A Novel Adaptive Inertia Particle Swarm Optimization ... - IEEE Xplore

3 downloads 0 Views 161KB Size Report
Institut TELECOM; TELECOM SudParis,. France. Abstract—In this paper, we present an efficient feature se- lection scheme for biometric authentication (for both ...
A Novel Adaptive Inertia Particle Swarm Optimization (AIPSO) Algorithm for Improving Multimodal Biometric Recognition R.Raghavendra

Bernadette Dorizzi

Istituto Italiano di Tecnologia, Italy.

Institut TELECOM; TELECOM SudParis, France.

Abstract—In this paper, we present an efficient feature selection scheme for biometric authentication (for both unimodal and multimodal systems) that allows selecting the dominant features and increase the performance of the overall system. More precisely, we propose an Adaptive Inertia Particle Swarm Optimization (AIPSO) algorithm such that the particle inertia weights are iteratively updated according to the particle fitness value. We then use AIPSO for selecting Log Gabor features for the face and palmprint modalities independently and on the fused Log Gabor space of these two modalities considered for fusion. Final classification (in both schemes) is performed on the projection space of the selected features using Kernel Direct Discriminant Analysis (KDDA). Extensive experiments are carried out on 250 users selected from FRGC face database, PolyU palmprint database and a virtual person multimodal biometric database built from the considered face and palmprint databases. We compare the proposed selection method with well known feature selection schemes such as Sequential Floating Forward Selection (SFFS), Genetic Algorithm (GA), Adaptive Boosting (AdaBoost) and Normal PSO in terms of both number of features selected and performance. Experimental result results show better performance of our AIPSO compared to all other techniques with an improvement of around 5% in performance and a reduction of around 62% of features compared to the initial system (with full features).

I. I NTRODUCTION Multi-biometrics is the subject of studies nowadays as it is considered as a way to deal with some of the drawbacks of monomodal systems [1] like: noise, lack of universality, intra class variation, vulnerability to forgeries and so on. In particular, fusing different hand-based biometrics like palmprint, hand vein, palm vein [2] is gaining more popularity [2]. Among the available multimodal biometric systems, the fusion of face and palmprint are widely addressed in the literature [1]. The main advantage of these modalities apart from their complementarity is the possibility of combining the information at various level like sensor [3], feature[4], score[5] and decision level[5]. Thr majority of the state of the art face and palmprint multimodal biometric system are based on score fusion, but as reported in [1], a biometric system that integrates information at an earlier stage of processing is expected to provide better performance than systems that integrate information at a later stage [4] [6]. The main drawback of feature level fusion is the high complexity resulting from the high dimension of the input vector coding the data to be fused. A first way to

tackle this problem is to use some feature reduction methods such as linear or non linear projections on subspaces [7], [8] [4]. The second way is to use feature selection schemes like SFFS [9], Genetic Algorithm [10] and PSO [4]. An interesting analysis on the impact of feature level fusion with regards to score level fusion of face and palmprint biometric is presented in [4]. Indeed these two modalities can be considered as independent ones and the same 2D-Gabor transformation can be used to extract from the original images, efficient features, corresponding to well-chosen Gabor coefficients. This is at the price of a high dimension of the feature vectors coding the images. Here, the evolutional technique of Binary Particle Swarm Optimization was used to further reduce the dimension of the fused feature space without affecting the performance of the overall system. From this work [4], it also appears that, while the classical PSO approach was quite efficient in reducing the number of feature vectors, it could not bring any improvement in performance over the use of the complete set of features. This result was also confirmed in [6], where a comparison between Adaboost and PSO for feature selection is performed. PSO was shown to perform better compared to Adaboost as a feature selector but none of the two approaches were able to provide any improvement of performance over the use of the full set of features. In this paper, we propose a new scheme for the feature level fusion of face and palmprint using a novel algorithm (called AIPSO) for feature selection and we will show on several experimental protocols that this new scheme is able at the same time to reduce the feature space and to improve the performance of the initial (full-features-based system). This AIPSO algorithm is an improvement of the conventional Binary PSO, which uses identical inertia weights for all the particles. In AIPSO, each particle assigns iteratively a value for its inertia weights in relation to its fitness function. We explore this scheme to perform; (1) feature selection in a Log Gabor features space of face and palmprint separately (unimodal case)(2) feature selection in a Log Gabor fused feature space of face and palmprint (multimodal case). We conduct extensive experiments on the proposed scheme using the same databases and protocols as in [4]. We also compare the proposed method with state-of-the-art feature selection schemes such as Sequential Floating Forward Selection (SFFS)

