ISSN (Print) : 2319-5940

ISSN (Online) : 2278-1021

International Journal of Advanced Research in Computer and Communication Engineering

Vol. 2, Issue 4, April 2013

Image Processing Tasks using Parallel Computing

in Multi core Architecture and its Applications in
Medical Imaging
Sanjay Saxena¹, Neeraj Sharma², Shiru Sharma³
Research Scholar, School of Biomedical Engineering, IIT (BHU), Varanasi, UP, India¹
Associate Professor, School of Biomedical Engineering, IIT (BHU), Varanasi, UP, India²
Assistant Professor, School of Biomedical Engineering, IIT (BHU), Varanasi, UP, India³

Abstract: To find accurate & reliable result in image analysis, it is important that image is processed and analyzed using
image processing suitable AI technique further at the same time it is highly desired that processing time must be minimum.
Preprocessing of the image makes it more clear and visible, while parallelizing of the algorithm optimizes the speed at
which the image is processed. This paper explores current multi-core architectures available in commercial processors in
order to speed up the image processing tasks. Parallel Implementation of Many sequential algorithms of Image processing
was examined and analyzed in test and achieved good result if all the recourses are efficiently used. Main objective of this
paper is to design some parallel image processing algorithms like segmentation, noise reduction, features calculation,
histogram equalization etc by using Multi Core architecture and comparative study with some sequential image processing
algorithm. These parallel algorithms are able to work with different number of thread, so as to take all the benefits of the
upcoming processors having any number of cores. As medical imaging refers to view the human body in order to diagnose,
monitor and treatment planning. This paper also describes the application of parallel computing applied in different Medical
Imaging techniques like CT, PET scans etc.

Keywords: Parallel Computing, Image Processing, Histogram, Multi Core, CPU, Latency, CT, PET.

I. INTRODUCTION thereof) for clinical purposes or medical purpose [5]. There

are so many different medical image modality are present
Image is the 2 dimensional distributions of small image like CT, PET, MRI etc. These Modalities are having
points called as pixels. It can be considered as a function of different characteristics and used as per requirements.
two real variables, for example, f(x,y) with f as the Because size of these images are very large so for analysing
amplitude (e.g. brightness) of the image at position (x,y). these modalities take so much time to process sequentially
Image Processing is the process of enhancing the image and and give result after some time. So if we divide this
extraction of meaningful information from an image. Image sequentially processing to efficient parallel processing than
processing is gaining larger importance in a variety of we can find good result in very reasonable time or if we are
application areas. Active vision, e.g. for autonomous able to process basic steps like Image Enhancement,
requires substantial computational power, in order to be able Morphological operation, feature calculation quickly than it
to operate in real time[1]. According to Jain [3], digital will be beneficial for the medical practitioner. So by the
image processing consists of the application of functions that parallel computing we can save time and/or money, we can
transform a two dimensional image using a computer. Others solve larger problems in very short time periods. Parallel
authors as Crane [4] define this task as a science that computing provides concurrency and by this we can use non
manipulates digital images that covers an extend set of – local recourses very efficiently. It also removes the limit of
techniques to enhance or distort them. serial computing. Matlab 2010a onwards finally enables the
Medical imaging is the technique and process used to „Parallel Computation Toolbox‟ for student use. It‟s not part
create images of the human body (or parts and function of the core Matlab student package, but it is now available as

Copyright to IJARCCE www.ijarcce.com 1896

 ISSN (Print) : 2319-5940
ISSN (Online) : 2278-1021

International Journal of Advanced Research in Computer and Communication Engineering

