Synthesizing texture involves the ordering of pixels in a 2D arrangement so as to display certain known spatial correlations, generally as described by a sample texture. In an abstract sense, these pixels could be gray-scale values, RGB color values, or entire spectral curves. The focus of this work is to develop a practical synthesis framework that maintains this abstract view while synthesizing texture with high spectral dimension, effectively achieving spectral invariance. The principle idea is to use a single monochrome texture synthesis step to capture the spatial information in a multispectral texture. The first step is to use a global color space transform to condense the spatial information in a sample texture into a principle luminance channel. Then, a monochrome texture synthesis step generates the corresponding principle band in the synthetic texture. This spatial information is then used to condition the generation of spectral information. A number of variants of this general approach are introduced. The first uses a multiresolution transform to decompose the spatial information in the principle band into an equivalent scale/space representation. This information is encapsulated into a set of low order statistical constraints that are used to iteratively coerce white noise into the desired texture. The residual spectral information is then generated using a non-parametric Markov Ran dom field model (MRF). The remaining variants use a non-parametric MRF to generate the spatial and spectral components simultaneously. In this ap proach, multispectral texture is "grown" from a seed region by sampling from the set of nearest neighbors in the sample texture as identified by a template matching procedure in the principle band. The effectiveness of both algorithms is demonstrated on a number of texture examples ranging from greyscale to RGB textures, as well as 16, 22, 32 and 63 band spectral images. In addition to the standard visual test that predominates the literature, effort is made to quantify the accuracy of the synthesis using informative and effective metrics. These include first and second order statistical comparisons as well as statistical divergence tests.

Library of Congress Subject Headings

Image processing--Digital techniques; Optical pattern recognition; Time-series analysis; Statistical decision

Publication Date


Document Type


Department, Program, or Center

Computer Science (GCCIS)


Anderson, Peter

Advisor/Committee Member

Schott, John

Advisor/Committee Member

Rhody, Harvey


Note: imported from RIT’s Digital Media Library running on DSpace to RIT Scholar Works. Physical copy available through RIT's The Wallace Library at: TA1637 .T977 2002


RIT – Main Campus