978-1-4577-0490-1/11/$26.00 ©2011 IEEE

[9], Genetic Algorithm (GA) [10], Adaptive Boosting (AdaBoost) and Normal PSO [6] in terms of both number of features selected and performance. The rest of the paper is organized as follows: Section II presents the proposed AIPSO scheme in details in relation with the standard PSO scheme. Section III discusses the application of AIPSO for unimodal and multimodal biometric systems. Section IV presents the results of the proposed method as well as a comparative analysis with different state-of-the-art schemes. Section V draws the conclusion. II. A DAPTIVE I NERTIA PARTICLE S WARM O PTIMIZATION (AIPSO) In this section, we first describe the general PSO approach then we describe the novel proposed AIPSO algorithm. A. Principle of PSO PSO is a stochastic, population based optimization technique aimed at finding a solution to an optimization problem in a search space [11]. The main objective of PSO is to optimize a given function called fitness function. PSO is initialized with a population of particles distributed randomly over the search space and evaluated using the fitness function. Each particle is treated as a point in the N-Dimension space. The ith particle is represented as Xi = {x1 , x2 , . . . , xN }. At every iteration, each particle is updated by two best values called pbest and gbest. pbest is the best position associated with the best fitness value of particle ’i’ obtained so far and is represented as pbesti = {pbesti1 , pbesti2 , . . . , pbestiN } with fitness function f (pbesti ). gbest is the best position among all the particles in the swarm. The rate of the position change (velocity) for particle ’i’ is represented as Vi = {vi1 , vi2 , . . . , viN }. The particles velocities are updated according to the following equations [11]: Vidnew = W ∗ Vidold + C1 ∗ rand1 () ∗ (Pbestid − xold id ) +C2 ∗ rand2 () ∗ (gbestd − xold id )

(1)

xid = xid + Vidnew

(2)

Where, ’W’ is the inertia weight, d = 1, 2, . . . , N is the dimension of the particle space and ’i’ is the index of the particle. A suitable selection of the inertia weight provides a good balance between global and local explorations, and results in fewer iterations on average for finding near optimal results. C1 and C2 are the acceleration constants used to pull each particle towards pbest and gbest. Low values of C1 and C2 allow the particle to roam far from target regions, while high values result in abrupt movements towards or past the target regions. rand1 () and rand2 () are random numbers between 0 and 1. The first part of Equation 1 ( i.e. (w ∗ Vidold ) ) provides the ”flying particles” with a memory capability allowing the exploration of new search space areas. The second part (i.e. (C1 ∗ rand1 () ∗ (pbestid − xid ))) is the ”cognition” part, which represents the private thinking of the particle itself. The third part (i.e. (C2 ∗ rand2 () ∗ (gbestd − xid )) ) is the

”social” part, which represents the collaboration among the particles. Equation 1 is used to calculate the new velocity of the particle according to its previous velocity and the distances of its current position from its own previous best experience (position) and the group’s best experience. Then the particle flies towards a new position according to Equation 2. Performance of each particle is measured according to a pre-defined fitness function. B. Principle of AIPSO In this Section, we describe our new method for iteratively estimating the value of the inertia weights for each particle. In the literature, very few works [12] [13] report assigning the value of inertia weights in PSO algorithm. In [12], a linearly decreasing scheme for assigning the value of inertia weights is proposed. This value is first set to a maximum value and it is linearly decreased through the course of PSO runs (number of iteration). This protocol is not really adapted to biometric applications where the fitness function that we want to optimize does not increase linearly during the different iterations. In [13], a dynamic inertia weighting scheme using a Fuzzy controller is proposed. Note that this approach is complex since it requires a prior knowledge about the characteristics of the data which is not available in real applications. In this work, we propose a novel way of assigning iteratively the values of the inertia weights of each particle in relation to its fitness function. It results in an adaptation of Equation 1 where the weights are updated for each particle using the relative value of its fitness function in comparison to the values of the best fitness function in the swarm. This way, we obtain directly a suitable value for the inertia weights. Moreover we add, as in Equation 1, a random number between 0 and 1 so that the weights do not get stuck at local minimum. Finally, we have also introduced a minimum threshold for the weights (tolerance value) to avoid their premature convergence. By considering all these factors, the inertia weights are calculated as follows:   fi − fg (3) Wi = 1 − r ∗ fg Where, Wi is the calculated inertia weight of particle i, fi is the fitness value of particle i, fg is the global fitness value (best fitness value in the swarm), r is a random number between 0 and 1. Once Wi is calculated using Equation 3, its value is compared with Wtol , where Wtol is the minimum tolerance value which is set experimentally. If the value of Wi is bigger than Wtol then the weight is equal to Wi or else the weight will be equal to Wtol . C. Adaptive Inertia Particle Swarm optimization (AIPSO) Algorithm The proposed method can be summarized in the following steps, for each particle P: 1) Initialize randomly the particle position.

