J. Cent. South Univ. (2017) 24: 2360-2371

DOI: https://doi.org/10.1007/s11771-017-3648-9

Fuzzy least brain storm optimization and entropy-based Euclidean distance for multimodal vein-based recognition system

Dipti Verma¹, Sipi Dubey²

1. Department of CSE, Vishwavidyalaya Engineering College, Lakhanpur, Sarguja University,Chhattisgarh 497001, India;

2. Department of CSE, Rungta College of Engineering and Technology (RCET), Bhilai, Chhattisgarh 490024, India

Central South University Press and Springer-Verlag GmbH Germany 2017

Abstract: Nowadays, the vein based recognition system becomes an emerging and facilitating biometric technology in the recognition system. Vein recognition exploits the different modalities such as finger, palm and hand image for the person identification. In this work, the fuzzy least brain storm optimization and Euclidean distance (EED) are proposed for the vein based recognition system. Initially, the input image is fed into the region of interest (ROI) extraction which obtains the appropriate image for the subsequent step. Then, features or vein pattern is extracted by the image enlightening, circular averaging filter and holoentropy based thresholding. After the features are obtained, the entropy based Euclidean distance is proposed to fuse the features by the score level fusion with the weight score value. Finally, the optimal matching score is computed iteratively by the newly developed fuzzy least brain storm optimization (FLBSO) algorithm. The novel algorithm is developed by the least mean square (LMS) algorithm and fuzzy brain storm optimization (FBSO). Thus, the experimental results are evaluated and the performance is compared with the existing systems using false acceptance rate (FAR), false rejection rate (FRR) and accuracy. The performance outcome of the proposed algorithm attains the higher accuracy of 89.9% which ensures the better recognition rate.

Key words: multimodality; brain storm optimization (BSO); least mean square (LMS); score level fusion; recognition

1 Introduction

Due to the rapid development of computer techniques and information sciences, the person identification is the facilitating and challenging task for access control systems. Biometric recognition is done by the distinct human traits such as facial images, hand vein, fingerprints and palm prints [1]. In unimodal biometric system, it consists of some constraints like noisy data, intraclass variations, unacceptable error rates, spoof attack and restricted degree of freedom, which lead to degradation of the recognition rate. To resolve this problem, nowadays the biometric recognition system is performed by the multimodalities for the person identification [2]. Thus, the multimodal biometric recognition becomes the prominent role for the security purpose where more than one modality is fused together. The verification and identification are the two major concerns in the biometric recognition system. The verification phase is used to validate whether the person is accepted or rejected by the predetermined features. Then, the identification is employed to verify the person identity with its corresponding features [3].

Among the different biometric traits, the vein based recognition system is employed recently to enhance the recognition performance rate since every person has various vein patterns even differentiate the identical twins also [4]. Normally, the vein patterns of the human are formed by the subcutaneous blood vessels. It has unique characteristics since it lies inside the skin which has the advantage of contact less manner, ensuring the liveliness and robust against the forgery [5]. The main characteristics of vein patterns are uniqueness, stability and vigorous immunity, which provide better secure and reliable recognition for the person identification. As the initial step of image acquisition, the vein patterns of the human are captured by the infrared imaging technologies [6]. In general, the vein images are extracted from the finger part, dorsal part and palm part of the hand. It gives the most significant part of texture information, increasing the differentiating ability among the people and improving the recognition performance [7].

The vein based recognition system constitutes region of interest (ROI) extraction, image enhancement, vein extraction and feature matching. The ROI extraction is one of the preprocessing steps to suppress the noisy pixels which acquire the better quality image and enhance the image contrast [8]. Then, the vein patterns are extracted significantly from the finger, dorsal and palm part of the hand. Thus, the vein extraction is performed by the global or adaptive threshold mechanism [9]. The feature vector is the main aspect in the biometric recognition system. Thus, the features of the vein patterns, such as line features, point features, geometry features, texture and statistical features, are extracted by the effective feature extraction method [10]. Finally, the fusion of the features is done by the feature extraction level, matching score level and decision level, which is then used for the matching and decision making module for the person identification [11].

