Factoring repeated content within and among images.

We reduce transmission bandwidth and memory space for images by factoring their repeated content.  A transform map and a condensed epitome are created such that all image blocks can be reconstructed from transformed epitome patches.  The transforms may include affine deformation and color scaling to account for perspective and tonal variations across the image.  The factored representation allows efficient random-access through a simple indirection, and can therefore be used for real-time texture mapping without expansion in memory.  Our scheme is orthogonal to traditional image compression, in the sense that the epitome is amenable to further compression such as DXT.  Moreover it allows a new mode of progressivity, whereby generic features appear before unique detail.  Factoring is also effective across a collection of images, particularly in the context of image-based rendering.  Eliminating redundant content lets us include textures that are several times as large in the same memory space.