Vol. 2, Issue 4, April 2013

an add-on toolbox. Of course, it‟s been available for III. PROPOSED ALGORITHMS FOR PARALLEL
commercial users for some time.[6] There is an increasing IMAGE PROCESSING TASKS
recognition that High-Level Languages, and in particular
scripting languages such as MATLAB, provide enormous We have gone through following sequential image
productivity gains in developing technical and scientific processing algorithms and developed its parallel versions
code [11]. with different approaches in core i3 processor and found
good results with variable numbers of threads passed as
II. IMAGE PROCESSING WITH PARALLEL input parameters. We focus on the most promising and well
COMPUTING supported techniques with an emphasis in SIP application
examples, namely: MatlabMPI [2], bcMPI [12], pMatlab
Image Processing with Parallel computing is an alternative [13], Star-P [14] and the MATLAB Distributed Computing
way to solve image processing problems that require large Toolbox (DCT) [15].
times of processing or handling large amounts of Load Distribution between Processors:
information in "acceptable time" (according to each Suppose we are taking the following image (1) as an
criterion). As we can see that medical imaging requires lots example.
of memory space and time to process so by parallelizing we The main step of these algorithms is to determine the
can find efficient and fast result. In parallel processing, a number of tiles to be generated. The number of tiles
program is able to create multiple tasks that work together to corresponds to the amount of threads. If only a thread exists,
solve a problem [7]. The main idea of parallel image the computation is just sequential computation. Otherwise if
processing is to divide the problem into simple tasks and there are two or more threads than the image is divided into
solve them concurrently, in such a way the total time can be distinct areas, as shown in Figure 2, 3, 4. Each thread is
divided between the total tasks (in the best case). Parallel responsible for processing the pixels included in its tile and
Image processing cannot be applied to all problems, in other to execute different tasks but considering to maintain
words we can say that not all the problems can be coded in a synchronization between all the processor otherwise there
parallel form. A parallel program should must have some will be the situation of deadlock between processors.
features for a correct and efficient operation; otherwise, it is
possible that runtime or operation does not have the
expected performance. These features include the following
[8]:

Granularity.- It is defined as the number of basic units and

it is classified as:
Coarse-grained.- Few tasks of more intense computing.
Fine grain.- A large number of small parts and less intense
computing.

Type of parallel processing:

Explicit.- The algorithm includes instructions to specify
which processes are built and executed in parallel way.
Implicit.- The compiler has the task of inserting the
necessary instructions to run the program on a parallel Fig. 1
computer.

Synchronization.- This prevents the overlap of two or more

processes.

Latency.- This is the time transition of information from

request to receipt.

Scalability.- It is defined as the ability of an algorithm to

maintain its efficiency by increasing
the number of processors and the size of the problem in the
same proportion [9]. Fig. 2

Copyright to IJARCCE www.ijarcce.com 1897

 ISSN (Print) : 2319-5940
ISSN (Online) : 2278-1021

International Journal of Advanced Research in Computer and Communication Engineering

Vol. 2, Issue 4, April 2013

evenly spaced on a grid, etc.) for different regions on client

processor.[5]
3. Count number of seed points and activate same
number of kernels.
4. Send copy of image and respective seed pixels on
each kernel processor.
5. Regions are then grown from these seed pixels to
adjacent depending on pixel intensity, gray level texture, or
color, etc on each kernel processor for this image
information is also important because region grown is based
on the membership criteria.
6. Calculation of different features of different ROI
(region of interest) on individual kernel.
7. Send segmented regions and information of ROIs to
the client processor and deactivation of worker kernels.
8. Reconstruct the image on client processor and
Display of all the information.

Fig. 3. b. Parallel Segmentation by Global Thresholding

and Calculation of Fourier and Regional Descriptors of
an Image

This algorithm consists of the following steps.

1. Divide the image into Quad Tree Structure in client
processor.

2. Send a flag from client processor to different worker

processor and activation of four worker processors.

3. Send all the parts of the image on different kernels or

different labs or different worker processors.

