Full Paper

Int. J. on Recent Trends in Engineering & Technology, Vol. 05, No. 01, Mar 2011

ANFIS Based Color Image Segmentation for Extraction of Salient Features: A Design Approach
Ms. Anuja Bujurge1, Mr. Parag Puranik2, Dr. P. R. Bajaj3
G. H. Raisoni College of Engineering, Nagpur, India Email: anujabujurge@yahoo.co.in 2 G. H. Raisoni College of Engineering, Nagpur, India Email: paragvpuranik@gmail.com2, principal@ghrce.edu.in3
Abstract - Image segmentation is very essential and critical to image processing and pattern recognition. In this paper, a technique for color image segmentation called Adaptive Neuro-Fuzzy Color Image Segmentation (ANFIS) is proposed. Adaptive Neuro-Fuzzy system is used for automatic multilevel image segmentation. This system consists of multilayer perceptron (MLP) like network that performs color image segmentation using multilevel thresholding. Threshold values for detecting clusters are found automatically using Fuzzy C-means (FCM) clustering technique. Fuzzy entropy is used as a tool to decide the number of clusters. ANFIS uses saturation and intensity planes of HSV (hue, saturation, intensity) color space for segmentation. Neural network is employed to find the number of objects automatically from an image. This system does not require a priori assumptions whatsoever are made about the image (type, features, contents, stochastic model, etc.). Index Terms- Adaptive thresholding, Color image segmentation, Neuro-fuzzy system, fuzzy entropy and clustering.
1

I.INTRODUCTION Image segmentation is a process of dividing an image into different regions in which each is homogeneous but the union of any two adjacent regions is not homogeneous. In the past decades, attention had been paid to monochrome image segmentation and many algorithms had been proposed in the literature. Basically, color image segmentation algorithms are frequently based on monochrome image segmentation approaches operating in different color space. The color distinguished by human eye is a combination of basic colors such as red (R), green (G), and blue (B). Compared to monochrome image, a color image provides in addition to intensity the additional information (hue and saturation) in the image. In this paper, a system capable of performing adaptive multilevel color image segmentation based on thresholding and FCM clustering technique is proposed. Clusters and their labels are automatically found out using FCM clustering technique. The proposed Neuro-fuzzy segmentation system consists of a multilayer perceptron (MLP)-like network that performs image segmentation by adaptive thresholding of the input image using labels automatically pre-selected by a fuzzy clustering technique. The proposed architecture is feed 2011 ACEEE

forward, but unlike the conventional MLP, the learning is unsupervised. The output status of the network is described as a fuzzy set. Fuzzy entropy is used as a measure of the error of the segmentation system. One of the main roadblocks toward full automation of the segmentation system is the problem of automatically choosing the correct number of labels. This parameter, of crucial importance to most labelling methods, is usually very hard to determine automatically, and in most cases it is left as a parameter which has to be provided by the user. Artificial neural network (ANN) is a powerful computing technique which consists of number of interconnected and nonlinear computing elements. Its processing capability and nonlinear characteristics are used for classification and clustering the data. FCM is one of the well-known clustering technique. However, it has some drawbacks like adjacent clusters often overlap in color space, which causes incorrect pixel classification. The integration of neural networks and fuzzy logic has emerged as a promising field of research in recent years. This has led to the development of a new branch called neurofuzzy computing. Neuro fuzzy system combines the advantages of both the uncertainty handling capability of fuzzy systems and the learning ability of neural networks. The present work is an attempt to design an adaptive neurofuzzy system capable of performing multilevel segmentation of color images in HSV color space. Clusters and their labels are found automatically using FCM clustering technique. Neural network is used to detect multiple objects within an image. The network architecture is the same in principle to. It consists of five main layersinput, 3-hidden and output layers. Each layer consists of a fixed number of neurons equal to the number of pixels in the image. The activation function of neuron is multisigmoid. The number of objects in the image is found out automatically. The paper is organized as follows: Section II contains the image segmentation problem. Section III contains a Neurofuzzy approach. Section IV contains a detailed description of the proposed system. Finally, Section V and VI ends the paper with several conclusions and some intermediate experimental results drawn from the design of and the work with the proposed system.

