Face transformation using StyleGAN

• Main Author: Cui, Naichuan
• Other Authors: Tan Yap Peng
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Electrical and electronic engineering::Electronic systems::Signal processing
• Online Access: https://hdl.handle.net/10356/153150
• Institution: Nanyang Technological University

With the development of the application of computer vision technology, face editing applications become common in various scenarios, and they are widely used in areas such as image beautification, live video streaming, and face confrontation attacks. In recent years, with the emergence of generative adversarial networks, the quality of face image generation has notably improved, making face editing methods even more popular. The proposal of StyleGAN in 2018 has led to great progress in the resolution and quality of face generation, and face-related research using StyleGAN has become a hot topic. The editing of a real facial image requires first obtaining its latent code, then using the vector of corresponding attributes for editing, and finally converting the edited latent code into a face image. In this dissertation, the whole process of face transformation using StyleGAN is investigated, and some improvements are made to the existing methods. The main work of this dissertation is as follows: This dissertation presents a method for computing the latent code of a real image using a combination of encoder and optimisation. In order to edit a real face image, its projection in the latent space needs to be obtained first. In this work, we first use a ResNet50-based network to obtain the approximate latent code from the original image and then use an optimisation algorithm to iterate over the latent code so that the corresponding image gradually approaches the original synthesized image, and eventually becomes close enough for subsequent face editing. The experimental results show that this method can significantly improve the speed of calculating the latent code while maintaining accuracy. We also test a method for separating attribute control vectors from face latent code and transformed faces, and compares it with other face editing methods. Specifically, the face latent code and the labels of several features are obtained and the vectors controlling specific attributes are separated from the face latent code by methods such as logistic regression. The face latent code is adjusted by the attribute control vectors, and the adjusted latent code is transformed into a face by StyleGAN to complete the transformation of specific face attributes. The experimental results show that this method can produce clearer and higher quality edited images of faces and is effective in face transformation.