Modified Self Organizing Feature Map Neural Network (MSOFM NN) Based Gray Image Segmentation

(1)

Procedia Computer Science 54 ( 2015 ) 671 – 675

Peer-review under responsibility of organizing committee of the Eleventh International Multi-Conference on Information Processing-2015 (IMCIP-2015) doi: 10.1016/j.procs.2015.06.078

ScienceDirect

Eleventh International Multi-Conference on Information Processing-2015 (IMCIP-2015)

Modiﬁed Self Organizing Feature Map Neural Network

(MSOFM NN) Based Gray Image Segmentation

Pankaj Upadhyay

∗

and Jitendra Kumar Chhabra

Department Computer Engineering, NIT Kurukshetra, Kurukshetra 136 119, India

Abstract

Image segmentation is the most important and crucial part of image analysis system. The accuracy of any image analysis system is highly dependent on the accuracy of image segmentation. In this paper, we propose an image segmentation method that is based on Modified Self Organizing Feature Map. The proposed method classifies image pixels based on their intensity values for image segmentation. In proposed method, there is no burden of feature extraction and training data set. First the image is converted into one dimensional vector (input vector) and image pixel intensity values directly feed to the input layer of SOFM NN. To classify each pixel class, we modified the SOFM NN by adding an extra layer of neurons. This extra layer does not contribute in the process of weight updation. After the termination of self organizing network and weights updation process, this extra layer of neurons checks for input intensities to closest match and credited that input intensity to concerned class. The proposed algorithm is tested on the publicly available Berkeley Segmentation Data set. The results clearly show that the algorithm present in this paper is better than existing algorithms

Peer-review under responsibility of organizing committee of the Eleventh International Multi-Conference on Information Processing-2015 (IMCIP-2015).

Keywords: Classiﬁcation; Image segmentation; Self organizing feature map neural network unsupervised.

1. Introduction

Image segmentation is crucial and most important part of low level vision. Segmentation is a process of dividing the image into multiple distinct parts containing each pixel with similar attribute. More precisely, we can deﬁne image segmentation as a process of labeling pixels in an image such that each pixel with the same label shares similar characteristic.

There are several approaches available for image segmentation. Traditional image segmentation methods are mainly based on the uniformity of image pixels feature values (intensity, color etc.)1, 2_{. Thresholding is an old and}

simple image segmentation technique, based on the global information (e.g. Histogram of the whole image) or local information of the image. However for most of the images it is very hard to ﬁnd out the exact value of threshold for good segmentation. Segmentation may be obtained using edge detection of various regions.

To produce meaningful segments region segmentation results and edge detector output can be fused3. Many graph theoretic techniques also have been proposed for cluster (segments) analysis. The clustering technique presented in4_is

∗_{Corresponding author. Tel.: +08950156401.} E-mail address: [email protected]

Peer-review under responsibility of organizing committee of the Eleventh International Multi-Conference on Information Processing-2015 (IMCIP-2015)

(2)

Fig. 1. (a) Graphical presentation of SOFM14; (b) Neighborhood nodes.

based on network ﬂow theory. Geometric level set methods for image segmentation given by5. Normalized cut based image segmentation method proposed in6. In6image segmentation was treated as a graph partitioning problem and used normalized cut criterion which measure the inter-similarity and intra-similarity between the groups. Mohamed

et al.8_{has given fuzzy c-means method for MRI image segmentation.}

Neural network based method is an attempt to achieve robustness towards the random noise and to have real time output. Huge connection architecture gives the system robustness while parallel processing provides real time functioning. In literature, there are various algorithm based on neural networks10–13for image segmentation. Neural network based image segmentation methods are usually designed for particular application. Self organizing feature map (SOFM) presented in9_{has also been used for image segmentation. In}7_{authors proposed SOFM neural network}

based color image segmentation. Traditional self organizing feature map neural networks are very powerful clustering method but it requires prior image information (large training data set, feature extraction etc.). Statistical and neural network based methods have a major drawback that they require prior knowledge about the images & performance of these techniques mostly depends on the prior information. At the same time large amount of sample space (training data) is required for these approaches.

We propose an image segmentation method, which used Modiﬁed Self-organizing Feature Map Neural Network (MSOFM-NN) approach. The proposed technique is unsupervised in nature which removes the limitations of existing segmentation techniques. In this proposed method there is no need of prior knowledge (no. of segments) of the image and feature extraction. Proposed method is very useful when we don’t have much data and cluster information. This method provides the power of SOFM without any prior information about images.

