Comparative Study of Image Segmentation using Variants of Self Organizing Maps (SOM)
Sanchari Sengupta
1, Sonal Verma
2, Srishti Mull
3and Sourav Paul
41
Sanchari Sengupta, Computer Science & Engineering /Sikkim Manipal Institute Of Technology, Sikkim, India
1
[email protected]
2
Sonal Verma, Computer Science & Engineering /Sikkim Manipal Institute Of Technology, Sikkim, India
2
[email protected]
3
Srishti Mull, Computer Science & Engineering /Sikkim Manipal Institute Of Technology, Sikkim, India
3
[email protected] and [email protected]
4
Sourav Paul , Assistant Professor, Computer Science & Engineering /Sikkim Manipal University, Sikkim, India
4
[email protected]
ABSTRACT
Image segmentation is a very crucial step in the field of image processing which helps us to simplify the representation of the image, to make it easier to analyze. This paper deals with the comparison of image segmentation techniques based on unsupervised artificial neural network technique, known as Kohonen’s Self Organizing Maps (SOM). We first present image segmentation using Kohonen’s Self Organizing Map. Secondly, we focus on the Threshold Self Organizing Map (TSOM) and finally multilevel Self Organizing Maps. We then evaluate each of the outputs based on different image segmentation parameters.
Finally, we compare the parameters of each output and present our conclusion.
Keywords — Artificial Neural Network, Unsupervised, Segmentation, Self-Organizing Maps.
1. INTRODUCTION
Segmentation is an essential part of image processing technique. Image segmentation divides the image into several parts making it more meaningful and easier to analyze.
Artificial Neural Network (ANN) has been used for image segmentation based on its supervised and unsupervised methods. In case of supervised methods, some prior knowledge is required for successful segmentation of the image. This prior information may not always be available, whereas the unsupervised method provides a way to segment an image without having all the prior information about the image. Self-Organizing Map (SOM) is an unsupervised artificial neural network technique that is used to produce a low dimensional representation of the input space. SOM differs from other ANN techniques as it preserves the topological properties of the input space using neighborhood
and Mapping. Training builds the map from the input and Mapping classifies a new input vector.
2. KOHONEN’S SELF ORGANIZING MAP
The artificial neural network introduced by the Finnish professor Teuvo Kohonen in the 1980s is sometimes also known as the Kohonen map or network. It contains of components, namely nodes or neurons. Each of these nodes has some weight vector of the same dimension as the input space associated with it. SOM is used for visualizing low- dimensional views of a higher dimensional data set.
2.1. Architecture of Self Organizing Map (SOM)
The number of neurons in the first neural layer can be chosen in a task specific manner. Each neuron in the first neural layer has its own weight vector which is dimensionally equal o the input layer. Each neuron is connected to its adjacent neuron by a suitable neighborhood relation which determines the topology of the map. This neighborhood function is assigned by a special function also known as the topological neighborhood.
Fig 2.1: Architecture of SOM
2.2. Components of Self Organizing Map
Self-organizing process involves four major components:
2.2.1 Initialization: All the connection weights of the nodes are initialized with small random values.
2.2.2 Competition: The discriminant function, which forms the basis of the competition, is computed by the neurons for each input pattern. The neuron with the smallest value of the discriminant function is declared as the winner.
2.2.3 Cooperation: The winner neuron determines the spatial location of the topological neighborhood.
2.2.4 Adaptation: The excited neurons, i.e. the neighboring nodes of the winner neuron decrease the value of their individual discriminant function according to the input pattern through suitable adjustments associated with the connection weights.
Fig 2.2: Components of SOM
2.3. Self-Organizing Map Segmentation Process
Initially, all weight vectors of the first neural layer are set to random values. After that some input vectors D(t) from the input space are selected and set as an input for the neural network. Then the difference between the neuron vectors Wv and the input vectors is computed by traversing each node in the map using the Euclidean distance formula. The node with the least difference between the input vector and neuron vector is selected. This node is the winner neuron in correspondence to which the neighboring neurons adjust themselves such that they are closer to the input vector using the following formula:
Wv(s + 1) = Wv(s) + Θ(u, v, s) α(s)(D(t) - Wv(s))
Then Wv(s+1) and Wv are compared for detection of any significant changes. If there exists any change then again the neuron vectors are updated such that they are closer to the input vector. This process is carried out till the last iteration is reached.
Figure 2.3: SOM Segmentation Process
3. THRESHOLD SELF ORGANIZING MAP
Threshold SOM image segmentation is a two stage image segmentation technique that aims at eliminating the over segmentation problem of SOM process. In this technique the output of SOM segmentation is taken as the input for the threshold segmentation .The cluster centers are found after the SOM process and the threshold is set by the user .The distance
individual clusters. In the merging process the smaller cluster merges with the bigger cluster. This merging process reduces the number of clusters that were formed during SOM segmentation.
