Faculty of Science - Cairo University

Name: Mohammed Abdel-Fattah Abdel-Aziz Mohammed.

Date of Birth: 7/5/1973.

Previous Degree: B.Sc.

Registration Date: 11 / 7 / 2000.

Awarding Date:     17 / 2 / 2008.

Supervisors:

1.     Prof. Ahmed Badr El-Din Khalil, Department of Mathematics, Faculty of science, Cairo University.

2.      Dr. Reda Abdul-Wahab El-khoribi, Department of Information technology, Faculty of computer and information science, Cairo University.

Examiners:

1.     Prof. Simon Y. Berkovitch, Department of Computer Science, George Washington University.

2.     Dr. Mahmoud Ahmed Ismail, Department of information technology, Faculty of Computer & Information Science, Cairo University.

3.     Prof. Ahmed Badr El-Din Khalil (Supervisor), Department of Mathematics, Faculty of science, Cairo University.

Title of Thesis:

Unsupervised Texture Segmentation of Digital Images””

Key Words:

Unsupervised/Supervised Texture Segmentation, Hidden Markov Tree, Multiresolution, Wavelet-Transform, Contourlet-Transform, Ranklet-Transform, Markov Random Field, Mixture Model, Gibbs Filter, Deterministic Annealing, Active Regions, Level Set, Active Polygons, Wavelet Frames.

Summary:

The goal of this thesis is to develop an algorithm for unsupervised texture images segmentation. The segmentation of a texture image aims to labeling each texture in the

input image by a different gray level. For example, if the input image consists of four textures, we have to deal with four different gray levels in the output image. It can be classified as supervised and unsupervised segmentation according to the availability of texture prototypes. In this work, the unsupervised segmentation issue will be studied. It is widely involved in a variety of computer vision tasks, such as medical diagnosis, remote sensing, and automatic surveillance/detection systems, etc.

The proposed unsupervised texture image segmentation method is exploits district local texture behaviors by using the likelihood disparity with respect to a global statistical model. The wavelet transform, contourlet transform and ranklet transform are specifically investigated to provide a multiresolution decomposition of image. The Hidden Markov Tree (HMT) models are developed as statistical models that impose a tree-structured markov chain across scales to capture inter-scale dependencies of wavelet coefficients /Contourlet coefficients/ Ranklet coefficients. Wavelet-domain Hidden Markov Tree (WHMT)/Contourlet Hidden Markov Tree (CHMT)/Ranklet Hidden Markov Tree (RHMT) are specifically studied in this work.

Unsupervised segmentation approach is developed by capturing the likelihood disparity of different texture features with respect to Wavelet-domain HMT/ Contourlet HMT/ Ranklet HMT. The K-means clustering is used to produce a raw segmentation map based on the likelihood values, so that the unsupervised segmentation problem can be converted into a self-supervised one. The segmentation process can accomplish by using the multiscale supervised Bayesian segmentation algorithms. A dual-model framework is suggested. which can take advantage of two different WHMT\CHMT\RHMT to implement unsupervised segmentation.

The simulation results of synthetic mosaics indicate that the proposed algorithm can achieve high segmentation accuracy which is much close to that of the supervised case.

Designed and Directed by A.H.A