Parallel processing technique for high speed image segmentation using color

Proceedings of the International Conference on Emerging Trends in Engineering and Management (ICETEM14)
30 – 31, December 2014, Ernakulam, India
118
PARALLEL PROCESSING TECHNIQUE FOR HIGH
SPEED IMAGE SEGMENTATION USING COLOR
Rasiq S.M., S. Krishnakumar
School of Technology and Applied Sciences, Mahatma Gandhi University Regional Centre, Edappally, Kochi-24, India
ABSTRACT
In this work, we introduce a novel method for segmenting a color image at a very high speed. The system is
based on self learning high speed parallel processing devices. The system processes video streams at speed of 1000
frames per second or more. For high speed image segmentation using sequential computing from an image of a video
having thousands of frames per second and each image frame consists of thousands of pixels, we need very much time
for executing complicated algorithms. In the traditional way of computing and segmentation systems are very time
consuming compared to our system because the traditional systems use sequential computation for segmenting, with
some complicated functions. If we use other types of parallel processors like ANN forprocessing each pixel or group of
pixels, those systems need programming and giving data to such large number of processors are practically difficult.
Here we have used a self learning parallel processor device which is made for doing some kinds of particular jobs. This
parallel processing devices are easy to manipulate and can be trained simultaneously. It contains memory for storing data
comparators for comparing with previously stored memory etc. Training as well as functioning is in real time even if the
system process thousands of image frames per second.
Keywords: Image Segmentation, Parallel Processors, Self Learning.
I. INTRODUCTION
Traditional ways of segmentation of an image use sequential or complicated computing [1, 3].Most of the
existing work, see e.g. [5,6], is based on the sliding window approach, where detection windows of various scales and
aspectratios are evaluated at many positions across the image. This approach becomes computationally very expensive
when rich representations are used.
These processes are very time consuming. Here we have designed a device, which can be used for parallel
processing, for fast computing and fast object recognition systems. This device is easy to manipulate. It learns to
compute from its previously learned memory. For parallel processing a large number of such devices can be used.
All devices which are working parallel can be simultaneously learned from an experience. After certain
number of learning or experiences a device can produce an output high (1) or low (0) from an input variable. The device
checks whether the particular variable value is in a particular range and produce the output. The range has an upper limit
and a lower limit. If the particular variable has a value within the range then the device output will be high. Learning is
the function of finding this upper limit and lower limit for a particular variable.
The self learning equipment like Perceptron needs more number of learning sets than our device and it is
complicated. The Perceptron manages more than one variable. It will also produce an output high or low depending on
the inputs, weights and threshold value [6]. Here the input may be affected by some noises. And these are used only for
some specific applications. Our device processes a single variable and it learns to compute the output by the help of a
trainer which will tell the device whether the output is high or low for a particular variable value.
INTERNATIONAL JOURNAL OF ELECTRONICS AND
COMMUNICATION ENGINEERING & TECHNOLOGY (IJECET)
ISSN 0976 – 6464(Print)
ISSN 0976 – 6472(Online)
Volume 5, Issue 12, December (2014), pp. 118-123
© IAEME: http://www.iaeme.com/IJECET.asp
Journal Impact Factor (2014): 7.2836 (Calculated by GISI)
www.jifactor.com
IJECET
© I A E M E

119
Here we can see the simple diagram of the device and its working, the device learning algorithm, the gray level
recognition and color recognition methods,
II. PROCESSOR ANDMETHOD
A processor will learn to produce an output from its previous memory. Learning is the function of finding the
values of lower limit k and upper limit k`, if the inputs X(n)and the corresponding out puts Y(n) are given for learning.
Figure 1 shows the diagram of a self learning device or processor. We can use op-amps as comparators, L and H as high
resolution analog memory locations [2] These L and H are used for storing the values of lower limit k and the upper
limit k`. We have used three AND gates in which two of them are one input bubbled. The one input bubbled AND
gates are used for generating memory ( L and H locations) enable signals. When memory enable signal becomes high,
the corresponding analog memory locations stores the value of Xm at that time if and only if the value of Ym is high in
that training period. Here Xm and Ym are the element of a training set.
After training completed the device will produce an out put Y`=1 if the analog input X` is within the lower
limit k and the upper limit k`, if X` is beyond the limit k and k` the device output Y` will be zero.
2.1. Learning Algorithm
1. initialize k=Xmax and k`=X min
2. mth
training step:check whether Ym=0 or 1
If Ym =1 and k ≥ Xm store Xm to the location L
If Ym=1 and k` ≤ Xm store Xm to the location H
3. Go to step 2 until the learning process is completed
4. decrease k= k- ∆X and increase k`= k`+∆X
End
Fig 1: Self learning device
2.2 Example for a learning procedure
Here it is given an example of finding the values of k an k` from a set.
Example : Find k and k’ for the set {(8,1),(6,1),(4,0),(14,0),(12,1)}, Let k=15 and k`=0;

