Edge-based Inverse Procedural Texture Synthesis
Texture synthesis is a technique of synthesizing larger texture images from limited input. It is widely used in various fields such as digital art, games, real-time applications, and scientific visualizations that provide understanding of underlying principles about the appearance. One good application is soil profile visualzation. Soil scientists in the USA have created a large national database of written soil profile descriptions that follow a well-defined set of rules for describing soil morphological properties. Writing a soil description is straightforward although it is a skill that requires considerable practice. However, recreating a visual representation of a soil profile from a written description is very difficult. Procedural modeling is an effective way to visualize soil profile since its key properties, data compression and amplification, enable us to generate a large variety of detailed content using a set of rules without pre-generated data (Smelik, Tutenel, Bidarra, & Benes, 2014). Translating soil profile descriptions into a set of rules or parameters produces good visual and functional representations (Kim, Dorantes, Schulze, & Benes, 2016). However, designing a procedural model that fits a user’s intent is extremely difficult due to its complex generative processes and complicated evaluation. A casual user should take a cumbersome trial-and-error approach in order to obtain desired results. Thus, the inverse procedural modeling has become significant and it attempts to obtain appropriate set of parameters from given existing data. The key observation is that visually salient features (i.e., edges) can be extracted in an organized structure called edge groups, with clearly specified mutual relationships within the group and among the groups. The edge groups are then used as basic building blocks for the generation of procedural textures with similar visual appearance. Our approach detects and extracts the edge groups automatically from an input image (i.e., soil profile images). The user controls placement of edge groups from one or more input images to define its appearance. Pixels are automatically mapped from the input images and blended into the final structure. Our method also allows for a fully automatic procedural texture generation by controlling only a few input parameters of the final synthesis. We show our method on a variety of examples of stochastic textures. Our implementation analyzes textures within tens of seconds and the synthesis of a new texture is under a second.
Benes, Purdue University.
Off-Campus Purdue Users:
To access this dissertation, please log in to our