19

DOI: 01.IJRTET.05.01.138

Full Paper
Int. J. on Recent Trends in Engineering & Technology, Vol. 05, No. 01, Mar 2011 II.IMAGE SEGMENTATION PROBLEMS Image segmentation is a process of dividing an image into different regions in which each is homogeneous but the union of any two adjacent regions is not homogeneous. In the segmented image, each object is labelled in a way that reflects the actual structure of the data and facilitates the description of the original image so that it can be interpreted by the system that handles the image further. Depending on whether spatially separated objects of the same kind have to be labelled the same or not, the image segmentation problem may be regarded as a classification problem or a clustering one, respectively. Several authors employed the following formal definition of image segmentation. Formal definition: Let F denote the grid of all the pixels in the image, i.e., the set of all the pairs: FMxN = {( i,j )}; i=1,2,....,N; j=1,2,.....,N; where M and N are the number of rows and columns of the matrix representing the image, and let P(.) be a uniformity (or homogeneity) predicate which assigns the value TRUE or FALSE to a nonempty subset of F, depending only on properties related to the value of the pixels in the subset. P(.) also has the property that given a subset of F, say Y, and a subset of Y, say Z, P(Y) = TRUE implies always that P(Z)=TRUE. A segmentation of the grid F for a uniformity predicate P is a partition of F into n disjoint nonempty subsets F1, F2,....,Fn such that: 1) Uni=1 Fi=F with (=> every pixel must be in one, and only one, segment); 2) Fi for i=1,2,....,n is connected (i.e., composed of contiguous grid points); 3) P ( Fi ) = TRUE for i=1,2,....,n; (the segments should be uniform, in terms of the chosen P); 4) P (Fi U Fj) = FALSE when Fi is adjacent to Fj. In what regards constraint above, the reader familiar with the problem might argue that some of the segmentation algorithms produce non connected (or non-convex) segments (e.g., classical thresholding algorithms). On the other hand, these non-convex segments may be treated as sets of several connected segments that were given the same label. More important, it should be noted that the definition above only defines what are the constraints on a possible segmentation of an image, and not what a correct segmentation of an image is. III.NEURO-FUZZYAPPROACH In this system, the proposed system use the ANIFIS algorithm for the color image segmentation. ANFIS implements a Takagi-Sugeno FIS and has a five layered architecture as shown in Figure 1. For first order Takagi-Sugeno model, a typical rule set with two fuzzy if-then rules can be expressed as Rule 1: If x is A1 and y is B1, then f1=p1x+q1y+r1 Rule 2: If x is A2 and y is B2, then f2=p2x+q2y+r2 As shown in ANFIS architecture nodes of the same layer have similar functions, as described below. (Here we denote the Output node i in layer l as Ol,i.) Layer 1: Every node i in this layer is an adaptive node with a node output defined by (1 ) where x (or y) is the input to the node and Ai (or Bi-2) is a fuzzy set associated with this node. In other words, outputs of this layer are the membership values of the premise part. Here the membership functions for Ai and Bi can be any appropriate parameterized membership functions. For example, Ai can be characterized by the generalised bell function: (2) where {ai, bi, ci}is the parameter set. Parameters in this layer are referred to as premise parameters. Layer 2: Every node in this layer is a fixed node labelled which multiplies the incoming signal and outputs the products. For instance, (3) Each node output represents the firing strength of a rule. (In fact, any other T-norm operators that perform fuzzy AND can be used as the node function in this layer.) Layer 3: Every node in this layer is a fixed node labelled N. The i-th node calculates the ratio of the i-th rules firing strength to the sum of all rules firing strengths: (4) For convenience, output of this layer will be called normalised firing strengths. Layer 4: Every node i in this layer is an adaptive node with a node function (5) where is the output of layer 3 and {pi, qi, ri} is the parameter set. Parameters in this layer will be referred to as consequent parameters. Layer 5: The single node in this layer is a fixed node labelled , which computes the overall output as the summation of all incoming signal: (6) ANFIS uses back-propagation learning to determine premise parameters (to learn the parameters related to membership functions) and least mean square estimation to 20