In this work, the fuzzy least brain storm optimization (FLBSO) algorithm is proposed for the vein based recognition system. The input images of hand, finger and palm vein images are considered for the recognition system. Initially, the input images are fed into the preprocessing step where the ROI extraction is utilized. The ROI extraction provides the most significant part of the vein images assist by the neighbourhood search algorithm. Then, the resultant image is given as input to the enlightened image and circular averaging filter. The holoentropy is used to determine the optimal threshold value for the vein extraction. Thus, the feature information of vein images is obtained. Finally, the features got fused by the newly proposed FLBSO algorithm. The score level fusion is employed where the new evaluation method of entropy based Euclidean distance (EED) is proposed for matching the query and input image with their weight score value. To enhance the performance, the optimal weight score is determined iteratively by the proposed FLBSO algorithm.

The main contributions of this work are as follows: The entropy Euclidean distance (EED) is proposed for the fusion of three modalities feature by the score-level fusion. The matching weight score in the fusion is optimally calculated by the newly proposed fuzzy least brain storm optimization (FLBSO) where the LMS is integrated with the FBSO algorithm.

This work is structured as follows: Section 2 discusses the vein based recognition system from eight research papers. The problem statement and challenges behind the system are described in Section 3. Section 4 briefly explains the proposed methodology using EED and FLBSO algorithm. Experimental results are validated and performance is analyzed in Section 5. Finally, this paper concludes in Section 6.

2 Literature review

JOARDAR et al [12] presented a palm dorsa subcutaneous vein pattern (PDSVP) for the real-time recognition system. It had been developed by the data acquisition using the two-axis pan-till mechanism where the palm dorsum was positioned. The NIR images of the PDSVP were acquired, in the aforementioned methodology the vein pattern did not represent with appreciable clarity and discernibility. Then, the input image underwent the preprocessing step to extract the vein patterns. Finally, the recognition was performed by the collaborative representation based classification. Thus, the performance of PDSVP was analyzed which was robust even in the presence of artifacts.

SHEKHA et al [2] described the multimodal sparse representation method, which represents the test data by a sparse linear combination of training data, while constraining the observations from different modalities of the test subject to share their sparse representations. A multimodal quality measure was also employed to weigh each modality as it gets fused. Furthermore, we also kernelized the algorithm to handle nonlinearity in data. Thus, the optimization problem was solved using an efficient alternative direction method. Various experiments showed that the multimodal sparse representation method was compared favourably with competing fusion-based methods.

WANG et al [11] explained multimodal personal identification system using palmprint and palm vein images with their fusion applied at the image level. The palmprint and palm vein images were fused by an edge-preserving and contrast-enhancing wavelet fusion method in which the modified multiscale edges were combined. Then, the features were extracted from the fused image by the developed Laplacianpalm representation using locally preserving projections. “Laplacianpalm” was utilized to find an embedding that preserves the local information and provides a palm space best detecting the essential manifold structure. Experimental results provided a better representation and achieved lower error rates in palm recognition. Furthermore, the Laplacianpalm representation multimodal method outperformed any of its individual modality.

GUPTA [5] demonstrated the efficient multi-modal authentication system which makes use of palm-dorsa vein pattern. Multi-algorithm fusion was applied to extract genuine vein patterns from a vein image by using various vein extraction algorithms. Then, three types of features are obtained from each extracted vein pattern which was shape features, minutiae and features obtained of the hand. Third level of fusion was at feature level to fuse minutiae and shape feature. Finally, fused features and hand boundary shape features were matched to obtain matching scores which were fused at score level. Thus, the multimodal recognition achieved the accuracy of 100%. Experimental results proved that it performs better than other existing systems.

