Face Image Editing using Generative Adversarial Networks

Abstract

This dissertation presents novel methods and applications for face image editing using Generative Adversarial Networks (GANs). GANs have significantly advanced the field of face image editing, and this work delves into the challenges and possibilities of this technology, with a focus on improving segmentationguided image editing, GAN embedding for real face editing, domain adaptation, and hairstyle transfer. The research introduces new techniques, including the Semantic Region-Adaptive Normalization (SEAN) block for GANs conditioned on segmentation masks, an improved GAN embedding algorithm, a single-shot domain adaptation method, and the Barbershop approach for hair transfer. The dissertation also proposes pose-invariant hairstyle transfer techniques for images with varying poses. These techniques are evaluated on multiple datasets, and results demonstrate significant improvements over existing state-of-the-art techniques. By providing a comprehensive understanding of GAN-based face image editing, this work contributes to the ongoing evolution of face image editing and GAN applications, paving the way for future research and development.