4. Chose Initial Thresholds T0, T1, T2, T3 in different labs

individually.
Fig. 4. 5. Calculate the mean value of each lab µ of the pixel
According to the requirement we can divide an image and below the threshold and the pixel above the threshold value.
send it different processor.
6. Compute a new threshold as
a. Parallel Segmentation by Region Growing
Technique and Calculation of different Features of T = (µ1 + µ2)/2 in each lab.
Segmented Regions 7. Repeat steps 5 and 6 until there is no change in
threshold in each lab or worker.
Region growing is one of the very famous techniques of
segmentation. Main drawback of the region growing is that 8. Send segmented region to client from the different labs.
it is time consuming. This algorithm is designed to reduce
timing of this algorithm. These are the steps involved in it. 9. Deactivate worker processors.
1. Take Input image on client processor. 10. Reconstruct the Segmented Image.
2. Selection of the Seed pixels(It is based on some
user criterion like pixels in a certain gray level range, pixels

Copyright to IJARCCE www.ijarcce.com 1898

 ISSN (Print) : 2319-5940
ISSN (Online) : 2278-1021

International Journal of Advanced Research in Computer and Communication Engineering

Vol. 2, Issue 4, April 2013

11. Calculate the Regional and Fourier descriptors of 6. Send all equalized histogram to the client
Segmented Image on client processor. processor.

IV. RESULT AND DISCUSSION

c. Complex Noise reduction usingParallel This section describes the results obtained with the parallel
Computing: implementation of above algorithms. This also deals with the
environment used in the experiments, the test images, and
This algorithm consists of the following steps: results, along with the evaluation of the performance
1. Take input Image on Input Image on client obtained with the parallelization.
processor.

2. Generate Copies of the Image on client processor.Test Environment

The experiments were all performed on an Intel Core i3 –
3. Activate Number of Kernels or Workers or Labs as
2350M Processor 2.30 GHz, 3 GB of RAM, Hard Disk
per requirement.
Drive 320 GB
4. Send Copy of the Image into different Kernels. Software used MATLAB R2011a and JAVA JDK 1.6.0_21,
64 bit operating system.
5. All the labs consist of filters for e.g.

Lab 1 – Median filter Parallelization Scheme in MATLAB

Lab 2 - Wiener Filter Multithreaded Parallelism
Lab 3 – Order statistical filter Explicit Parallelism
Lab 4 – Min – Max filter etc.

6. Perform Convolution and find PSNR and MSE of Following Table 1 illustrates the performance of parallel
the image on individual kernels. image processing algorithm on different size images.

7. Send back Values of MSE and PSNR value to the

client Time(T
Time(T1) 2)
8. Compare the result of received data on client Alg Taken by Taken Observed
processor. Image Size orit Sequenti by Speed
s hms al Parallel Up
9. Display the image with Highest PSNR and lowest Algorith Algorith µ =
MSE Value. m(in sec) m(in T1/T2
sec)
10. Deactivation of kernels a 0.3267 .15691 2.082085
Came 256 b 0.268 0.0956 2.803347
raman x c 0.9132 0.9005 1.014103
d. Histogram Equalization of an Image by Parallel .tif 256 d 0.259 0.0823 3.1567
Computing
a 1.0896 .00783 1.081134
1. Divide the image in quad tree format or in row wise Blure 256 b 4.6117 .10005 2.196
or in column wise format and activate required number of dtext.j x c 0.1394 .06983 1.9965
kernel processors. pg 768 d 0.9857 0.3279 3.006
2. Form the cumulative histogram on each part of a 3.98154 2.99105 1.331149
image on each kernel processor. Colorl 128 b 4.4722 3.9612 1.129
3. Normalize the value by dividing it by total number eaf.jp 1 x c .319738 3.665 1.9972
of pixels. g 843 d 1.793 0.8932 2.007
4. Multiply these value by the maximum gray level
value and round off the value. Table 1: Performance Evaluation Table
5. Map the original value to the result of step 3 by a
one to one correspondence.

