Texture synthesis using dimensionality-reduced appearance space

Abstract

A dimensionality-reduced appearance space system and method that transforms an exemplar image from a traditional three-dimensional space of pixel colors to a low-dimensional Euclidean space of appearance vectors. The transformation of an exemplar is a preprocessing step, and the transformed exemplar becomes the starting point for high-quality texture synthesis. The exemplar transformation begins by computing a high-dimensional appearance vector using one or a combination of several attribute channels. These attribute channels provide additional information to further distinguish exemplar pixels from each other. These attribute channels includes spatial pixel neighborhoods, feature distance, and radiance transfer information. Dimensionality reduction is applied to the resulting high-dimensional appearance vector to generate the transformed exemplar in low-dimensional Euclidean appearance space. Because much of the information contained in the high-dimensional appearance vector is redundant or coherent, dimensionality reduction can be applied to drastically reduce the dimensionality of the appearance vector with little loss of information.