3. THE PARALLEL ANALYSIS AND GRID
STRATEGY OF REGIONAL GROWING
SEGMENTATION
3.1 Traditional algorithm limitations
For regional growth division, several seed selection criteria and
growth criteria of traditional algorithm have its shortcomings.
Such as the iterative method and clustering method selected
seeds add the computational burden, but each growth criteria
use up a lot of computing resources and can only solve region
segmentation that satisfy certain conditions. Remote sensing
images have its characters such as regional, multi-band etc, so
ideal growth guidelines should consider a variety of criteria for
the growth of integrated application. To achieve this purpose, a
stronger and more computing resource is needed, and single
machine must develop to multi-machine integration.
Grid features of the impact of remote sensing are enormous.
Firstly, Grid can solve the massive remote sensing data storage
and access issues. Secondly, the computation time of some
large image can be reduced by Using Grid computing resources.
And then, Software and human resources to the high-sharing
agreement for the remote sensing images efficient, accurate
processing is a useful way to achieve.
3.2 The parallel analysis and Strategy of image
segmentation
Grid is an environment an tool including parallel characteristics,
and it can realize multi-machine parallel processing for remote
sensing image segmentation. Its feasibility mainly includes the
following:
3.2.1 Data parallelism: RS pixels are relative independence,
and the more senior RS data processing has greater relation.
Because image segmentation is the preliminary work for image
classification and feature extraction, so the large volume of data
in remote sensing image data can parallel compute on the basis
of block.
The whole image pixel is ordered by gray level:
IJ . T_T
1 rain ' 11 max
• The F m[n grey pixel is initialized of the first seed. By
searching through four Neighborhood pixel and stopping
until f (x, y) < Z , the first segment region is formed.
• Then according to f > f m \ n + Z , the most recent pixel
distance from the last region is selected to go on searching
until /(x, y) < IZ .
• There are not stop processing repetitively
until /(x, y) + nZ > F max .
• The Image loopholes is detected and the pixel value is
computed by [/(x,/) - (/ min + YlZ)] < Z The
pixel f (x, /) is classified to the n-recursive segmenting
region if meeting the conditions.
• Each preliminary Segmentation region’s basic parameters
are calculated, such as regional area and regional
average value .
• If satisfying the condition < S , the segment region is
combined with the most neighboring region.
Repeating the process, image segmentation is completed.
(3) Some explanations
In order to avoid over-segmentation problems, we can make use
of smoothing factor, for example Gaussian smoothing function,
to preprocess the image in the segmentation process.
We can utilize the mean gradient method to obtain the
segmental threshold on the basis of the image pre-processing.
4. GRID PROCESSING FOR PARALLEL
IMPLEMENTATION AND PERFORMANCE
EVALUATION
3.2.2 Algorithm parallelism: Regardless of what kind of
growth segmentation criteria, the processes of seed points are
targeted at the iteration process. So parallel operation is
possible.
3.2.3 Segmentation parallel strategy: Various
segmentation criteria can be integrated use. Based on the
growth of gray value segmentation results, the test for gray
structural features and characteristics of the border region
division can be simultaneously process to achieve better
segmentation results.
3.2.4 Parallel algorithm:
(1) Parameter settings:
There are many experiential parameters such as Z(Gray
threshold of Segmentation), S (the minimum segmentation
region size threshold), D C! M 2 ( a complete image of the
scope of M x M), j \ D —> M ( / denote the pixel Gray
value), h max and h mi „(respectively the smallest and the largest
gray value among/). Recursive process is growth from h min to
h max gray level and using point f as seed to form division
followed.
(2) A step-by-step approach of parallel segmentation
methodology for RS data is described below:
In this paper, the regional growth segmentation algorithm was
put forward, and during conducting of parallel program in the
grid environment we can make use of functional and data
hybrid parallel programming model, the concrete realization of
the algorithm described in the following:
Parallel function has the basic idea: "Copy Image divided
parameters", the various functional module handles the data is
the same. Assuming there are N number of processors and R
number of the parameters who need for the calculation,
parameters will be average allocated to the N processors and
each processor calculate the parameter that distributed, then a
result sent to the main system processor, and the main processor
is responsible for the integration of the final results. Parallel
data has the basic idea: "classified images, reproduction
parameters", the image is divided into N blocks average. Each
processor only put up all steps for the correlation coefficients
calculated.
The algorithm is a dimensional growth, and the seed choice is a
process of small to large. Once identified segmentation
threshold, the seed points sequence of pixels generate
automatically. Every one related to the growth of regional was
generate. This process is repetitive, and seed growth process is
independent of each other. So we can design data parallel
algorithm. In order to further provide the, Segmentation of the
region can achieve their own parallel gray regional
characteristics and supply parameters to further union region.
289