Retina recognition system using adaptive neuro fuzzy inference system

IEEE International Conference on Computer, Communication and Control (IC4-2015).

Retina Recognition System using Adaptive Neuro Fuzzy Inference System Tripti Rani Borah1

Kandarpa Kumar Sarma2

Pran Hari Talukdar3

Gauhati University, Guwahati, India. [email protected]

Gauhati University, Guwahati, India. [email protected]

Gauhati University, Guwahati, India. [email protected]

Abstract -- Biometric based technologies including retina, face, fingerprint, speech, hand geometry, handwriting, iris and typing rhythm are used to deal high security problems because, they have reached a high degree of maturity such as applications on secure authentication. Artificial Neural Networks (ANN) are nonparametric prediction tools based on the analogy of biological nervous systems consisting of interconnected group of artificial neurons. They have information-processing units based on connectionist approach that can be used for a host of pattern classification and recognition applications. Fuzzy Logic (FL) is a powerful problem solving methodology receiving widespread acceptance for a range of applications. It provides a simple way to reach the definite conclusion from ambiguous, imprecise and vague information. Like ANN models, some Fuzzy Inference System (FIS) has the quality of universal approximation. ANN along with FL constitutes Adaptive Neuro Fuzzy Inference System (ANFIS) which can be used to model a system with nonlinear, random and uncertain data. An ANFIS consists of the advantages of ANN and Fuzzy System (FS). In this paper, a biometric recognition system based on retina using ANFIS as classifier is described. A number of retina samples have been collected and used in this retina recognition system as inputs for proper identification. Here, Principal Component Analysis (PCA) is used for the feature extraction of the blood vessels of the retina. The specialty of the work is associated with the fact that if the ANFIS is configured properly in terms of number and types of membership functions (MFs), it can tackle the variations in the retinal images properly. Experimental results show that the system is reliable for such kind of identification system. Keywords - retina; artificial neural network; fuzzy logic; adaptive neuro fuzzy inference system

I. INTRODUCTION Of late, biometrics has been accepted as an emerging field of technology using unique and measurable ‘physical and behavioural’ characteristics that can be processed electronically to established identification, perform verification and automated recognition of a person. These physical attributes include facial appearance, fingerprint, hand geometry, handwriting, iris, retina, veins and voice. Behavioral biometrics is related to the behavior of a person, including but not limited to: typing rhythm, gait etc. Retina identification is an automatic method that provides true identification of the person by acquiring an internal body image which is difficult to counterfeit [1]. Retina identification has found application in high security environments. An ANN is easily configured and trained to handle such variations observed in the texture of the retina.

But in some cases the capabilities of the ANN can be enhanced by combining fuzzy systems. Fuzzy systems are known to be capable of dealing with uncertainty and finer variation. While ANN is able to learn from processed data, retain the learning and use it subsequently. Therefore the combination of ANN and fuzzy systems together called neurofuzzy system (NFS) that combines the advantages of both the tools. Here, we report the application of a NFS configured as an adaptive NFS for inference generation using fuzzy reasoning. The system is based on adaptive neuro-fuzzy inference system (ANFIS) which is trained to perform as a reliable biometric authentication system using retina as an input. Here, we report the design of an ANFIS based biometric authentication system which provides fast and reliable detection of a range of retina inputs and shows better performance compared to an ANN. Various researches on biometric identification based on retina recognition using fuzzy have been going on all over the world. Some of the relevant works can be mentioned below: In [2], Retinal vessel segmentation using Gabor filter and artificial neural network is a work proposed which demonstrates an automated segmentation scheme of retinal vasculature using Gabor filter bank, which is optimized on the basis of entropy. In [3], Supervised segmentation of vasculature in retinal images using neural networks is another work proposed, where the authors reported about a neural network based supervised segmentation algorithm for retinal vessel delineation. Self Organizing neural network based pathology classification in retinal images is another work proposed in [4], where the authors designed an automated system based on Self-Organizing neural network (Kohonen network) for eye disease classification. In [5], Retinal vessel segmentation using the 2-D Morlet wavelet and neural network is a work proposed where the authors reported a new method for automatic segmentation of the vasculature in retinal images. In [6], Emulation of salamander retina with multilayer neural network is another work proposed where the authors used multilayer cellular neural network platform to emulate the functions of salamander retina. In [7], Retina vessel detection using Fuzzy Ant Colony Algorithm is another work proposed, where the authors described a fuzzy clustering method based on Ant Colony Algorithm. In [8], Vascular landmark classification in retinal images using fuzzy RBF is a work proposed where the author suggested the use of radial based neural networks for classification of the landmark points from retina vessels in the retinal vascular images to diagnose the disease in the diabetic retinopathy patients and to

