Emotional morphing

Abstract

Deep Learning has been on the rise for the past decade, and Generative Adversarial Networks (GANs), one of the deep learning models, had experienced exceptional successes across numerous fields, especially computer vision. With facial images as one of the most commonly used data in computer vision, the project focuses on the morphing of facial images from one emotion to another using deep learning networks. CycleGAN, a deep learning model based on GANs, focuses on translating images from one style to another. This project seeks to better understand the architectural structure of CycleGAN and to make use of it for emotion morphing. The project also researched on the implementations of spectral normalisation in CycleGAN. Two new version of CycleGANs are implemented in this project and specified data are sorted from existing datasets and used for testing. The results of these new versions will be analysed with the results from the original CycleGAN as baseline. Results from the experiment show that both versions of CycleGAN implemented with the spectral normalisation has a faster speed of convergence and produce more realistic images.