Scalable Data Parallel Algorithms for Texture Synthesis and Compression using Gibbs Random Fields

David A. Bader
dbader@umiacs.umd.edu

Joseph JáJá
joseph@src.umd.edu

Rama Chellappa
chella@eng.umd.edu

Department of Electrical Engineering, and
Institute for Advanced Computer Studies,
University of Maryland, College Park, MD 20742

October 4, 1993

Abstract:

Permission to publish this abstract separately is granted.

Keywords: Gibbs Sampler, Gaussian Markov Random Fields, Image Processing, Texture Synthesis, Texture Compression, Scalable Parallel Processing, Data Parallel Algorithms.

• Introduction
• Parallel Algorithm Analysis
  • Parallel Communications Model
  • CM-5 Communications Model
    • Layout of parallel arrays
    • Simple and Regular Grid Shift
  • CM-2 Communications Model
  • Complexity of Some Basic Operations
• Texture Synthesis
  • A Parallel Gibbs Sampler
  • Gaussian Markov Random Field Sampler
    • Iterative Gaussian Markov Random Field Sampler
    • Direct Gaussian Markov Random Field Sampler
• Parameter Estimation for Gaussian Markov Random Field Textures
  • Least Squares Estimate of Parameters
  • Maximum Likelihood Estimate of Parameters
• Texture Compression
• Conclusions
• Example Texture Figures
• Equations used for the Computation of the Maximum Likelihood Estimate
• Timings for parallel image processing techniques
  • Tables for CM-5 with 4th order model
  • Tables for CM-5 with higher order model
• References
• About this document ...