IEEE International Conference on Computer, Communication and Control (IC4-2015). track the periodic differences in retinal vessel images. In [9], Locating the Optic Nerve in a Retinal image using the Fuzzy convergence of the blood vessels is a work proposed where the author used an automated method to locate the optic nerve in images of the ocular fundus. They used a novel algorithm called fuzzy convergence to determine the origination of the blood vessel network. In [10] Xu et. al, obtained vector curve of blood vessel’s skeleton using the green channel gray-scale retina images. In [11], another approach is reported where the authors used a fuzzy circular Hough transform to localize the optical disk in the retina image. This algorithm is computationally more efficient with respect to the algorithm reported in [10]. Here, we noticed that the major literature have been restricted to the preprocessing techniques with no focus on the variations of the blood vessels. These are restricted to the popular NFS learning and decision making arena with no focus on performance variations that may be derived from the application of fuzzy based hybrid systems with varying fuzzification norms and inference rules. The specialty of the work is associated with the fact that if the ANFIS is configured properly in terms of number and types of membership functions (MF) it can tackle the variations in the retinal images. This way the approach provides the insights for developing certain learning aided system which is autonomous and requires these samples for verification and authorization during the training stage only. Experimental results show that the proposed approach is fast, reliable and robust for a range of conditions. Here, we used samples from databases of ISI Kolkata, MESSIDOR, Sri Sankardeva Nethralaya, Assam. Some of the issues involved are acquisition of proper quality retina images, so that these can be used as inputs in the recognition system, preprocessing of the input images for making them suitable for manipulation and interpretation, use of different kinds of techniques for the enhancement of input images, application of efficient feature extraction techniques to develop a prototype model for retina based biometric identification system and application of neuro-fuzzy classifier for improvement of the overall performance of the recognition system. The rest of the paper is organized as follows: Section II provides the brief description of a generic retina recognition system. Section III provides the background principles related to the working of the proposed model. All experimental results and related discussions are provided in Section IV. This paper is concluded by summing up the work in Section V. II. BASIC THEORETICAL ASPECTS RELATED TO THE PROPOSED SYSTEM A. Retina Retina is the vascular pattern of the eye which is not easy to change and replicate. The patterns are different for right and left eye. The retina of a person is unique and remains unchanged over a lifetime [12]. “Fig. 1,” shows the structure of a retina.

Fig.1 Structure of a retina.

B. Retina Recognition System (RRS) RRS captures and analyzes the patterns of blood vessels on the thin nerve on the back of the eyeball that processes light entering through the pupil. Retinal patterns are highly distinctive traits. Every eye has its own totally unique pattern of blood vessels. The eyes of identical twins are also distinct [1]. The strength of the retina recognition is difficulty in destroying its features. C. Artificial Neural Network (ANN) These are non-parametric prediction tools based on the analogy of biological nervous systems consisting of interconnected group of artificial neurons. They also have information processing units based on connectionist approach that can be used for a host of pattern classification and recognition applications. ANNs are trained to perform complex functions in various fields, including pattern recognition, identification, classification, speech, vision and control systems and to solve problems that are difficult for conventional computers or human beings [13]. D. Fuzzy Logic (FL) Zadeh introduced Fuzzy sets in 1965 [14] as a means of representing and manipulating data that was not precise, but represent variations between the limits defined by traditional crisp sets. The variations are statistical in nature which enables fuzzy based systems to track minor variations. FL starts with and builds on a set of user-supplied human language rules. It provides a simple way to arrive at a definite conclusion based upon vague, ambiguous, imprecise, noisy or missing input information. FL is a set of mathematical principles for knowledge representation and reasoning based on degrees of membership [15]. The primary components of a fuzzy based system are fuzzification, fuzzy inference system (FIS) consisting of MFs and fuzzy rules and defuzzification stages as shown in “Fig.2”. In FIS, fuzzifier takes the crisp inputs to a fuzzy system and converts them into fuzzy inputs. Fuzzy rule base consists of fuzzy IF-THEN rules that form the heart of a fuzzy system. FIS makes use of FL principles to combine the fuzzy IF-THEN rules. Defuzzifier extracts a crisp value from a fuzzy set. Fuzzy logic allows decision making with estimated values under incomplete or uncertain information. Fuzzy sets are characterized by membership functions which assigns to each element x in a fuzzy set a number, A(x), in the closed unit interval [0, 1]. The number, A(x) represents the degree of membership of x in A. In the case of Crisp sets, the members of a set are either out of the set, with membership degree of zero or in the set with the value one being the degree of membership. Therefore, Crisp sets are special cases of fuzzy sets. There are mainly two types of FISs that have been widely employed in various applications. These are Mamdani FIS (1975) and Sugeno FIS (TSK,1985). The first Mamdani model includes linguistic models based on collections of IF - THEN rules, whose antecedents and consequents utilize fuzzy values. It uses fuzzy reasoning and