120
Table 1: Learnig of the device
Trail
(n)
X(n) Y(n) k k`
1 8 1 8 8
2 6 1 6 8
3 4 0 6 8
4 14 0 6 8
5 12 1 6 12
suppose ∆X = 0.1, then
k = k-∆X
= 6-0.1 = 5.9
k`= k`+∆X
= 12+0.1 = 12.1
By observing the above table we can understand the learning procedure from the 5 trails. Figure 2 shows a
representation of the trained device.
Fig 2: A representation of trained device
The device gives most accurate output if it is trained with correct set of X,Ypairs. If one of the elements Ym in a
pair in the training set is not correct and variable Xm is beyondthe expected limits k and k`, then the pair in that set is not
suited for learning. And if one of the elements Ym in a pair in the training set is not correct and variable Xm is within the
expected limits k and k`, then the pair in that training set may be suited
We may be confused that why the machine knowledge (k and k`) are in a range. The answer is simple that
almost all the variables are affected by some errors or noises, getting the exact values are difficult. That is why we use
this kind of device.
III. GRAY LEVEL RECOGNITION
The object in an image may be a region with certain color, texture or some particular gray level value (for a
gray level image), our device can be used to identify that gray level value. Consider the image is a matrix having m
number of rows and n number of columns then we need mn number of devices. From the training images the devices
will learn the gray level intensity value of that particular object. Thus after the learning process is completed if we give
an input image having the same object then all the outputs of the devices representing the object become high.
IV. COLOR RECOGNITION
Here it is explained how a region with certain color can be detected. In a color image each pixels are having the
primary colors red, green and blue. If the device is considered as a block we can give each primary color as an input to
the device as shown figure 3.
Fig 3: Basic color matching device

121
Using the training set of a primary color, the device will learn to identify the intensity of that primary color.
After training the device, it will produce an output high when that particular primary color intensity value is given as its
input. Then the three devices representing three primary colors are given to an AND gate as shown in figure 3 and its
output will give whether the color matches or does not match. Consider the image is a matrix having m number of rows
andn number of columns then we need 3mn number of devices. After training is completed if we are giving an input
image with the same colored object then all the outputs of the AND gates representing the object become high.
V. RESULT AND DISCUSSION
The term self learning is used in our parallel Processors is due to the fact that with the help of a training system
the device can study to compute the output. For gray level image object recognition the best result will be obtained if we
are processing the image having an object with a light gray level background. But for color images, the best result will
be obtained if the object has different color from the surroundings. The image should be taken from a scene with almost
uniform light intensity. Here we have used an object with a single gray level intensity or a single colored object. So we
need to teach only one device for gray level object and only three devices for a colored object. The trained device or
devices will pass the gray level intensity or color to all other devices in the image matrix.
Here we have shown some simulated results for segmenting images using colors.
Fig 4: yellow object
Fig 5: gray object
Fig 6: yellow object

122
Fig 7: light blue object
Here images figure 4 to figure 7 are in png format. We are not using images in the jpeg format for simulation
due to blocking artifacts [4]. Let us see some examples in figure 8 and 9.
Fig 8: red letters
Fig 9: red object
VI. CONCLUSION
The color, length, breadth and shape of an object can be analyzed using this method and also recognized easily.
It can be used for identifying regions having a particular texture in an image. It is very easy to manipulate. A large
number of devices are required for image processing application for parallel processing. But for some application
related to artificial intelligence and machine learning, a few numbers of devices are only needed. It can also be used as a
logical gate whose output will be high for the input in a range and its output will be low for the input outside the range.
The noises in a binary image can be eliminated by using our processors if we use a system with a two input AND gate
and two processors.
It is easy to segment an image using color by giving RGB matrices to the processor array.

123
REFERENCES
[1] B. Catanzaro, B. S. amd N. Sundaram, Y. Lee, M. Murphy and K. Keutzer. (2009) Efficient, high-quality
image contour detection. In ICCV.
[2] Chris Diorio, SunitMahajan, Paul Hasler, Bradley Minch, Carver Mead,(1995), A High- Resolution
Nonvolatile Analog Memory Cell, California Institute of Technology Pasadena, California 91125 (818)
395-6996.
[3] N. Dalal and B. Triggs. ( 2005 ),Histograms of oriented gradients for human detection. In CVPR.
[4] Y. L. Lee, H. C. Kim, and H. W. Park, (1998) ,"Blocking effect reduction of JPEG images by signal adaptive
filtering", IEEE Trans. Image Processing, vol. 7, pp.229 -234
[5] P. Felzenszwalb, R. Grishick, D. McAllester, and D. Ra- manan. (2010) Object detection with discriminatively
trained part based models. PAMI, 32(9),
[6] Radford M. Neal (1995), Bayesian Learning for Neural Network, Department of Computer Science, University
of Toronto.
[7] Gunwanti S. Mahajan and Kanchan S. Bhagat, “Medical Image Segmentation using Enhanced K-Means
and Kernelized Fuzzy C- Means”, International Journal of Electronics and Communication Engineering &
Technology (IJECET), Volume 4, Issue 6, 2013, pp. 62 - 70, ISSN Print: 0976- 6464, ISSN Online:
0976 –6472.
[8] Gaganpreet Kaur and Dr. Dheerendra Singh, “Pollination Based Optimization for Color Image Segmentation”,
International journal of Computer Engineering & Technology (IJCET), Volume 3, Issue 2, 2012, pp. 407 - 414,
ISSN Print: 0976 – 6367, ISSN Online: 0976 – 6375.

Parallel processing technique for high speed image segmentation using color

More Related Content

What's hot (20)

Viewers also liked (17)

Similar to Parallel processing technique for high speed image segmentation using color (20)

More from IAEME Publication (20)

Recently uploaded (20)

Parallel processing technique for high speed image segmentation using color