Programming Research Group Technical Monograph PRG-99

C Philip Winder

DPhil thesis May 1992, 273 pages, ISBN 0-902928-77-5

Visual texture is a fertile source of information about the physical environment. Texture models should form rich but accessible descriptions of image composition and appearance. Markovian representations make explicit the variability of natural textures, but manipulation of current models is computationally demanding. This practical limitation enforces approximations and use of the simplified auto-normal form.

We propose two novel frameworks for Markovian texture analysis, and illustrate and quantify their advantages by adopting Bayesian classification of 33 Brodatz textures as a benchmark.

1. Simple spatially-parallel image filtering is computationally attractive, but our experiments demonstrate the limitations of segmentation algorithms responding to first-order differences of Gabor amplitude. We harness the efficiency of Gabor filtering within a hybrid Gabor-Markov framework by describing arrangements of local image features with random field models.

2. Our experimental appraisal of Gabor-Markov models established the importance of pre-processing image data prior to statistical analysis. We propose two families of Sampled-Markov models employing concise representations derived directly from the image data.

Both paradigms are more efficient and robust than a conventional Markovian analysis. Without reducing classifier accuracy, computational load was decreased by 88% using Gabor-Markov, and by 96% using Sampled-Markov models. Despite the improvements achieved by Gabor-Markov models, Smooth-Sampled Markov models perform better and have achieved 100% accuracy in our tests. We explain their superior performance by showing a strong correlation between classification accuracy and fidelity of the statistical modelling.