vanilla GANs, DCGAN, WGAN, and WGAN-GP) were applied to three different training datasets including MNIST, NLVR, and Oxford-102 flowers. The models were successfully implemented through Keras and the code of the project can be efficiently executed by Python scripts. The limitations and the evolution of vanilla GANs were explored in the experiments of the dissertation. We can find that firstly DCGAN alleviated the non-convergence problem and produced good quality images on MNIST and Oxford-102 flowers; next, WGAN mitigated the mode collapse problem but failed on Oxford-102 flowers, using unsuitable method to restrict its discriminator. Lastly, WGAN-GP overcame all limitations and synthesised compelling images. In the end, DGAN produced the best results of Oxford-120 flowers among all models and yielded Fr{\~A}{\copyright}chet Inception Distance (FID) of 79 and a human error rate of 27.78%. On the other hand, WGAN-GP synthesised the best quality images on MNIST and NLVR. MNIST got FID of 6 and an error rate of 72.22%. NLVR obtained FID of 94 and an error rate of 19.44%.", }