WANG et al [3] presented a biometric identification system based on near-infrared imaging of dorsal hand veins and matching of the keypoints that were extracted from the dorsal hand vein images by the scale-invariant feature transform. In addition to several constraints introduced to minimise incorrectly matched keypoints, a particular focus was placed on the use of multiple training images of each hand class to improve the recognition performance. Then, the multiple keypoint sets were extracted from multiple training images of each hand class into three sets, namely, the union, the intersection and the exclusion, based on their inter-class and intra-class relationships. This work shows the contribution made by each set to the recognition performance and demonstrates the feasibility of achieving 100% correct recognition by combining the three sets, based on the experiments conducted using more than 2000 dorsal hand vein images.

LAJEVARDI et al [13] described an automatic dorsal hand vein verification system using a novel algorithm called biometric graph matching (BGM). Veins were extracted here using a maximum curvature algorithm. The matching performance of BGM for hand vein graphs was tested with two cost functions and compared with the matching algorithms, iterative closest point (ICP) and modified Hausdorff distance. Experiments were conducted on two public databases captured using far infrared and near infrared (NIR) cameras. BGM’s matching performance was competitive with state-of-the-art algorithms on the databases despite using small and concise templates. For both databases, BGM performed at least as well as ICP. For the small sized graphs from the NIR database, BGM significantly outperformed point pattern matching.

MURUKESH et al [4] presented a score level fusion of palmprint and finger vein for the biometric recognition system. Palmprint and finger vein images were fused according to the normalization scores of the individual traits. Then, the features of the palmprint were extracted by the discrete cosine transform (DCT) which was classified by multi-class linear discriminant analysis (LDA) and self-organizing maps (SOM). A multimodal biometric authentication system integrated the information from multiple biometric sources to compensate for the limitations in performance of each individual biometric system. These systems could significantly improve the recognition performance of a biometric system.

LEE et al [1] designed a directional filter bank to extract vein patterns and the minimum directional code was employed to encode line-based vein features into binary code. In addition, there were many non-vein areas in the vein image, which were not meaningful for vein recognition. To improve accuracy, the non-vein areas were detected by evaluating the variance of the minimum directional filtering response image and were considered non-orientation code. In total, 4280 dorsal hand vein images from 214 persons were used to validate the proposed dorsal hand vein recognition approach. A high accuracy (>99%) and low equal error rate (0.54%) were obtained using the filter bank approach, which proved that the approach was feasible and effective for dorsal hand vein recognition.

3 Motivation behind recognition system

3.1 Problem description

The main problem is to identify the person through the vein patterns of finger, palm and hand. Thus, recognition of human becomes the challenging technique throughout the world. Let D be the input database consisting of B number of persons, which is expressed asThen, every person poses V number of finger, palm and hand vein images, which is defined by . The image X is given as input to the proposed algorithm of size e×f. Here, the main challenge is to achieve the better recognition performance by the vein images of finger, hand and palm.

3.2 Challenges

Due to security problems such as uniqueness, acceptability and permanence, the biometric recognition system has been extensively studied to ensure the reliable system for the person identification [14].

The vein recognition system becomes the challenging task since the vein image contains large number of features, which is burden to the biometric system [15].

Fusion and matching [7] are the major challenges in the recognition system since the features of different modality are of great complementary and then matching between testing and training image ensures the authorized person significantly.

4 Proposed methodology: Vein-based recognition system using entropy Euclidean distance and FLBSO algorithm

The ultimate aim of this work is to recognize the individuals by their vein images of hand, finger and palm. Here, the V number of vein images is considered the input image. The proposed methodology constitutes four following steps: I) preprocessing, II) vein extraction, III) feature extraction and IV) recognition. Figure 1 depicts the diagrammatic representation of proposed methodology. Initially, the input images are fed into the preprocessing steps to acquire the image concisely for the subsequent steps. Here, the ROI extraction is utilized in the preprocessing step based on the neighbourhood search. Then, the vein regions are extracted by the enlightening of image, circular averaging filter and optimum threshold using holoentropy based mechanism. The features are extracted from the acquired hand vein, finger vein and palm vein images. After the features are extracted, the entropy Euclidean distance (EED) is proposed for the matching using score level fusion. Also, fuzzy least brain storm optimization (FLBSO) is newly designed where the least mean square is integrated with the fuzzy based brain storm optimization. Thus, the novel FLBSO algorithm is used to determine the optimal score value iteratively for the recognition system.