2011 ACEEE

DOI: 01.IJRTET.05.01.138

Full Paper
Int. J. on Recent Trends in Engineering & Technology, Vol. 05, No. 01, Mar 2011 determine the consequent parameters. A step in the learning procedure has got two parts: In the first part the input patterns are propagated, and the optimal consequent parameters are estimated by an iterative least mean square procedure, while the premise parameters are assumed to be fixed for the current cycle through the training set. In the second part the patterns are propagated again, and in this epoch, back-propagation is used to modify the premise parameters, while the consequent parameters remain fixed. This procedure is then iterated.

Fig. 1 Five Layered Architecture of ANFIS.

IV.PROPOSED SYSTEM A. Overview of the Proposed System The proposed system consists of a multilayer neural network that performs adaptive, multilevel thresholding of the image using labels automatically preselected by a fuzzy clustering technique. The learning technique employed is self supervised allowing, therefore, automatic adaptation of the NN. The output status of the network is described as a fuzzy partition. Fuzzy entropy is used as a measure of the error of the system as well as a criterion for determining potential edge pixels. Given an input image, the system is forced to evolve toward a minimum fuzzy entropy state in order to obtain image segmentation. Pixels most affected by the consecutive training iterations (due to the amount of their contribution to the fuzzy entropy of the system) are labelled as edge pixels. B. System Description A general flowchart of the proposed algorithm is depicted in Fig. 2. First, labels are found by applying the FCM algorithm to the image histogram. Then, the information about the labels is employed to build the networks activation and error functions. The input to a neuron in the input layer is normalized between [01], proportionally to the gray value of the correspondent pixel. The image information is first propagated forward using ( 8 ) to get the output status of the network. The output value of each neuron lies in the interval [0-1]. Then, the output error is calculated and then backpropagated to update the weights. Training continues either until a minimum error or until a maximum number of iterations is reached.
2011 ACEEE
Fig. 2 General Flowchart

The output of the system at this stage constitutes the segmented image. Integrating the thresholded differences between the output at consecutive epochs yields the edge image. The proposed algorithm consists of two main processing blocks (Fig. 3): 1) The fuzzy error function definition block (A), 2) The adaptive thresholding block (B).

Fig. 3 Block Diagram of Proposed System.