2) Calculate the fitness value for particle ’i’, fi , and also the global fitness value of the swarm fg . 3) Calculate the value of Wi using Equation 3. 4) Replace the value of W in Equation 1 with Wi . 5) Update the particles values according to Equation 2. 6) Repeat steps 2 to 6 till stopping criteria is satisfied. III. T HE PROPOSED SCHEME FOR BIOMETRIC AUTHENTICATION USING AIPSO Figure 1 shows the block diagram of the proposed scheme for biometric authentication, using AIPSO. This framework is used on both unimodal and multimodal biometric systems using face and palmprint as described in [4]. In case of unimodal biometric, the image is represented by Log Gabor coefficients resulting in a vector of size 3200×1. In the case of multimodal biometric feature level fusion system, the feature level fusion is obtained by concatenating the Log Gabor features of face and palmprint and result in a fused feature space of dimension 6400 × 1. Then, a statistical normalization [8] of the fused features is carried out in order to compensate for the imaging conditions for face and palmprint. The proposed AIPSO scheme is employed individually on the features of face and palmprint (in case of unimodal biometric system) or on the fused features space (in case of multimodal biometric system) to select the best subset of features. For performing verification and closed identification, we use the same framework as in [4] [6] except that we replace the standard binary PSO by AIPSO. 1) Fitness Function: In this work, we will evaluate two different performance measures corresponding to two biometric modes, namely verification and closed set identification [4] . For this reason, we need to define two different types of fitness function for AIPSO implementation. 1) Fitness function corresponding to Verification [4]: In Biometric verification, it is difficult to identify a single function that would characterize the matching performance across a range of False Acceptance Rate (FAR) and False Reject Rate (FRR) values i.e. across all matching thresholds [1]. Thus in our experiments, we first compute the distance between reference and test samples to get the match scores and then we compute FAR and Genuine Acceptance Rate (GAR) by setting different values of the thresholds. In order to optimize the performance gain across a wide range of thresholds, we define the objective function to be the average of 12 GAR values corresponding to 12 different FAR values (85%, 70%, 50%, 30%, 10%, 0.8%, 0.6%, 0.4%, 0.2%, 0.09%, 0.05%, 0.01%). Thus, the main objective of the verification fitness function is to maximize this averaged GAR value. 2) Fitness function corresponding to Closed Identification [4]: The main objective of the closed identification fitness function is to maximize the recognition rate. Given the test sample, we compute its distance against all the samples in the reference dataset to obtain the match scores. Then, we select the sample from the

reference dataset with the lowest distance value and we check whether it belongs to the same class (same person) as the testing sample. In that case it is a success or else a failure. We repeat this for all testing samples and count the number of successes and failures. Finally, we define the fitness function as being the recognition rate calculated as follows: RecognitionRate =

Nsuccess NDB

(4)