IEEE International Conference on Computer, Communication and Control (IC4-2015). the system behavior can be described in natural terms. The knowledge is represented as Ri : If x1 is Ai1 and x2 is Ai2…….and xn is Aim (1) (2) : then yi is Bi where Ri (i = 1, 2,….l) denotes the ith fuzzy rule, xj (j = 1, 2,….l) is the input, yj (j = 1, 2,….l) is the output of the fuzzy rule Ri and Ai1, Ai2,… Aim, Bi (i = 1, 2,….l) are fuzzy membership functions usually associated with linguistic terms. The second category, based on Sugeno–type system uses the rule structure that has fuzzy antecedent and functional consequent parts. This can be viewed as the expansion of piecewise linear partition represented as Ri : If x1 is Ai1 and x2 is Ai2…….and xn is Aim (3) : then yi =ai0 + ai1x1 + ..... ainxn (4) Fuzzy control has proved to be the most successful application of fuzzy logic. Fuzzy controllers, by the utilization of human knowledge, revolutionized the field of control engineering so as to provide solutions for situations which had troubled designers for a long period of time. The solutions derived were precise and inexpensive. A typical Fuzzy Control algorithm would proceed as follows: x Obtaining information: Collect information of measurements related to all relevant variables. x Fuzzification: Convert the collected measurements data into appropriate fuzzy forms to include the uncertainties in the measurements. x Running the Inference Engine: Use the fuzzified forms to examine the control rules and formulate the set of possible actions. x Defuzzification: Convert the set of likely functions actions into a numerical value. x Repeat the above. Stability, learning capability, membership functions and fuzzy rules are the some problems associated with fuzzy systems as well.

Fig. 2 Block diagram of a fuzzy system.

E. Neuro-Fuzzy System A neuro-fuzzy system is a neural network which is functionally equivalent to a fuzzy inference system. It can be trained to develop IF-THEN fuzzy rules and determine membership functions for input and output variables of the system. It combines the advantages of ANN with those of fuzzy systems. The structure of a NFS is similar to a multilayer neural network. In general, a NFS has input and output layers, and three hidden layers that represent membership functions and fuzzy rules. The hybrid neuro-

fuzzy system combines the parallel computation and learning abilities of neural networks with the human-like knowledge representation and explanation abilities of fuzzy systems. As a result, NN becomes more transparent while fuzzy system becomes capable of learning. F. Adaptive Neuro-Fuzzy Inference System Fuzzy systems provide as expert level decision making system by following the variations in the texture of the input data. ANN and Fuzzy System (FS) are the important components of soft computing. They have received large acceptance in the diverse fields of uncertainty. ANNs are comfortable with the situations that have sufficient model free measurement data [16] whereas FSs work where sufficient process data are available and expert knowledge is necessary [17]. Thus ANN and FS having the numeric quantitative capability and symbolic-qualitative capacity respectively can be applied to solve problems where higher precision of estimation is difficult. ANN and fuzzy formed hybrid systems that have adaptability, parallelism, non-linear processing, robustness and learning in data rich environment for modeling uncertainty [15]. ANFIS has found numerous applications in a variety of fields. The fuzzy model can either be Mamdani fuzzy model or TSK fuzzy model. However, TSK model is smoother and computationally efficient representation than Mamdani’s model. Let, the considered FIS is given two inputs ‘x’ and ‘y’ and one output ‘f’. For a first-order Sugeno fuzzy model, a common rule set with two fuzzy if-then rules is as follows: Rule 1: If (x is A1) and (y is B1) then (f1 =p1 x +q1y+r1) (5) Rule 2: If (x is A2) and (y is B2) then (f2 =p2 x +q2y+r2) (6) where Ai and Bi are the fuzzy sets, fi are the outputs within the fuzzy region specified by the fuzzy rule, pi, qi and ri are the design parameters that are determined during the training process. “Fig. 3,” shows the Type-3 fuzzy reasoning mechanism for a first order Sugeno fuzzy model. The corresponding equivalent ANFIS architecture to implement these two rules is shown in “Fig. 4”. in which a circle indicates a fixed node, whereas a square indicates an adaptive node [18]. The general architecture of ANFIS, which is primarily a fuzzy system having five layers as described below: The first layer performs fuzzification. The second layer generates the response of the network due to fuzzy inputs. The Third layer performs normalization. For the fuzzy responses adaptive update is done by the fourth layer. The last layer performs the task of summation of all the outputs.