21

DOI: 01.IJRTET.05.01.138

Full Paper
Int. J. on Recent Trends in Engineering & Technology, Vol. 05, No. 01, Mar 2011 1) Error Function Definition (Block A): The purpose of this stage is to provide the objective error function to be used by the adaptive thresholding stage (for unsupervised training). First, the input image is fuzzified based on its graylevel histogram and then the error function is obtained by determining the contribution of each gray level to the fuzzy entropy of the partition. In this paper, the FCM algorithm was employed to create a fuzzy partition that properly describes the image, using only the pixel intensity feature. In the proposed system, this was done by iterating the FCM algorithm for a range of hypothesized numbers of clusters and choosing the best option based on a cluster validity measure. Although in general cluster validity measures are not considered very reliable, some of them yielded surprisingly good results for some of the test images. This might be partially due to the special nature of the data, which is not common in clustering problems: the data is 1-D and at least one entry exists at each possible point. Still, in general, the problems involved in this choice are (yet) unsolvable, thus leaving this parameter as the only part of the system requiring the (yet) non-replaceable human intervention. Supposing that someone does rely on one of the cluster validity measures, another decision has to be made: on one hand, the correct C can be chosen as the one providing the external value for the specific validity measure employed. On the other hand, accepting the fact that different values of C might fit the given data, a threshold on the validity measure should be chosen below/above which C is accepted. There is (yet) no theoretical way to define such a threshold. Attempts to do this are done based on application/data specific heuristics, like how many different Cs are acceptable, or how far can the threshold be from the external value of the validity measure. 2) Adaptive Thresholding (Block B): This stage contains the adaptive thresholding system itself, the fuzzy entropy calculation block and the training/tuning algorithm of the adaptive thresholding system. Its inputs are the input image and the error function determined by stage A. And its output is the segmented image. The adaptive thresholding system is based on a self-organizing neural network. This part of the system is based on the image segmentation system proposed by Ghosh et al.. They developed a self-organizing MLP that has an inherent unsupervised learning ability. The network architecture proposed by them is similar to the one used in the present work, except that in their system the outputs of the network were fed back as inputs to the network after the first forward pass. Their system suited only the task of objectbackground separation. Furthermore, they assumed that the membership values of the image pixels in the two classes (object and background) attain their maximum at one of the ends of the intensity value range and descend monotonically toward the other end, attaining the value of 0.5 exactly at the middle of the range. In the proposed system, the neuron activation function and the error measuring function were completely modified in order to deal with multilevel segmentation. 22 The network consists of an input layer, an output layer and at least one hidden layer. Each layer consists of M x N neurons, every neuron corresponding to an image pixel. Each neuron in one layer is only connected to the corresponding neuron in the previous layer and the neurons in its d-th order neighbourhood. A neighbourhood system over a M x N lattice L is defined as (7) where dij called the d-th order neighbourhood of (i ,j ), is such that ( i ,j ) dij; ( k, l ) implies ( i ,j ) dij. There are no connections between neurons in the same layer. The NNs weights cannot be randomly initialized or they will alter the input image. In this work, all weights were initialized to 1, but it is also possible to initialize the weights using some kind of weighting window within the neighbourhood of each pixel. 3) Activation function: In order to allow more than two stable states of the neuron output (i.e., more than two allowable intensity levels), a special activation function was developed. It consists of a sigmoid-like function with multiple levels. The multisigmoid is obtained by the superposition of k= C-1 shifted sigmoid functions (C being the number of classes). The rationale behind using such a function was to enable stability around the multiple targeted pixel values. The multisigmoid function is defined as (8) where

The thresholds and the target values are obtained from the error function, as the gray levels with the maximal and with minimal levels of fuzziness respectively. Because the range of the neuron input levels depends on the number of neurons in the previous layer to which it is connected (the size of the neighbourhood), the threshold values are adapted to reflect this dependency (by multiplying them by d2, the number of input links). 4) Training: The back-propagation algorithm is employed for training. The weights are updated as follows:

(9)

2011 ACEEE

DOI: 01.IJRTET.05.01.138

Full Paper
Int. J. on Recent Trends in Engineering & Technology, Vol. 05, No. 01, Mar 2011 where Ii total input to the i-th neuron; wji weight of link from neuron i in one layer to neuron j in the next layer; oi output of the i-th neuron in the previous layer; E error in the networks output (relative to the desired target image); learning rate. For a multisigmoid as previous defined ( 10 ) and the equations for become V. EXPERIMENTAL RESULTS Some of the intermediate results of proposed system are given below. The intermediate results are obtained by applying FCM clustering algorithm on different HSV images. Further ANFIS method is to be applied on those results so that we can get segmented image output.

( 11 )