where NDB is the total number of samples in the whole database (i.e. reference and testing) and Nsuccess is the number of successes in the whole database. IV. R ESULTS AND D ISCUSSION In this section, we present the results of the proposed biometric authentication schemes using AIPSO. Here, we first provide a brief description of the database and experimental protocol used in this work, which are the same as in [4]. We then discuss the efficacy of the proposed method against stateof-the-art schemes mentioned in [9], [10], [6], and [4]. We employed a multimodal biometric database of face and palmprint as mentioned in [4]. This database consists of 250 users such that each user has 12 samples each corresponding to face and palmprint. In order to have a consistency with our previous work we employ the same protocols as suggested in [4] for conducting the experiments. According to this protocol, the whole database of 250 users is divided into two sets as Set-I (200 users) and Set-II (50 users). Set-II is used as a development database to tune all PSO parameters. Set-I is used as a validation data set to measure the performance of the proposed scheme. Set-I is divided into two equal partitions, non overlapping providing 6 reference data and 6 testing data for each of the 200 persons. In order to reduce the bias in the performance evaluation of the proposed schemes, the reference and testing partition was repeated ’m’ times (where m = 10) using Holdout crossvalidation and there is no overlapping between these two subsets. Thus, in each of the 10 trials we have 1200 (= 200×6) reference samples and 1200 (= 200×6) testing samples resulting in 1200 genuine matching scores and 238800 (= 200 × 199 × 6) impostor matching scores, as for each user, all other users are considered as an impostor. This scheme is valid for verification performance evaluation. In closed identification, we calculate the recognition rate using the 1200 reference samples and the 1200 testing samples that will give 1200×1200 matching scores. Note that as the persons are exactly the same in reference and test sets, this is why we speak of closed identification. Finally, results correspond to the mean of all runs and a 90% parametric confidence interval [6] which gives a better estimation of standard deviation than that which we can obtain by cross-validation. Figure 2 shows the performance of the proposed method along with other state-of-the-art schemes for the face modality. Table I indicates performance and also the number of features selected by the proposed method compared with state-ofthe-art schemes of features selection with verification fitness

Fig. 1.

Block diagram of the proposed scheme for biometric authentication, using AIPSO

TABLE I C OMPARATIVE PERFORMANCE OF THE DIFFERENT FEATURE SELECTION SCHEMES ON FACE V ERIFICATION (M EAN GAR AT FAR = 0.01%) Methods

Dori

DF S

DKDDA

Full features FS-SFFS [9] FS-AdaBoost [6] FS-GA [10] FS-PSO [4] FS-AIPSO

3200 3200 3200 3200 3200 3200

3200 2856 2676 2215 1715 1417

104 98 89 72 56 47

GAR at 0.01% of FAR(%) with 90% confidence interval 67.58 [65.37; 69.79] 68.72 [66.70; 70.74] 70.33 [68.19; 72.47] 70.25 [68.08; 72.42] 70.00 [69.09; 70.91] 75.98 [75.34; 76.62]

TABLE II C OMPARATIVE PERFORMANCE OF THE DIFFERENT FEATURE SELECTION SCHEMES ON FACE R ECOGNITION (M EAN R ECOGNITION R ATE )(C LOSED I DENTIFICATION ) Methods

Dori

DF S

DKDDA

Full features FS-SFFS [9] FS-AdaBoost [6] FS-GA [10] FS-PSO [4] FS-AIPSO

3200 3200 3200 3200 3200 3200

3200 2701 2489 2175 1647 1217

104 96 85 78 51 44

100

95

90

GAR %

85

80

75

70

Full Features PSO NAPSO SFFS AdaBoost GA

65

60 −2 10

−1

10

Fig. 2.

0

10 FAR %

1

10

ROC curves of face biometric

2

10

Recognition Rate with 90% confidence interval 70.12 [68.957; 72.28] 70.89 [68.72; 73.06] 73.01 [70.94; 75.08] 71.85 [69.75; 73.95] 73.45 [71.47; 75.43] 79.19 [77.42; 80.92]

function. In Table I, DOri indicates the initial features dimension, DF S indicates the dimension after applying the feature selection scheme and DKDDA indicates the final dimension of the KDDA projection space and also Full features indicate the performance of the system obtained using all the Log Gabor features. Table II indicates the performance of the proposed AIPSO with closed identification (Recognition Rate) fitness function. From Figure 2, Table I and Table II, we can observe that, the use of AIPSO for selecting the features allows improving the verification performance compared to all the other schemes with a verification rate of 75.98% at FAR=0.01% and a closed identification rate of 79.19%. Further, it is also observed that the proposed AIPSO based scheme will use around 38% of the full features. In this experiment, we have

also noted that, using state-of-the-art selection schemes like SFFS, AdaBoost, Genetic Algorithm and PSO do not provide improvement in performance even if they are able of reducing the Log Gabor feature space up to 30%. This indicates the efficacy of the proposed feature selection scheme using AIPSO which not only allow a high reduction of the feature space but also a significant improvement of the authentication scores.

identification. From Table V and VI, we can observe that the proposed method uses around 34% of the global features as compared to that of full features and achieve nearly 5% gain in the performance with 90% confidence interval. These figures clearly indicate the efficacy of the proposed scheme using AIPSO both in terms of performance gain and number of features selected. V. C ONCLUSION

100

95

90

GAR %

85

80

75

70

Full Features SFFS GA AdaBoost PSO NAPSO

65

60 −2 10

−1

10

Fig. 3.