IEEE International Conference on Computer, Communication and Control (IC4-2015).

Fig.3 TSK type -3 Fuzzy reasoning.

Fig.4 Equivalent ANFIS.

G. Fuzzification Fuzzification is one of the primary blocks of our proposed approach. Here, we have used Generalized Bell MFs and Gaussian MFs. A generalized bell MF is specified by three parameters {a, b, c} and is expressed as: ଵ Bell(x ; a, b, c) = (7) ೣష೎ మ್ ଵାቚ

ೌ

ቚ



where ‘c’ determines the center of the curve , ‘a’ is the half width and ‘b’ (together with ‘a’) controls the slopes at the crossover points (where MF value is 0.5) and the parameter ‘b’ is usually positive. A desired bell MF can be obtained by a proper selection of the parameter set {a, b, c}. Specifically, we can adjust ‘c’ and ‘a’ to vary the center and width of the MF and then use to control the slopes at the crossover points. We have also considered the Gaussian MF . Gaussian MF is specified by two parameters {c, s} and is expressed as: ଵ

୶ିୡ ୫

Gaussian (x; c, s, m)=exp ቂെ ቚ ቚ ቃ (8) ଶ ୱ ‘c’ represents the center of the MF’s , width of the MF’s and m is the fuzzification factor (e.g.; m=2). Because of their smoothness and concise notation, gauss MF’s and bell MF’s are becoming increasingly popular methods for specifying fuzzy sets. Gaussian functions are well known in the fields of probability and statistics and they possess useful properties such as invariance under multiplication and Fourier transform. The bell MF has one more parameter than the Gaussian MF, so it can approach a non fuzzy set if b→∞ [17]. A detailed set of experiments are carried out in this work. Respective performances observed derived by considering these two MFs and their impact in ascertaining the system performance. III. DESIGN AND IMPLEMENTATION OF THE PROPOSED SYSTEM A retina based biometric identification system designed using ANFIS is shown in “Fig. 5”.

Fig.5 Flow logic of the proposed system.

It involves image acquisition, image preprocessing, feature extraction and classification. In this work, the focus is to study the performance of the retina recognition system that provides reliability, accuracy and reduced overall match speed. The decision obtained from the system is used to generate the response. During image acquisition the operations are performed separately. Retina images captured by the Fundus camera are pre-processed for subsequent manipulation. Retina image preprocessing includes color image conversion, resizing the original images to required size, noise removing and filtering. In this work, median filter is used to remove the noise from the input images. Median filter is used because of its demonstrating ability to reduce random impulsive noise without blurring edges. The next stage is the feature extraction. Retina features include patterns of blood vessels [12]. In this proposed model, we have used Principal Component Analysis (PCA) to extract the retina features. PCA involves a mathematical procedure that transforms a number of correlated variables into a number of uncorrelated variables called principal components. The first principal component accounts for as much of the variability in the data as possible and each succeeding component accounts for as much of the remaining variability as possible. PCA is a common technique for finding patterns and size reduction in data of high dimension. The feature vector represents a unique set of data providing relevant information regarding shape, size, morphology etc of the inputs. For this retina recognition system, the feature length is 101 due to the principal components number 101 giving best optimal performance during training. Here, ANFIS is used as classifier for recognition. It applies a combination of the least-squares (LS) method and the back propagation gradient descent (BPGD) method for training FIS membership function parameters to emulate a given training dataset. The ANFIS considered have five layers and its key specifications are provided in Table I. Retina feature of length 101 is used for the training of the ANFIS for 10 to 200 epochs. The results obtained are average values of at least ten trials for the epoch considered. We have used the Gaussian-curve MF for doing fuzzification. This is because the Gaussian-curve MF provides the most reliable real world to fuzzy conversion as has been known from experiments. We have considered ten such MFs which give optimum results. Also, we consider the root mean squared error (RMSE) to be the cost functional for training the ANFIS. IV.

