PhD Proposal: Repetitive pattern synthesis and analysis
Peihan Tu
Monday, April 24, 2023, 3:00-5:00 pm 
Abstract
Repetitive patterns are ubiquitous in various fields, including textile design, digital art, web design, and graphic design. The ability to analyze and synthesize these patterns automatically can have significant practical and creative implications. The proposal presents my research on synthesizing and analyzing repetitive patterns. Repetitive vector patterns can be challenging and tedious to create.
Computational methods can automate parts of the manual creation process, but are mainly tailored for discrete pixels or elements instead of more complex patterns. To address these issues, first, we propose a computational method to synthesize continuous structures in vector patterns, by exemplar. Our main idea is to extend prior sample-based discrete element synthesis methods to consider not only sample positions (geometry) but also their connections (topology). Second, we propose an example-based method to synthesize more general patterns with diverse shapes and structured local interactions. Our main idea is to add explicit clustering as part of neighborhood similarity and iterative sample optimization for more robust sample synthesis and pattern reconstruction.
On the other hand, image texture is also an important visual component in real photos and computer-generated imagery. However, unlike vector patterns, editing image textures on pixels is unintuitive. We observe an image texture can be defined as an image with self-repeating visual elements: texels. Therefore, in the third part, we propose a neural tree representation for a generative model of textures that enables fast and intuitive user editing. The image texture is represented by a neural tree whose structure captures the recursive embedding of regions: each tree node represents an image region; the children of a node indicate a subregion within the node region. In the neural tree, the texels appear as subtrees with similar structures, with nodes having similar photometric and geometric properties. The representation is both expressive and easy to edit in that textures are modeled as discrete elements, resembling object-centric representation for a scene. We will provide initial results from this ongoing project.
Computational methods can automate parts of the manual creation process, but are mainly tailored for discrete pixels or elements instead of more complex patterns. To address these issues, first, we propose a computational method to synthesize continuous structures in vector patterns, by exemplar. Our main idea is to extend prior sample-based discrete element synthesis methods to consider not only sample positions (geometry) but also their connections (topology). Second, we propose an example-based method to synthesize more general patterns with diverse shapes and structured local interactions. Our main idea is to add explicit clustering as part of neighborhood similarity and iterative sample optimization for more robust sample synthesis and pattern reconstruction.
On the other hand, image texture is also an important visual component in real photos and computer-generated imagery. However, unlike vector patterns, editing image textures on pixels is unintuitive. We observe an image texture can be defined as an image with self-repeating visual elements: texels. Therefore, in the third part, we propose a neural tree representation for a generative model of textures that enables fast and intuitive user editing. The image texture is represented by a neural tree whose structure captures the recursive embedding of regions: each tree node represents an image region; the children of a node indicate a subregion within the node region. In the neural tree, the texels appear as subtrees with similar structures, with nodes having similar photometric and geometric properties. The representation is both expressive and easy to edit in that textures are modeled as discrete elements, resembling object-centric representation for a scene. We will provide initial results from this ongoing project.
Examining Committee
Bio
Peihan Tu is a 4th-year PhD student in Computer Science at UMD, advised by Prof. Matthias Zwicker. He is interested in graphics, image processing and HCI. Currently, he is particularly interested in developing novel algorithms/UIs for content authoring in graphics.
This talk is organized by Tom Hurst