Copyright to IJARCCE www.ijarcce.com 1899

 ISSN (Print) : 2319-5940
ISSN (Online) : 2278-1021

International Journal of Advanced Research in Computer and Communication Engineering

Vol. 2, Issue 4, April 2013

V. CONCLUSION & FUTURE SCOPE dissertation, Departamento de Sistemas Informáticos y Computación.

Universidad Politécnica de Valencia, Valencia, España, 2009.
[10]. Thomas Bräunl,” Tutorial in Data Parallel Image Processing,” in
This work presents parallel implementation of different Australian Journal of Intelligent Information Processing Systems (AJIIPS),
sequential image processing algorithm. It was developed an vol. 6, no. 3, 2001, pp. 164–174 .
algorithm using OpenMP threads in order to leverage the [11]. A. Edelman, P. Husbands, S. Leibman, “Interactive
parallel processing capability of current processors with Supercomputing‟s Star-P Platform: Parallel MATLAB and MPI Homework
Classroom Study on High Level Language Productivity,” HPEC, 2006.
multiple cores and we can see that the speed up is good. [12].bcMPI:http://www.osc.edu/hpc/software/apps/bcmpi.shtml
The focus of this implementation was to improve the [13] pMatlab: http://www.ll.mit.edu/pMatlab/
performance of segmentation, de noising, histogram [14] Star-P: http://www.interactivesupercomputing.com/
processing keeping the reproducibility of results. In terms of [15]DCT: http://www.mathworks.com/products/distribtb/
performance, parallel implementation was about two and a
half times faster than the sequential segmentation. This is a BIOGRAPHY
very promising result since it allows the exploitation of the
vast processing power of current processors with multiple Sanjay Saxena is a research scholar at
cores. In the future, the intention is to use the same principle shool of Biomedical Engineering, IIT(BHU)
of division of work in tiles to write an out-of-core version of Varanasi, Uttar Pradesh, India. His research
the segmentation algorithm. This version would allow the interests are Medical Image Processing,
segmentation of images that do not fit in main memory. Parallel Computing, Computer Vision.
Thus, it is expected that the image segmentation can handle
extremely large data efficiently and without requiring special
hardware. It is also expected to propose others parallel Dr Neeraj Sharma is associate professor
versions for different hardware like clusters and GPUs in school of Biomedical Engineering,
(Graphics Processing Units). IIT(BHU). His specialization area is
Biomedical Signal & Image Processing &
Bio instrumentation.
VI. APPLICATION IN MEDICAL IMAGING

Medical imaging require more computing power than a Dr Shiru Sharma is assistant professor in
traditional sequential computer can do and we also know school of Biomedical Engineering,
that for medical imaging, it is important that the image is IIT(BHU). His specialization area is
clear and be obtained as quickly as possible. We can achieve Biomedical Control System Analysis.
through the process of parallelizing. Parallelizing optimizes
the speed at which the image is produced. Normal a CT
scan, PET scan, etc take much time to produce the result so
by parallelizing our code we can find good result in very
reasonable time.

REFERENCES
[1]. Thomas Bräunl, “Tutorial in Data Parallel Image Processing,”
Australian Journal of Intelligent Information rocessing Systems (AJIIPS),
vol. 6, no. 3, 2001, pp. 164–174
[2]. MatlabMPI: http://www.ll.mit.edu/MatlabMPI/
[3]. A. K. Jain, Digital Image Processing. Prentice Hall, 1989.
[4]. R. Crane, A simplified approach to image processing: classical and
modern techniques. Prentice Hall, 1997.
[5]. http://en.wikipedia.org/wiki/Medical_imaging
[6]. J. Fung and S. Mann, “Using graphics devices in reverse: Gpu-based
image processing and computer vision," in 2008 IEEE International
Conference on Multimedia and Expo. IEEE, June 2008, pp. 9-12.
[7]. P. Pacheco, An Introduction to Parallel Programming. Morgan
Kaufmann, 2011.
[8]. A. G. López, J. Delgado, and S. Castañeda, “Metodologías de
paralelización en la supercomputadora cicese2000,” Departamento de
Cómputo Dirección de Telemática Centro de Investigación Científica y de
Educación Superior de Ensenada, Tech. Rep., February 2000.
[9]. R. T. Rasúa, “Algoritmos paralelos para la solución deproblemas de
optimización discretos aplicados a ladecodificación de señales,” Ph.D.

Copyright to IJARCCE www.ijarcce.com 1900