EXPERIMENTAL DETAILS AND RESULTS

The performance of RRS is analyzed in terms of computational speed and reliability. The overall computational time taken by the system is reduced to a greater level. A total of 100 identical retina images have been provided to the system for training, validation and testing and from them, a few set noise corrupted samples are also formed for testing. After extensive training, the system is subjected to certain

IEEE International Conference on Computer, Communication and Control (IC4-2015). variations with signal to noise ratio (SNR) range between 0 to 3 dB to achieve robustness and proper recognition. The obtained results are compared with [19] and found that the ANFIS generate comparable RMSE values but the computational speed is significantly better. Further, the RMSE convergence is completed in less number of epochs. Table II shows the time and minimum RMSE for 200 epochs training with the ANFIS. Table III shows the average success rates achieved among few numbers of training epochs with ANFIS and ANN. Thus, the proposed approach is better for verification and authorization for the design of RRS. A sample retina image, its gray image, resized, noise mixed image and its filtered images are shown in “Fig.s 6 to 10,” respectively. The average RMSE convergence plot shown by the ANFIS during training is shown in “Fig. 11”. The ANFIS based approach takes, on an average 50% less epochs than the ANN aided method and generates between 2 to 3% better success rates. The training time required is between 1.9 to 2.1 seconds for each sample. The results are derived by performing ten trials for the sample set and the average results are quoted. The Root Mean Squared Error (RMSE) is calculated using the expression: ଵ

ଶ RMSE= ට σே ௧ୀଵሺ‫ܣ‬௧ െ  ‫ܨ‬௧ ሻ

9 3 6 9

Gaussmf

1.99x10-3 2.82x10-3 2.08x10-3 1.97x10-3

2.1 1.9 2.0 2.1

TABLE III. AVERAGE SUCCESS RATES ACHIEVED AMONG FEW NUMBERS OF TRAINING EPOCHS WITH ANFIS AND ANN FOR RETINA Observation no.

Minimum Epochs by ANN

Epochs by ANFIS

%Success rate with ANN

%Success rate with ANFIS

1

15

10

86

88

2 3 4 5 6

30 45 60 74 86

20 30 40 50 60

90 91 92 92.5 93

92 93 94 95 96

(9)

ே

where At and Ft are actual and fitted values respectively and N is the number of training or testing sample. The RMSE is used as the primary criterion to ascertain the extent of learning acquired by the ANFIS.

Fig.6 Original sample of a retina image.

TABLE I. ANFIS SPECIFICATIONS

Input Data size

For retina-PCA features of length 101

SNR

0-3dB

ANFIS type

ANFIS with five layers

ANFIS training method Average training epochs Total no. of membership functions Type of membership function RMSE with 9 GaussMFs at 200 epochs

Least-squares with the back propagation gradient descent method

Fig.7 Gray image of the retina image.

10-200 10 Gaussmf, Gbellmf 1.97x10-3

Fig. 8 Resized image of the retina. TABLE II. CALCULATION OF TIME AND AVERAGE RMSE FOR 200 EPOCHS FOR RETINA

Input

Type of member ship function

No. of Membership Functions (MFs)

Time required to reach 200 epochs by (in sec.)

Minimum RMSE attained by 200 epochs

Retina

Gbellmf

3 6

1.9 2.0

2.82x10-3 2.05 x10-3

IEEE International Conference on Computer, Communication and Control (IC4-2015). Fig.9 Noise mixed image.

Fig.10 Filtered image of the retina.

Fig. 11 Average RMSE convergence of the ANFIS.

V.

