Informally, a texture is an image that is perceived and recognized partly as a statistical distribution, rather than a definite geometric shape. In texture perception, the interplay between statistics and geometry generally seems to go on at several scales, with different shapes and distributions involved at each level.
Texture synthesis (as opposed to texture acquisition from real-world images) is often necessary, for a variety of reasons: available texture samples may be too small or inhomogeneous, the surface to be textured has odd geometry or topology, or simply one needs extra ``knobs'' to customize or randomize the texture for a particular application.
We have focused on one particular method for texture synthesis, based on multiscale statistical imitation of a given sample texture. The basic (monoscale) algorithm can be summarized as follows:
The full (multiscale) algorithm consists of applying the basic algorithm above to a pyramid of texture pairs S[i], A[i], i = 1,2,...n, at different resolutions. Each sample texture S[i] is a shrunken copy of S[i-1], starting with the original sample S[0] = S and ending with a 1x1 texture S[n]. Each synthetic texture A[i] is obtained through the basic algorithm from the next smaller texture A[i+1], using the latter as the initial guess and S[i] as the reference texture.
Jorge's work in this project was supported in part by the CNPq research grant 301016/92-5. A substantial part of this work was performed by Carlos H. Q. Forster, supported by the FAPESP undergraduate research assistanship grant no. (grant 93/4681-7).