Figure 3.1: Merging process according to the distance between Cluster centers. [2]
Figure 3.2: T-SOM Segmentation Process
4. MULTILEVEL SELF ORGANIZING MAP
The Multi-Level Self Organizing map is a two stage process in which both the stages consist of the SOM process. Although SOM has been primarily used for dimension reduction, it also has applications in the field of image segmentation. Chang and Teng (2007) proposed a two stage SOM to overcome the noisy nature of various kinds of images such as medical images.[3]
Fig 4.1: Flowchart for Multilevel SOM
5. COMPARISON
The segmentation process was carried out in the standard data set, the Berkley’s image. Now the comparison of the above mentioned image segmentation techniques are carried out using the following parameters.
5.1 Time
The amount of time required to carry out the segmentation process is the first parameter that we have taken into consideration. The tic toc function in MATLAB is used to carry out the time estimation process.
5.2 Noise Immunity
It is the ability of the process to perform its operations when noise is present in the image. An image noise is the variation of the any information in an image and it usually pertains to any fluctuations in an electric signal. The peak-signal-noise- ratio (PSNR) technique has been used to find out the noise immunity of the input to the output. The higher the value of the PSNR, the more immune is the output to noise.
5.3 Spatial Information
Spatial Information is the physical location of object and the metric relation between the objects. The visualization technique has been used to compare the original image to the various outputs.
5.4 Object Connectivity
It is the measure of connected components or pixels of an image. Using the bwconncomp function in MATLAB, the number of components present in the image is found out.
6. INPUT OUTPUT
Fig 6.1: Original gray scale Image (Input Image)
Fig 6.2: SOM (Output Image)
Fig 6.3: HSOM (Output Image)
Fig 6.4: TSOM (Output Image)
7. CONCLUSION
After the analysis of different techniques used for image segmentation based on a type of artificial neural network known as Self Organizing Map, it is seen that the TSOM image segmentation technique was the most efficient technique amongst all the techniques used, as it was most immune to the noise. Moreover, the spatial information was highly preserved. However, depending on the nature of the application, the performance of the variants may vary, eg:
Multilevel SOM is very much suitable for medical image segmentation and SOM in comparison to TSOM and Multilevel SOM serve as a better option when time is a strict constrain. Finally, the techniques based on self-organizing maps are capable of maintaining spatial information of the image and are therefore preferred whenever preserving the topological features is a priority.
REFERENCES
[1] K. B. Vaishnavee, K. Amshakala, Study of Techniques used for Medical Image Segmentation Based on SOM,
[2] Mohamad Awad, An Unsupervised Artificial Neural Network Method for Satellite Image Segmentation, The International Arab Journal of Information Technology, Vol. 7, No. 2, April 2010
[3] Mostafa Jabarouti Moghaddam and Hamid Soltanian- Zadeh, Medical Image Segmentation Using Artificial Neural Networks, University of Tehran, Tehran
[4] A.M.Khan, Ravi.S, Image Segmentation Methods: A Comparative Study, International Journal of Soft Computing and Engineering (IJSCE), ISSN: 2231-2307, Volume-3, Issue-4, September 2013
[5] Simon K. Warfield*, Member, IEEE, Kelly H. Zou, and William M. Wells, Member, IEEE, Simultaneous Truth and Performance Level Estimation (STAPLE): An Algorithm for the Validation of Image Segmentation, IEEE Transactions on Medical Imaging, Vol. 23, No. 7, July 2004
[6] S. Chabrier, H. Laurent and B. Emile, Performance Evaluation of Image Segmentation. Application to Parameters Fitting. Laboratoire Vision et Robotique - UPRES EA 2078
[7] Nikita Sharma, Mahendra Mishra, Manish Shrivastava, Colour Image Segmentation Techniques and Issues: An Approach, International Journal of Scientific &
Technology Research Volume 1, Issue 4, May 2012 [8] Sourav Paul , Mousumi Gupta, Image Segmentation By
Self Organizing Map With Mahalanobis Distance , International Journal of Emerging Technology and Advanced Engineerin (ISSN 2250-2459, ISO 9001:2008 Certified Journal, Volume 3, Issue 2, February 2013) [9] Salem Saleh Al-amri, N.V. Kalyankar and Khamitkar
S.D, Image Segmentation by Using Threshold Techniques, JOURNAL OF COMPUTING, VOLUME 2, ISSUE 5, MAY 2010, ISSN 2151-9617
[10] R.Yogamangalam, B.Karthikeyan, Segmentation Techniques Comparison in Image Processing, International Journal of Engineering and Technology (IJET)
[11] E.J. Palomo, E. Dominiguez, R.M. Luque, J.Munoz, A New GHSOM Model Applied to Network Security,University of Malaga,Spain.