From rule-based to learning-based image-conditional image generation

Yi, Zili (2018) From rule-based to learning-based image-conditional image generation. Doctoral (PhD) thesis, Memorial University of Newfoundland.

[English] PDF - Accepted Version Available under License - The author retains copyright ownership and moral rights in this thesis. Neither the thesis nor substantial extracts from it may be printed or otherwise reproduced without the author's permission. Download (58MB)

Abstract

Visual contents, such as movies, animations, computer games, videos and photos, are massively produced and consumed nowadays. Most of these contents are the combination of materials captured from real-world and contents synthesized by computers. Particularly, computer-generated visual contents are increasingly indispensable in modern entertainment and production. The generation of visual contents by computers is typically conditioned on real-world materials, driven by the imagination of designers and artists, or a combination of both. However, creating visual contents manually are both challenging and labor intensive. Therefore, enabling computers to automatically or semi-automatically synthesize needed visual contents becomes essential. Among all these efforts, a stream of research is to generate novel images based on given image priors, e.g., photos and sketches. This research direction is known as image-conditional image generation, which covers a wide range of topics such as image stylization, image completion, image fusion, sketch-to-image generation, and extracting image label maps. In this thesis, a set of novel approaches for image-conditional image generation are presented. The thesis starts with an exemplar-based method for facial image stylization in Chapter 2. This method involves a unified framework for facial image stylization based on a single style exemplar. A two-phase procedure is employed, where the first phase searches a dense and semantic-aware correspondence between the input and the exemplar images, and the second phase conducts edge-preserving texture transfer. While this algorithm has the merit of requiring only a single exemplar, it is constrained to face photos. To perform generalized image-to-image translation, Chapter 3 presents a data-driven and learning-based method. Inspired by the dual learning paradigm designed for natural language translation [115], a novel dual Generative Adversarial Network (DualGAN) mechanism is developed, which enables image translators to be trained from two sets of unlabeled images from two domains. This is followed by another data-driven method in Chapter 4, which learns multiscale manifolds from a set of images and then enables synthesizing novel images that mimic the appearance of the target image dataset. The method is named as Branched Generative Adversarial Network (BranchGAN) and employs a novel training method that enables unconditioned generative adversarial networks (GANs) to learn image manifolds at multiple scales. As a result, we can directly manipulate and even combine latent manifold codes that are associated with specific feature scales. Finally, to provide users more control over image generation results, Chapter 5 discusses an upgraded version of iGAN [126] (iGANHD) that significantly improves the art of manipulating high-resolution images through utilizing the multi-scale manifold learned with BranchGAN.

Actions (login required)

View Item