الفهرس 
SEGMENTATION, QUALITY MEASURE AND KNOWLEDGEOnly 14 pages are availabe for public view
 |  | from | 184 |  |  |
| | | from | 184 | | |
Abstract
The objective of this thesis is to propose a developed and
efficient approach to solve segmentation problem of texture image
”natural image” to achieve the performance levels in the speed of
execution in image processing for time critical problems. The
problem of segmentation requires a grouping and separation of
structured or unstructured pixels to form a natural object of
subimage. Segmentation has been carried out by so many
techniques. However in texture segmentation, the algorithms are not
robust and their performances are inefficient.
In our contribution, we developed the problem formulation
based on stochastic analysis for its human perception similarity.
The problem is divided into two independent subproblems. The first
one is the iterative seeking for the best parameters that represent the
stochastic model assumed for the textures comprising the tested
images. The second one is to find the optimum segmentation of the
domain similar to the tested properties.
In our solution of the formulated problem, we have taken two
paths. In the first one, we uied to solve the problem by simulating
annealing (SA) for optimum solution of the two parts of the
problem. We adopted an existing algorithms and modified it to
improve its convergence. However, the results of both the adopteand modified algorithms were inefficient and slow to describe the
feature of texture images. The SA optimization does not take in
consideration all parameters globally and depends on initial
conditions. In the second part, it leads to inaccurate segmentation
of tested texture images.
In the second path, we have proposed a new approach to solve
the problem based on genetic algorithm. The algorithm is developed
to solve both first and second part of the problem based on decimal
numbers. We have reached the global solution for modeling and
segmentation of texture images. Afterwards the algorithm is applied
on four test images taken from Brotaz manual for Dog fur, grass,
river pebbles, and cork. A comparison of parameters using genetic
algorithm, simulating annealing and least square for optimization is
presented, then combined the four images to test the segmenator.
Finally, simulated results demonstrate the efficiency of the
proposed method where it can be applied in parallel for
segmentation of remote sensing image using CD-ROM