Associative Memory Model for Color Image Clustering

(1)

International Journal of Emerging Technology and Advanced Engineering

Website: www.ijetae.com (ISSN 2250-2459,ISO 9001:2008 Certified Journal, Volume 5, Issue 4, April 2015)

335

Associative Memory Model for Color Image Clustering

Manish Maheshwari

1

, Sindhu Shrivatri

2

Computer Department, Makhanlal Chaturvedi National University of Journalism & Communication, Bhopal, India Abstract— Pattern Association is the process of forming

association between related patterns that maps a set of input patterns to a set of output patterns. The pattern that has to be associated may be of same type or of a different type. Associative memory net can be seen as a simplified model of a human brain which stores and retrieves patterns by association. In this paper, an efficient new fuzzy model for association of color image is introduced. A new color quantization ordering scheme that focuses on color as feature and considers Hue-Value and Saturation (HVS) space is proposed. Using fuzzy if-then rules pixel color is quantized into 54 colors. Fuzzy histogram of these 54 colors is calculated and stored in feature database. We propose a simplified associative memory model to store associations between feature vectors.

Keywords—Associative Memory, Clustering, Fuzzy Set, HSV Color Model, Image Retrieval, Neural Network.

I. INTRODUCTION

Brain is the central organ responsible for the learning process taking place in an individual. Learning can be defined as the process of acquiring new concepts, knowledge, and understanding about the environment. One of the fundamental characteristics of the human brain is the capability of recognizing and classifying patterns. Human brain is capable to associate different memories to specific events.

A neural network is a processing device, whose design was inspired by the design and functioning of human brain and their components [1]. Neural network has many applications. The most likely applications for the neural networks are classification, Association and reasoning etc. A pattern recognition task involves an ability of the brain to categorize / classify objects, sounds, feelings or ideas such that the degree of association is high among structures of the same category and low between structures of different categories [2].

Association is one of the fundamental characteristics of the human brain. The human memory operates in an associative manner; that is a portion of recollection can produce an associated stream of data from the memory. The human memory can retrieve a full image from a partial or noisy version of the image as the query image.

Furthermore, given a query image as the input, the human brain can recall associated images that have been stored in the past. Associative Memories, one of the major classes of neural networks, are faint imitations of the human brain’s ability to associate patterns. Associative memory is such a fundamental and encompassing human ability (and not just human) that the network of neurons in our brain must perform it quite easily [3].

The rapid progress in computer technology for multimedia system has led to a rapid increase in the use of digital images. Rich information is hidden in this data collection which is potentially useful in a wide range of applications like crime prevention, military, home entertainment, education, cultural heritage, medical diagnosis, and World Wide Web [4, 5]. Exploring and analyzing these images and locate target images in response to user queries has become a significant problem.

In this paper we propose a data mining approach to cluster the images based on color feature. Concept of color histogram is used to obtain the features. RGB color space is converted to HSV color space. Based on hue, saturation and value, image is quantized to 54 colors and histogram of these 54 color is formed. These histogram values are input features of each image to the clustering algorithm. A simplified associative memory model is designed to cluster the images and results are compared with k means algorithm. The rest of paper is organized as follows: In section two we provide overview of image retrieval. Section three introduces the concept of clustering and section four about associative memory. In section five we present the proposed work and result of our experiments in section six.

II. IMAGE RETRIEVAL

(2)

International Journal of Emerging Technology and Advanced Engineering

336

An n-dimensional feature vector represent an image where n is the selected number of extracted attributes. Color information is the most intensively used feature for image retrieval because of its strong correlation with the underlying image objects. A commonly used one is the RGB space because most digital images are acquired and represented in this space However, due to the fact that RGB space is not perceptually uniform, color space such as HSV (Hue, Saturation, and Value), HSL (Hue, Saturation and Luminance), CIE L*u*v* and CIE L*a*b* tend to be more appropriate for calculating color similarities. Color Histogram [5, 6, 7] is the commonly and very popular color feature used in many image retrieval system. The mathematical foundation and color distribution of images can be characterized by color moments [8]. Color Coherence Vectors (CCV) has been proposed to incorporate spatial information into color histogram representation [9].

III. CLUSTERING

There are techniques such as clustering for unsupervised learning or class discovery that attempt to divide data sets into naturally occurring groups without a predetermined class structure. The cluster analysis is a partitioning of data into meaningful subgroups (clusters), which the number of subgroups and other information about their composition or representatives are unknown. Cluster analysis does not use category labels that tag objects with prior identifiers i.e. we don’t have prior information about cluster seeds or representatives. The objective of cluster analysis is simply to find a convenient and valid organization (i.e. group) of the data [10,11]. Intelligently classifying image by content is an important way to mine valuable information from large image collection. Reference [12] explore the challenges in image grouping into semantically meaningful categories based on low-level visual features. The SemQuery [13] approach proposes a general framework to support content-based image retrieval based on the combination of clustering and querying of the heterogeneous features. Reference [14] describe data mining and statistical analysis of the collections of remotely sensed image. Large images are partitioned into a number of smaller more manageable image tiles. Then those individual image tiles are processed to extract the feature vectors.

