We consider object recognition as the process of attaching meaningful labels to specific regions of an image, and propose a model that learns spatial relationships between objects. Given a set of images and their associated text (e.g. keywords, captions, descriptions), the objective is to segment an image, in either a crude or sophisticated fashion, then to find the proper associations between words and regions. Previous models are limited by the scope of the representation. In particular, they fail to exploit spatial context in the images and words. We develop a more expressive model that takes this into account. We formulate a spatially consistent probabilistic mapping between continuous image feature vectors and the supplied word tokens. By learning both word-to-region associations and object relations, the proposed model augments scene segmentations due to smoothing implicit in spatial consistency. Context introduces cycles to the undirected graph, so we cannot rely on a straightforward implementation of the EM algorithm for estimating the model parameters and densities of the unknown alignment variables. Instead, we develop an approximate EM algorithm that uses loopy belief propagation in the inference step and iterative scaling on the pseudo-likelihood approximation in the parameter update step. The experiments indicate that our approximate inference and learning algorithm converges to good local solutions. Experiments on a diverse array of images show that spatial context considerably improves the accuracy of object recognition. Most

The underlying aim of this research is to investigate the mathematical descriptions of homogeneous textures in digital images for the purpose of segmentation and recognition. The research covers the problem of testing these mathematical descrip- tions by using them to generate synthetic realisations of the homogeneous texture for subjective and analytical comparisons with the source texture from which they were derived. The application of this research is in analysing satellite or airborne images of the Earth's surface. In particular, Synthetic Aperture Radar (SAR) images often exhibit regions of homogeneous texture, which if segmented, could facilitate terrain classification.

The strong Markov random field (MRF) model is a sub-model of the more general MRF-Gibbs model. The strong-MRF model defines a system whereby not only is the field Markovian with respect to a defined neighbourhood, but all sub-neighbourhoods also define a Markovian system. A checkerboard pattern is a perfect example of a strong Markovian system. Although the strong Markovian system requires a more stringent assumption about the field, it does have some very nice mathematical properties. One mathematical property, is the ability to define the strong Markov random field model with respect to its marginal distributions over the cliques. This property allows a direct equivalence to the ANOVA loglinear construction to be proved. From this proof, the general ANOVA log-linear construction formula is derived.

We approach the object recognition problem as the process of attaching meaningful labels to specific regions of an image. Given a set of images and their captions, we segment the images, in either a crude or sophisticated fashion, then learn the proper associations between words and regions. Previous models are limited by the scope of the representation, and performance is constrained by noise from poor initial clusterings of the image features. We propose three improvements that address these issues. First, we describe a model that incorporates clustering into the learning step using a basic mixture model. Second, we propose Bayesian priors on the mixture model to stabilise learning and automatically weight features. Third, we develop a more expressive model that learns spatial relations between regions of a scene. Using the analogy of building a lexicon via an aligned bitext, we formulate a probabilistic mapping between the image feature vectors and the supplied word tokens. To find the best hypothesis, we hill-climb the log-posterior using the EM algorithm. Spatial context introduces cycles to our probabilistic graphical model, so we use loopy belief propagation to compute the expectation of the complete log-posterior, and iterative scaling and iterative proportional fitting on the pseudo-likelihood approximation to render parameter estimation tractable. The EM algorithm is no longer guaranteed to converge with an intractable posterior, but experiments show the approximate E and M Steps consistently converge to a local solution. Empirical results on a diverse array of images show that learning image feature clusters using a standard mixture model, feature weighting using Bayesian shrinkage priors and spatial context potentials considerably improve the accuracy of gen...

This paper introduces a new classification scheme called "open-ended texture classification." The standard approach for texture classification is to use a closed n-class classifier based on the Bayesian paradigm. These perform supervised classification, whereby all the texture classes have to be predefined. We propose a new texture classification scheme, one that does not require a complete set of predefined classes. Instead our texture classification scheme is based on a significance test. A texture is classified on the basis of whether or not its statistical properties are deemed to be from the same population of statistics as those that define a specific texture class. This new "open-ended texture classification" is considered potentially valuable in the practical application of terrain mapping of Synthetic Aperture Radar (SAR) images.

This thesis investigates texture classification using Non-Parametric Markov Random fields. Texture models using local image descriptors are investigated. Classification performance using such models is then reported upon and the results are used to guide development of future classifiers which take account of scale information within an image. The issues and effects of scale within texture modelling and classification are explored. From this investigation texture models which incorporate scalar information are developed. Results are presented upon these classifiers and the reasons behind them are analysed. The use of region detectors within texture classification is investigated and its role questioned. Finaly an investigation of the role that the zooming level plays within texture is investigated. i Acknowledgements I would like to thank my supervisor, Louis Atallah for all the invaluable help, support, commitment and enthusiasm he has given me throughout this project. Without his

In this paper we present noncausal, nonparametric, multiscale, Markov Random Field (MRF) model for syntheslslng and recognislng texture. The model has the ability to capture the characteristics of a wide variety of textures, varying from the structured to the stochastic. For texture synthesis, we use our own novel multiscale approach, incorporating local annealing, allowing us to use large neighbourhood systems to model some complex textures. We show how we are able to manipulate the statistical order of our high dimensional model without over compromising the integrity of the representation. Also by varying the statistical order of our model we are able to optimise it for the unsupervised recognition of textures with respect to textures that have not been modelled.

For applications in texture synthesis, we derive two approximate Bayes criteria for selecting a model from a collection of Markov random fields. The first criterion is based on a penalized maximum likelihood. The second criterion, a Markov chain Monte Carlo approximation to the first, has distinct computational advantages. Some simulation results are also presented. KEY WORDS AND PHRASES: fields, texture synthesis, Markov chain Monte Carlo.