0

10 FAR %

1

2

10

10

ROC curves of palmprint biometric

100

95

GAR%

90

85

Full Features SFFS GA AdaBoost PSO NAPSO

80

75 −2 10

−1

10

Fig. 4.

0

10 FAR%

1

10

2

10

ROC curves of feature level fusion

We observe the same behavior with the palmprint modality. Figure 3 shows the performance of the proposed AIPSO in terms of ROC curve for the palmprint modality with verification fitness function. Table III and Table IV indicate the performance of different feature selection schemes including AIPSO and also the number of features selected. From Figure 3, Table III and Table IV we can observe that, the AIPSO based scheme shows the best result in term of performance, with a GAR of 82.25% at FAR =0.01% in verification and a recognition rate of 87.47% in closed identification. Further, we can also observe that the AIPSO scheme uses the least amount of features of all the other schemes (less that 35% of full features). Figure 4 shows the performance of the proposed scheme on feature level fusion of face and palmprint in verification. Table V and Table VI indicates the number of features selected using the different schemes including AIPSO. In this experiment also, we can observe that the proposed method shows the best performance with GAR of 98.26% at FAR=0.01% in verification and a recognition rate of 99.92% in closed

In this paper, we have investigated the dimensionality reduction of high dimensional feature spaces, present in many biometric systems such as face, palmprint where the Log Gabor features have been shown as an efficient tool to represent the modalities at hand. We proposed a novel feature selection scheme based on an adaptation of the standard PSO algorithm. In the PSO algorithm, each particle adapts its position using the same inertial factor, which in general is chosen as linearly decreasing during the iteration process, while in the proposed AIPSO algorithm, each particle chooses its inertial factor iteratively according to its fitness value which prevent getting stuck in a local minimum. We performed extensive experiments on two unimodal biometric modalities, namely face and palmprint and also on the fusion of these modalities at feature level on large databases of 250 users. These experiments indicate that the proposed AIPSO outperforms the state-of-the-art schemes by reducing the feature space of around 62% of features and achieving nearly 5% gain in performance with 90% confidence interval. Let us note that when compared to the normal PSO approach which already allows reducing significantly the number of features, the proposed AIPSO scheme has also the advantage of improving the performance of the systems which could not be achieved by any of the other selection schemes that we tested. These figures clearly indicate the efficacy of the proposed selection scheme using AIPSO for unimodal and multimodal biometric context. R EFERENCES [1] A.Ross, K.Nandakumar, and A. Jain, Handbook of Multibiometrics, springer-Verlag ed., 2006. [2] D. Zhang, Palmprint Authentication, springer-verlag ed., 2004. [3] D. Kisku, J. K. Singh, M. Tistarelli, and P.Gupta, “Multisensor biometric evidence fusion for person authentication using wavelet decomposition and monotonic-decreasing graph,” in Proceedings of 7th International Conference on Adavnaces in Pattern Recognition (ICAPR-2009), Kolkata, India, 2009, pp. 205–208. [4] R.Raghavendra, B. Dorizzi, A. Rao, and G. H. Kumar, “Designing efficient fusion schemes for multimodal biometric system using face and palmprint,” Pattern Recognition, vol. 44, no. 5, pp. 1076–1088, 2011. [5] G. Feng, K. Dong, D. Hu, and D. Zhang, “When faces are combined with palmprints:a novel biometric fusion strategy,” in First International Conference on Biometric Authentication (ICBA), 2004, pp. 701–707. [6] R. Raghavendra, B. Dorizzi, A. Rao, and G. H. Kumar, “PSO versus adaboost for feature selection in multimodal biometrics,” in Proceedings of Third IEEE International Conference on Biometrics: Theory, Application and System- BTAS-2009, Washington, DC, USA., 2009, pp. 238–244. [7] Y.Yao, X. Jing, and H. Wong, “Face and palmprint feature level fusion for single sample biometric recognition,” Nerocomputing, vol. 70, no. 7-9, pp. 1582–1586, 2007. [8] X. Jing, Y. Yao, J. Yang, M. Li, and D. Zhang, “Face and palmprint pixel level fusion and kernel DCV-RBF classifier for small sample biometric recognition,” Pattern Recognition, vol. 40, no. 3, pp. 3209–3224, 2007.