IV. ASSOCIATIVE MEMORY

An associative memory is a storehouse that allows one data item to be associated with another so that access to one data item allows access, by association to the other. An associative memory belongs to the class of single layer feedforward or recurrent network architect depending on its association capability.

There are two types of associative networks: Hetero Associative (HA) and Auto Associative (AA). HA networks are capable of making associations between two or more different types of input signals. For instance, a HA network may associate a verbal command with an image or text. An Auto Associative memory is used to retrieve a previously stored pattern that most closely resembles the current pattern. Such a network can learn various patterns, and then recall the pattern based on a fractional part. For instance, it could recall an original pattern based on a corrupt or partially missing pattern [15]. The Hopfield network [16] is an auto associative fully connected network of a single layer of nodes based on Hebbian learning. Bidirectional Associative Memory is hetero associative memory consisting of two layers. It uses the forward and backward information flow to produce an associative search for stored stimulus-response association [17].

Based on the principle of recall, associative memory models may be classified into static and dynamic networks. While static networks recall an output given an input in one feedforward pass, dynamic networks recall through an input/output feedback mechanism which takes time. Static networks are non-recurrent and dynamic networks are termed recurrent.

V. PROPOSED WORK A. HSV Color Model

(3)

International Journal of Emerging Technology and Advanced Engineering

337

[image:3.612.54.274.216.469.2]

The boundary of the hex cone represents the various hues and it is used at the top of the HSV hex cone. In the hex cone, saturation is measured along a horizontal axis and value is along the vertical axis through the center of the hex cone.

Fig. 1. HSV Color Model

Hue is represented as an angle about the vertical axis, ranging from 0 at red through 360 vertices of the hexcone are separated by 60 intervals. Yellow is at 60, Green at 120 and Cyan opposite red at H=180. Complementary colors are 180* apart.

Saturation(S) varies from 0 to 1. It is represented in this model as the ratio of the purity of a selected hue to its maximum purity as S=1. Value V varies from 0 at the apex of the hexcone to 1 at the top. The apex represents black.

B. Fuzzification of the HSV values

The hue values range from 0 to 360 degrees and hue represents the dominant color of a pixel. The fuzzification of hue is done in such a way that the non-crisp boundaries between the colors can be represented much better. Six symbols are used in order to characterize the hue values at the distance of 60 degree

Hue = {RED, YELLOW, GREEN, CYAN, BLUE, MAGENTA}

Every symbol is characterized by its membership function. The membership functions, which were used for the six symbols, are classical triangular functions as shown in Fig 2.

The saturation & value range from 0 to 1. Three symbols are used to characterize these quantities as shown in Fig.3.

Saturation = {Small, Medium, Large} Value = {Small, Medium, Large}

A. Rules for Histogram Calculation

In the proposed work hue, value and saturation values of each pixel are considered as the input for the calculation of histogram.

Small If value is Medium Large

and saturation is Large Medium Small Red and Hue is Magenta

Blue Yellow Cyan Green

then Color is CH1 CH2 CH3 : : : CH54

(4)

International Journal of Emerging Technology and Advanced Engineering

338

The normalization allows us to assign the maximum, unitary typicality to the color that is dominant within the image, regardless its probability of appearance (which is not the case of the usual histogram). Histogram of each color is obtained by dividing each histogram bin by the largest number of histogram bin in that image.

Fig.3. Fuzzy Sets used for Saturation & Value

B. Fuzzy histogram

Color histogram as a set of bins where each bin donates the probability of pixels in the image being of a particular color. A color histogram H for a given image is defined as a vector:

H = {H[o], H [1], H [2]…H[c]…H [N]} (1) Where c represents a color in the color histogram H[c] is the number of pixels in color c in that image, and, N is the number of bins in the color histogram, i.e., the number of colors in the adopted color model.

Typically, each pixel in an image will be assigned to a bin of a color histogram of that image, so for the color histogram of an image, the value of each bin is the number of pixels that has the same corresponding color. In order to compare image of different sizes color histograms should be normalized. The normalized color histogram H' is defined as:

H' = {H'[0], H'[1], H'[2] ...H’[c] ...H’ [N]} (2) Where H'[c] = H[c] / P

& P is the total number of Pixels in an image.

