A natural image can take a significant amount of resources while processed by a modern computer, although it might appear straightforward to human at the first glance. The reason is that an image usually contains much redundant information which can be easily filtered out by human brain, but it turns out to be hard for a computer. Therefore, we often want a condensed version of the image, or a summary of a sequence of highly correlated images. We would like the representation to take as few resources as possible, while preserving the information sufficient for achieve our goals. Furthermore, this representation is preferably to be visually informative.
Color histograms and templates are two extremes in describing an image. the first one only summarizes the color information, and loses the spatial arrangement of an image. There have been great efforts to incorporate spatial information into the color histogram. The color coherence method proposed by Pass and Zabih [3] and correlogram by Huang, et al. [4] proved to be successful in applications such as image indexing and retrieval. But they are not visualized representations. On the other hand, methods based on templates and basis functions do maintain the geometrical properties of an image, but it suffers from large deformations in shape. They are too rigid to tolerate variations.
Jojic [1] presented a novel method ``epitome'' as the miniature of the image. It has considerably smaller size, but preserves most of the constitutive components of the image. The epitome can be considered as a generative model of the patches of an image. We can use appropriate statistical methods to extract the epitome1 from a single image, a collection of highly correlated images, or a video sequence. From the epitome and mappings we have learned, we can reconstruct the images with fairly good quality. Its potential savings in storage can be utilized in compression. As a probabilistic model which takes into account both color and texture properties, it is also beneficial to image retrieval. Furthermore, its simple relations to the original image defined by the associated mappings leads to efficient texture transfer techniques.
In section 2, we give both the conceptual definition and the statistical formulation of the epitome, followed by a description of the training and reconstruction procedure. In section 3, we present several applications using epitome.