CONCLUSION

In this work, we described a RRS where the ANFIS forms a critical decision support system. The experimental results show that the proposed approach is more efficient compared to ANN based approach reported earlier [19]. Computational speed, efficiency, robustness and high precision are the strength of the approach which shall make it suitable for such kind of applications. The future scope of this work is to improve the ANFIS architecture to achieve high classification accuracy. The overall performance of the system can be enhanced further by considering more number of samples and by using better preprocessing techniques with ANFIS based blocks. REFERENCES [1] R. B. Hill, “Apparatus and method for identifying individuals through their retinal vasculature patterns”, US Patent No.4109237, 1978. [2] M. Nandy and M. Banerjee, “Retinal vessel segmentation using Gabor filter and artificial neural network,” Proceedings of Third International Conference on Emerging Application of Information Technology (EAIT) ,Kolkata, India, pp. 157 – 160, 2012. [3] C. Ding, Y. Xia and Y. Li, “ Supervised segmentation of vasculature in retinal images using neural networks,” Proceeding of IEEE International Conference on Orange Technologies (ICOT), Xian, China, pp. 49 – 52, Sept 2014. [4] J. Anitha, D. J. Hemanth, C. Kezi Selva Vijila and A. Ahsina, “Self Organizing neural network based pathology classification in retinal images” Proceedings of World Congress on Nature & Biologically Inspired Computing (NaBIC), Patiala, India, pp. 1457 – 1462, Dec 2009. [5] R. Ghaderi, H. Hassanpour and M. Shahiri, “Retinal vessel segmentation using the 2-D Morlet wavelet and neural network,” Proceeding of International Conference on Intelligent and Advanced Systems (ICIAS), pp. 1251 – 1255, Kuala Lampur, Malayasia, Nov 2007. [6] Y. Sung-Nien, L. Chien-Nan and C. Chun-Chieh,. “Emulation of salamander retina with multilayer neural network,” Proceeding of IEEE International Symposium on Circuits and System (ISCAP), Taipei,Taiwan, pp. 2902-2905, May 2009.

[7] S. Hooshyar and R. Khayati, “Retina Vessel Detection Using Fuzzy Ant Colony Algorithm,” Proceeding of Canadian Conference on Computer and Robot Vision (CRV), Ottawa, Canada, pp. 239 – 244, May 2010. [8] C. Candemir, C. Cetinkaya, O. Kilincceker and M. Cinsdikici, “Vascular landmark classification in retinal images using fuzzy RBF,” Proceeding of Signal Processing and Communications Applications Conference (SIU), Haspolat, pp. 1 – 4, April 2013. [9] A. Hoover and M. Goldbaum, “Locating the Optic Nerve in a Retinal Image Using the Fuzzy Convergence of the Blood Vessels,” IEEE Transactions on Medical Imaging, vol 22, iss 8, pp. 951- 958, Aug 2003. [10] Z. W. Xu, X. X. Guo, X. Y. Hu and X. Cheng, “The Blood Vessel Recognition of Ocular Fundus” Proceedings of the 4th International Conference on Machine Learning and Cybernetics (ICMLC05), Guangzhou, China, pp. 4493 – 4498, August 2005. [11] M. Ortega, C. Marina, M.G. Penedo, M. Blanco and F. Gonzalez, “Biometric Authentication using Digital Retinal Images,” Proceedings of the 5th WSEAS International Conference on Applied Computer Science (ACOS 06), Hangzhou, China, pp. 422 – 427, April 2006. [12] Hill, R. B., “Retina Identification, Biometrics”, Springer, 2003. [13] S. Haykin, Neural Networks- A Comprehensive Foundation, Pearson Education, 2nd ed., New Delhi, 2003. [14] Zadeh, L.A., Fuzzy sets, Information and Control,vol.8, . pp. 338-353, 1965. [15] T. J.Ross, Fuzzy logic with Engineering Applications, 2nd edn., Wiley India, New Delhi, 2008. [16] S. Mitra, and Y. Hayashi, “Neuro-Fuzzy Rule Generation: Survey in Soft Computing Framework”, IEEE Transactions on Neural Networks, vol. 2, no.3, pp. 748-768. [17] J.S.R. Jang and C.T.Sun, “ Neuro Fuzzy Modeling and Control”, in the proceedings of the IEEE, vol. 83, pp. 378-406, 1995. [18] R.J. Jang, ANFIS: Adaptive-Network Based Fuzzy Inference System, IEEE Transaction on Systems, Man and Cybernetics vol.23, iss3. 1993. [19] Borah, T.R., K.K.Sarma and P.H.Talukdar, “Retina and Fingerprint based Biometric Identification System”, in International Journal of Computer Applications (IJCA), pp. 74-77, Feb 2013.