The value of each bin is thus the number of image pixels having the color c, or after normalization, is the probability that the color c appears in the image. From a numerical point of view H'[c] maps the color set C in the interval {0, 1}. According to Zadeh’s [18] theory, such a function is a fuzzy set. In order to fit semantically plausible description for the properties of the object, significance of the numbers attached to each color in the color space C and a method that assures that the numbers are well within the [0, 1] range, it requires slight modification in the construction of the normal color histogram.

The simplest approach is to normalize the histogram in "Eq.(2)" by the value of its largest bin, in such way that the most probable color will have a membership degree of 1 within the fuzzy set ―image‖ [19]. The most predominant color can be thus considered as the most typical for the given image and the constructed fuzzy histogram measures the typicality of a color within the image.

H" = {H"[0], H"[1], H"[2], ...H"[c], …H"[N]} (3) Where H"[c] = H[c]/Max (H[c])

The normalization by the model from "Eq.(3)" allows us to assign the maximum, unitary typicality to the color that is dominant within the image, regardless its probability of appearance ( which is not the case of the usual histogram)

A feature database is created by calculating the fuzzy histogram of these eight colors. Neural network is trained by giving eight fuzzy histogram value of the query image.

C. K Means Clustering Algorithm

K-means is one of the simplest unsupervised learning algorithms in which each point is assigned to only one particular cluster. The procedure follows a simple, easy and iterative way to classify a given data set through a certain number of clusters (assume k clusters) fixed a priori. The procedure consists of the following steps:

Step 1: Set the number of cluster k Step 2: Determine the centroid coordinate

Step 3: Determine the distance of each object to the centroids

Step 4: Group the object based on minimum distance Step 5: Continue from step 2, until convergence that is no object move from one group to another.

D. Associative Memory Model

Most associative memory models insist on binary or bipolar pattern pairs. [20] Proposed Bidirectional Associative Memeory network to associates patterns that are real coded. It is a single-layer network which does not require training.

In our work output of fuzzy histogram discussed above is already normalized, step 1 of BAM proposed in [20] is not require. After omiting step 1 algorithm for simplified Associative Memory Model (sAMM) is as follows:

sAMM(N, X, Y)

/* N is the numbers of images (X , Y) are the pattern set where X = (X1, X2, X3, ……..XN) and

(5)

International Journal of Emerging Technology and Advanced Engineering

339

/* Feature vector Xi = (Xi1, X i2, X i3, ……..X ip) and

Output Yi = (Yi1, Yi2, Yi3, ……..Yim) */

Step 1: Input A (Ai1, A i2, A i3, ……..A ik) the feature

vector whose associated pair is to be recalled. Zi = A  Xi /*  is the inner produt operator */

End

Step 2: Compute the inner product of A with X for i = 1 to N

Let Z = (Zi) i = 1,2….N

Step 3: Apply threshold function f on Z to obtain the correlation vector M

M = f(Z) = m1, m2, m3, ….mN

where f is equal to one for maximum of Z and zero elsewhere.

Step 4: Output Yk where k is such that

mk = max (mi) where k= 1,2,….p

VI. EXPERIMENTS

The proposed scheme has been performed using a image database of 1000 images including 10 classes, which can be downloading from the website http://wang.ist.psu.edu/iwang/test1.tar. Each class has 100 images. Each image is of size 384*256 pixels. The system is developed in Matlab. Steps to perform the work as follows:

 Convert Image from RGB to HSV color space

 As discussed in section V.C, using HSV values, convert pixel value to CH1 to CH54

 For each Image count CH1 to CH54 to calculate 54 color histogram.

 Apply Associative memory algorithm discussed in section V.F to group the images

 Calculate Recall and Precision

Based on commonly used performance measures in information retrieval, two statistical measures were computed to assess system performance namely Recall and Precision.

Recall consists of the proportion of target images that have been retrieved among all the relevant images in the database.

Recall = Number of Relevant Images Retrieved Total Number of Relevant Images

Precision consists of the proportion of relevant images that are retrieved.

Precision = Number of Relevant Images Retrieved Total Retrieved Images Table I and II shows the values of recall and precision of each classes. Figure 4 showing the sample images in Dinosaurs, Bus and Horses cluster.