2. Methodology

2.1 Self organizing feature map (SOFM)

Kohone9, given an algorithm which generates self organizing feature maps, that mimic the human brain’s self organizing map. The self-organizing feature map is basically a competitive vector quantizer to which real valued input presented sequentially to a linear or planer array of nodes. The basic model made up of two layers. The ﬁrst layer consists of input nodes and the second one consists of output nodes. The output nodes make a two dimensional lattice as shown in Fig. 1(a). Every input node is connected to every output node through adaptable weights.

The input vector Xk = (x₁k, x₂k, . . . , x_nk) ∈ nis feed to a (v × v) lattice of nodes. Because of the planer nature each node will be recognized by row-column index i j , where i, j = 1, 2, . . . , v. The i jthnode has an incoming weight vector W_{i j}k = (wk₁_,ij, wk₂_,ij, . . . , wk_n,ij) ∈ nat kthinput. The first step is to compare each weightwi j with input xi and find out the winning node, and therefore define a neighborhood around the winning node (Fig. 1(b)). The winning node can be found by comparing Euclidian distance.

The measurement of di j is a Euclidean distance deﬁned as:

di j = min

j xi− wi j

2

The criterion for winner is having minimum Euclidian distance. The weight adjustment is done only for output winning neuron and its neighborhood. A neighborhood may be rectangular, or a hexagonal array of nodes around the best

(3)

Fig. 2. Automatic image segmentation framework.

matching unit (winner). The size of neighborhood decreases with time. It require learning rate also, that decreases with time.

Algorithm can be described as follows:

Step 1: Initialize the weights to some small numbers, size of the neighborhood and learning rate. Step 2: Input a stream of training vector.

Step 3: Find out the winning neuron (i j ) on the basis of Euclidian distance.

Step 4: Update the weights of winning neuron i j and its all neighboring neurons, the neighborhood deﬁned by NI J. Weights update as follows:

W_{i j}k+1 = W_{i j}k + ηk(Xk− Wi jk), where j ∈ NI Jk Step 5: update learning rate and neighborhood size.

Step 6: Go back to step 2 until there is no observable change in map.

2.2 Proposed method

In proposed method, there is no burden of feature extraction and training data set. First the image is converted into one dimensional vector (input vector) and then pixel intensity values are directly fed to the input layer of SOFM NN. To classify each pixel class, we modiﬁed the SOFM NN by adding an extra layer of neurons. This extra layer does not contribute in the process of weight updation. After the termination of self organizing network and weights updation process, this extra layer of neurons checks for input intensities to closest match and credited that input intensity to concerned class. The framework for automatic image segmentation is shown in Fig. 2.

Modiﬁed Self Organizing Feature Map Algorithm is given as follows: Input: Gray scale image, Output: Segmented image

Step 1: Randomized the initial weights of SOM’s neurons.

Step 2: Feed the I/P image intensity vector X(t) to input layer of SOM. Step 3: Traverse each node of SOM

a. Find the similarity between input vector and SOM’s node using Euclidian distance. b. Choose the node with smallest distance (best matching node).

Step 4: Update the weights of SOM’s best matching node and its neighbors. Step 5: Repeat step 2–4 up to max. iterations.

Step 6: Map I/P intensities to nearest SOM’s node using additional layer of neurons. 3. Experimental Results

The proposed algorithm is implemented on a personal computer with dual core CPU of 2.30 GHz using MATLAB 8.0 programming language. Fuzzy C-Means clustering methods for image segmentation gives the promising

(4)

Table 1. The results of proposed method with comparison to FCM.

Image Method AE NED SNR

Image 1 MSOFM 11.2711 0.0664 1.5975 FCM 14.6000 0.0762 0.4027 Image 2 MSOFM 7.6039 0.0551 3.2252 FCM 8.6208 0.0681 2.8560 Image 3 MSOFM 7.2048 0.0536 3.4620 FCM 10.3853 0.0644 1.8697

Fig. 3. Segmentation results with modiﬁed self organising map NN(MOSFM) and fuzzy C-means methods(FCM).