TABLE III C OMPARATIVE PERFORMANCE OF THE DIFFERENT FEATURE SELECTION SCHEMES ON PALMPRINT V ERIFICATION (M EAN GAR AT FAR = 0.01%) Methods

Dori

DF S

DKDDA

Full features FS-SFFS [9] FS-AdaBoost [6] FS-GA [10] FS-PSO [4] FS-AIPSO

3200 3200 3200 3200 3200 3200

3200 2920 2610 2372 1872 1614

110 103 94 78 64 54

GAR at 0.01% of FAR(%) with 90% confidence interval 74.47 [72.41; 76.53] 74.58 [72.57; 76.59] 75.51 [73.41; 77.61] 74.75 [72.61; 76.89] 76.50 [74.58; 87.42] 82.25 [80.58; 83.92]

TABLE IV C OMPARATIVE PERFORMANCE OF THE DIFFERENT FEATURE SELECTION SCHEMES ON PALMPRINT R ECOGNITION (M EAN R ECOGNITION R ATE )(C LOSED I DENTIFICATION ) Methods

Dori

DF S

DKDDA

Full features FS-SFFS [9] FS-AdaBoost [6] FS-GA [10] FS-PSO [4] FS-AIPSO

3200 3200 3200 3200 3200 3200

3200 2834 2497 2204 1694 1320

110 98 89 72 59 49

Recognition Rate(%) with 90% confidence interval 78.41 [76.46; 80.35] 78.51 [76.41; 80.61] 79.01 [76.89; 81.13] 79.21 [77.05; 81.37] 80.21 [78.26; 82.16] 87.47 [85.77; 89.17]

TABLE V C OMPARATIVE PERFORMANCE OF THE DIFFERENT FEATURE SELECTION SCHEMES FOR F EATURE F USION (M EAN GAR AT FAR = 0.01%)(V ERIFICATION ) Methods

Dori

DF S

DKDDA

Full features FS-SFFS [9] FS-AdaBoost [6] FS-GA [10] FS-PSO [4] FS-AIPSO

6400 6400 6400 6400 6400 6400

6400 5286 4090 3855 3520 2954

224 207 184 170 139 94

GAR at 0.01% of FAR(%) with 90% confidence interval 93.70 [92.70; 94.70] 92.55 [91.32; 93.77] 92.88 [91.76; 93.28] 92.75 [91.52; 93.97] 94.72 [93.85; 95.59] 98.26 [97.26; 98.91]

TABLE VI C OMPARATIVE PERFORMANCE OF THE DIFFERENT FEATURE SELECTION SCHEMES FOR FEATURE FUSION (R ECOGNITION R ATE )(C LOSED I DENTIFICATION ) Methods

Dori

DF S

DKDDA

Full features FS-SFFS [9] FS-AdaBoost [6] FS-GA [10] FS-PSO [4] FS-AIPSO

6400 6400 6400 6400 6400 6400

6400 5172 3972 3701 3314 2870

224 201 179 165 128 89

[9] A. Ross and R.Govindarajan, “Feature level fusion using hand and face biometrics,” in Proceedings of SPIE Conference on Biometric Technology for Human Identification, 2004, pp. 196–204. [10] A. A. Altun, H. E. Kocer, and N. Allahverdi, “Genetic algorithm based feature selection level fusion using fingerprint and iris biometrics,” IJPRAI, vol. 22, no. 3, pp. 585–600, 2008. [11] J. Kennedy and R. Eberhart, “Particle swarm optimization,” in IEEE International Conference on Neural Networks, Perth, Australia, 1995, pp. 1942–1948. [12] Shi.Y and E. R. C, “A modified particle swarm optimizer,” in Proceedings of IEEE International Conference of Evolutionary Computation, Anchorage, Alaska., 1998, pp. 69–73. [13] Shi.Y and Eberhart.R.C, “Fuzzy adaptive particle swarm optimization,” in Proceedings of the Congress on Evolutionary Computation, Seoul, Korea, 2001, pp. 101–106.

Recognition Rate(%) with 90% confidence interval 95.76 [94.14; 96.17] 96.20 [95.00; 97.40] 95.87 [95.04; 96.69] 95.78 [96.64; 96.92] 96.01 [95.08; 96.93] 99.92 [99.84; 99.98]