5) Fuzziness measure/Error function: The error function defined at the first stage of the system is evaluated for the image output at each training epoch. As discussed before, the aim of the network is to reduce the degree of fuzziness of the input image. 6) Defuzzification: As it should be understood from the descriptions of the neural network, it is working directly on the gray level values, and not on their fuzzy membership values, as would be the case in a usual fuzzy image processing algorithm. In other words, the network does not change the fuzzy membership values of the pixels in order to reduce the error. Instead, it changes the initial pixel values to values that will decrease the amount of fuzziness, according to the initial fuzzy partition. Thus, the output of the neural network is initially obtained in terms of gray levels, which are then fuzzified in order to determine the error. In the ideal case when the network converges with no error at all (E = 0), the outputs have only values whos membership values are 1 or 0, defuzzification is not necessary. When the network does not converge completely, the fuzzification of the output image does not result in merely crisp membership values. The information about the membership values of the pixels might be useful for further processing, depending on the application at hand. If crisp labelling is required, a defuzzification stage must be added. For display purposes, the simplest defuzzification method is thresholding the fuzzy partition, so that each pixel is uniquely assigned to the class in which it has the highest membership value. After being assigned to a single class, the pixels can be labelled either with the value of the prototype of their class, or with the value of the prototype weighted by the membership values of the pixels in their class (thus preserving some of the fuzzy information).

VI.CONCLUSION In this method, ANFIS technique for color image segmentation is presented. The image segmentation can be obtained automatically using multilevel adaptive thresholding and FCM algorithm. The main part of proposed system is the neural network that performs the adaptive thresholding. The FCM algorithm is going to be used to determine threshold and error functions. The obtained results will be compare with those of other know method.

2011 ACEEE

23

DOI: 01.IJRTET.05.01.138

Full Paper
Int. J. on Recent Trends in Engineering & Technology, Vol. 05, No. 01, Mar 2011 REFERENCES
[1] Kanchan Deshmukh And G. N. Shinde An Adaptive NeuroFuzzy System for Color Image Segmentation, J. Indian Inst. Sci., sept.-oct. 2006, 86, 493-506, Indian Institute of science. [2] Jang, Neuro-Fuzzy Modelling: Architectures, Analysis and Applications, PhD Thesis, University of California, Berkley, July 1992. [3] J. S. R. Jang, ANFIS: Adaptive-Network-Based Fuzzy Inference System, IEEE Trans. on Systems, Man and Cybernetics,1993. [4] Victor Boskovitz and Hugo Guterman, An Adaptive NeuroFuzzy for Automatic Image segmenatation and Edge Detection, IEEE Transactions on Fuzzy Systems, VOL. 10, NO. 2, APRIL 2002 [5] Jyh-Shing Roger Jang, chuen-Tsai Sun, Neuro-Fuzzy Modeling and Control. [6] V. Boskovitz and H. Guterman, Neuro-fuzzy system for adaptive multilevel image segmentation, Proc. 19th IEEE Conf. EE Engineers Israel, pp. 208211, 1996. [7] Jaume Verges Llah, State-of-the-Art Survey on color Segmentation Methods. [8] R.C Gonzales and R.E. Woods, Digital image processing, pp.579,2*ed., Prentice Hall, New Jersey, USA, 2002 [9] Ajith Abraham Neuro Fussy Systems: State-of-the-art Modelling Techniques. [10] http://rsb.info.nih.gov/ij/images/ [11] Branimir B. JovanoviC, Irini S. Reljin, , and Branimir D. Reljin, Modified ANFIS Architecture Improving Efficiency of ANFIS Technique, 7th Seminar on Neural Network Applications in Electrical Engineering, NEUREL-ZOO4 faculty of Electrical Engineering, University of Belgrade, Serbia and Montenegro. September 23.25.2004 [12] L. Shamir, Human Perception-based Color Segmentation Using Fuzzy Logic, Proc. Of International Conference on Image Processing , Computer Vision and Pattern Recognition (IPCV 2006), Ls Vegas, USA, Jun 2006, pp. 496-505 [13] A. Soetedjo, K. Yamada, A New Approach on Red Color Thresholding for Traffic Sign Recognition System, Journal of Japan Society for Fuzzy Theory and Intelligent Informatics, Vol. 19, No. 5, 2007, pp. 457-465. [14] Abdallah A. Alshennawy, and Ayman A. Aly, Edge Detection in Digital Images Using Fuzzy Logic Technique, World Academy of Science, Engineering and Technology 51 2009.

2011 ACEEE

24

DOI: 01.IJRTET.05.01.138