TABLE I

RECALL OF K-MEANS & ASSOCIATIVE MEMORY

TABLE II

PRECISION OF K-MEANS & ASSOCIATIVE MEMORY

Classes Precision % K-Means Associative

Memory African People

and villages 41 37

Beaches 42 49

Buildings 32 37

Buses 51 49

Dinosaurs 84 100

Elephants 40 41

Flowers 77 95

Horses 78 94

Mountains and

glaciers 39 61

Food ₂₃ ₃₂

Classes Recall % K-means Associative

Memory African People

and villages 74 52

Beaches 25 65

Buildings 28 52

Buses 36 85

Dinosaurs 100 89

Elephants 27 34

Flowers 51 75

Horses 99 15

Mountains and

Glaciers 60 30

[image:5.612.346.572.255.464.2] [image:5.612.347.572.446.674.2]

(6)

International Journal of Emerging Technology and Advanced Engineering

[image:6.612.44.283.134.573.2]

340 Fig. 4. Sample clusters using Associative Memory

(I Dinosaurs, II Bus, III Horses )

VII. CONCLUSION

In image retrieval system, the content of an image can be expressed in terms of different features such as color, texture and shape. These low-level features are extracted directly from digital representations of the image and do not necessarily match the human perception of visual semantics. We proposed a framework of unsupervised clustering of images based on the color feature of image. Concept of color histogram is used to obtain the features.

RGB color space is converted to HSV color space. Based on hue, saturation and value, image is quantized to 54 colors and histogram of these 54 color is formed. A simplified associative memory model is designed to cluster the images.

REFERENCES

[1] Y.P.Singh, V.S.Yada, A.Gupta, A. Khare, ―BiDirectional Associative Memory Neural Network method in the character Recognition‖, JATIT 2009, pp 382-386

[2] Gose E. E., J. W. bacus and L. Ackerman, ―A Comparison of some computer-measured and human-measured pattern recognition properties,‖ Journal of Cybernetics, 1 (1971), 68-74

[3] Timoty J. Dasey and Evangelia Micheli-Tzanakou, ―Fuzzy Neural Networks,‖ CRC Press LLC (2000), 135-162

[4] H.J.Zhang et al., ―Video Parsing, Retrieval and Browsing: an Integrated and Content-Based Solution‖, Proc. ACM Multimedia 95, San Francisco, Nov 95

[5] B.Furht, S.w.Smoliar, and H.J.Zhang, ―Image and Video Processing in Multimedia systems, kluwer Academic Publishers, Norwell MA, 1995

[6] Wayne Niblack, Ron Barber, William Equitz, Myron Flickner, Eduardo H. Glasman, Dragutin Petkovic, Peter Yanker, Christos Faloutsos, Gabriel Taubin: ―The QBIC Project: Querying Images by Content, Using Color, Texture, and Shape‖, Storage and Retrieval for Image and Video Databases (SPIE) 1993: 173-187

[7] Alex Pentland, Rosalind W. Picard, Stan Sclaroff, ―Photobook: Tools for Content-Based Manipulation of Image Databases‖, Storage and Retrieval for Image and Video Databases (SPIE) 1994: 34-47 [8] M.Stricker and M.Orengo, ―Similarity of color images‖, Storage and

Retrieval for Image and Video Databases III (SPIE) 1995: 381-392 [9] Greg Pass, Ramin Zabih, Justin Miller: Comparing Images Using

Color Coherence Vectors. ACM Multimedia 1996: 65-73

[10] Han and M.Kamber, ―Data Mining concepts and Techniques‖, Morgan Kaufmann Publishers, 2002

[11] A.K.Pujari, ―Data Mining Techniques‖, University Press, 2001 [12] Y. Uehara, S. Endo, S. Shiitani, D. Masumoto, and S. Nagata, ‖A

computer-aided Visual Exploration System for Knowledge Discovery from Images‖, In Proceedings of the Second International Workshop on Multimedia Data Mining (MDM/KDD'2001), San Francisco, CA, USA, August, 2001.

[13] Gholamhosein Sheikholeslami, Wendy Chang, Aidong Zhang, ―SemQuery: Semantic Clustering and Querying on Heterogeneous Features for Visual Data‖, IEEE Trans. Knowl. Data Eng. 14(5): 988-1002 (2002)

[14] Krzysztof Koperski, Giovanni Marchisio, Selim Aksoy, and Carsten Tusk, "Applications of Terrain and Sensor Data Fusion in Image Mining", IEEE 2002, pp 1026-1028

[15] J.A. Starzyk, ―Associative Learning in Hierarchical Self-Organizing Learning Arrays, IEEE Trans on Neural Network 2006

[16] J. J. Hopfield, ―Neural Networks and Physical Systems with Emergent Collective Computational Abilities‖, Proceedings of the National Academy of Sciences, 79: 2554-2558, 1982.

[17] B. Kosko, ―Adaptive bidirectional associative memories,‖ IEEE Trans. on Systems, Man, and Cybernetics, vol. 18, no. 1, pp. 49–60, 1988.

Cluster I

Cluster II

Cluster III

(7)

International Journal of Emerging Technology and Advanced Engineering

341 [18] L.A.Zadeh, "Fuzzy logic, neural networks, and soft computing,"

ACM 37:77-84, 1994.

[19] J.C.Bezdek, "Fuzzy models - what are they and why?" IEEE trans. On Fuzzy systems, 1(1): 1-5, Feb 1993