results so proposed method is compared with Fuzzy C-Means segmentation method. We test our algorithm on the publicly available Berkeley Segmentation Data set15. Proposed method achieves better segmentation results compared to Fuzzy C-Means segmentation method Fig. 3 shows the segmentation results, using proposed algorithm and Fuzzy C-Means segmentation method. We also compared our result with Fuzzy C-Means segmentation method using AE, NED, and SNR. The deﬁnitions of AE, NED, and SNR are given as follows. If we deﬁne

I(i, j), i = 1, 2, . . . M and j = 1, 2, . . . N the input image and O(i, j) the output image, the criteria are deﬁned as

follows: a. Average Error(AE) A E= 1 M.N M i=1 N j=1 (I (i, j) − O(i, j))2

b. Normalized Euclidean Distance(NED)

NED= 1 M.N M i=1 N j=i (I (i, j) − O(i, j))2

c. Signal to Noise Ratio(SNR)

SNR= 10 log₁₀ ⎡ ⎢ ⎢ ⎢ ⎢ ⎣ M i=1 N j=1 I2(i, j) M i=1 N j=i (I (i, j) − O(i, j))2 ⎤ ⎥ ⎥ ⎥ ⎥ ⎦

(5)

4. Conclusion

Traditional segmentation algorithms need the prior information about segments. The proposed algorithm work well where we don’t know about the number of Segments (regions). In Proposed method, there is no need of feature extraction and training data set of images. The proposed method is not image speciﬁc; it can be used for different kind of images. Proposed method achieves better segmentation results compared to Fuzzy C-Means segmentation method. The proposed method requires higher computational time in comparison to existing techniques. So in future work, we will try to reduce the computational time.

References

[1] R. M. Haraliek and L. G. Shapiro, Image Segmentation Techniques, In CVGIP, pp. 100–132, vol. 29, (1985).

[2] N. R. Pal and S. K. Pal, A Review on Image Segmentation Techniques, Pattern Recognition, vol. 26, pp. 1277–1294, (1993).

[3] Yu Xiaohan and Juha Yia Jaaski, A New Algorithm for Image Segmentation Based on Region Growing and Edge Detection, IEEE

International Symposium on Circuits and Systems, vol. 1, pp. 516–519, (1991).

[4] Z. Wu and R. Leahy, An Optimal Graph Theoretic Approach to Data Clustering: Theory and it’s Applications to Image Segmentation, IEEE

Transaction on Pattern Analysis and Machine Intelligence, vol. 15, no. 11, pp. 1101–113, (1993).

[5] S. Osher and N. Paragios, Geometric Level Set Methods in Imaging Vision and Graphic, Springer Verlag, (2003).

[6] Jianbo Shi and Jitendra Malik, Normalized Cuts and Image Segmentation, IEEE Transaction on Pattern Analysis and Machine Intelligence, vol. 22, no. 8, pp. 888–905, (2000).

[7] Jun Zhang and Jinglu Hu, An Automatic Segmentation Technique for Color Image Segmentation based on SOFM Neural Network, IEEE

Conference on Neural Network, (2009).

[8] Mohamed N. Ahmed, Samesh M. Yamany, Nevin Mohamed, Aly A. Farag and Thomas Moriarty, A Modiﬁed Fuzzy C-Means Algorithm for Bias Field Estimation and Segmentation of MRI Data, IEEE Transaction on Medical Imaging, vol. 21, no. 3, pp. 193–199, (2002). [9] Teuvo Kohonen, The Self-Organizing Maps, Proceeding of the IEEE, vol. 78, no. 9, pp. 1464–1480, (1990).

[10] A. Ghosh, N. R. Pal and S. K. Pal, Image Segmentation using Neural Networks, Biological Cybernetics Springer Verlag, vol. 66, pp. 151–158, (1991).

[11] Sugata Ghosal and Rajiv Mehrotra, Application of Neural Networks in Segmentation of Range Images, Pattern Recognition, vol. 28, no. 51995, pp. 711–727.

[12] Javed Alirezaie, M. E. Jarnigan and C. Nahmias, Neural Network Based Segmentation of Magnetic Resonance Image of Brain, IEEE

Transactions on Nuclear Science, vol. 44, no. 2, pp. 194–198, (1997).

[13] T. Otani, K. Sato, H. Madokoro and A. Inugami, Segmentation of Head MR Images using Hybrid Neural Networks of Unsupervised Learning,

Proceeding of the IEEE International Conference (IJCNN), pp. 1–7, (2010).

[14] http://i.stack.imgur.com/IqbYZ.gif