4.1 Preprocessing

The preprocessing is the initial step of the vein based recognition system. The preprocessing is used to remove the noise present in the input image, which obtains the image concisely for the further steps. Here, the ROI extraction by neighbourhood search algorithm [16] is utilized. The ROI extraction is employed to extract the most significant vein part from the hand, finger and palm vein images. The neighbourhood search algorithm is performed by the Euclidean distance between the template image and extracted image by the reference point. Thus, the ROI extraction is apparently described as follows:

I) Here, m reference points are considered in the neighbourhood search algorithm. Thus, the significant regions of the vein are extracted using the reference points which are defined as

where represents the extracted regions based on the reference points and f is the function to generate the sub image.

II) Then, the Euclidean distance is calculated between the template image and extracted region from the input image.

III) Finally, the significant part of vein is extracted by the reference point which has the minimum Euclidean distance between the template and input image. It is determined by

4.2 Vein extraction using holoentropy based mechanism

The resultant ROI image is given as input to the vein extraction. Here, the vein patterns of hand, finger and palm part are determined by the three steps: I) enlightening of image, II) circular averaging filter and III) holoentropy mechanism. Thus, the vein extraction is described as follows.

Fig. 1 Diagrammatic representation of proposed methodology

I) Enlightening of image

The image enlightened is the process to enhance the darken pixels, which is used to extract the vein regions significantly. Here, the exponential kernel is used to convert the input image into x number of different images, which is used for edge detection, embossing and sharpening of the vein regions. Thus, the enlightened image is calculated as

where e₁ is the kernel vector; h represents the random integer of size 1×n and T_a is the user given threshold. The vein image is highly dependent on the kernel function which has the specific regions of the vein image. Based on the kernel, finally, the vein image is enhanced for the vein extraction which is determined by

II) Circular averaging filter

The circular averaging filter [17] is also employed to improve the pixel intensity of the vein image. The averaging filter is performed by calculating the average of each value in the vein extracted image. The idea behind this filter is to estimate the average value of each pixel with respect to its neighbours. The circular averaging filter is utilized in which the radius is fixed as five. Thus, the quality of the image is enhanced by each pixel value and replaced by the average value of the neighbour pixels in circular manner.

III) Holoentropy based mechanism

Finally, the vein parts of the finger, hand and palm are extracted by the optimum threshold using holoentropy mechanism [18]. The thresholding is an efficient way to extort the significant part of the vein image. Though the entropy is used as the global measure, which is not enough to segment the vein regions from the hand, palm and finger. In order to solve this problem, the holoentropy based mechanism is employed. Thus, the M and N are the two matrices obtained from the circular pixel matrix. Thus, CP matrix is generated by the threshold value t which is determined as I=[t×t] and .

The weight factor is incorporated with the holoentropy to attain the better recognition performance. Here, the reverse sigmoid function is used as the weight factor for the vein extraction. The main advantage of holoentropy measure is significantly and quickly extracting the vein regions rather than the other measure. Thus, the holoentropy is estimated for two matrices by the product of the entropy measure and weight function. It is derived as follows:

where is the holoentropy measure; represents the entropy value and is the reverse sigmoid function which is expressed by

Similarly, the holoentropy for the matrix N is calculated as

where

Then, the optimal threshold value is obtained by the summation of the holoentropy measure of the two matrices. Thus, the vein regions are extracted appropriately from the image using the optimal threshold value. It is defined by

where Z is the final optimal threshold value.

4.3 Feature extraction

The extracted vein regions are fed into the feature extraction process for the person identification. The feature vector is more important since it exhibits significant representation of the vein image and used to mitigate the non-linear complexity for the classifier. Due to the pattern of blood vein, the system provides robust against the security imposter. Thus, the extracted feature information represents by two coordinates. For the finger vein image is defined as the hand vein feature extraction is represented as and then finally, the extraction of palm vein feature is denoted by Then, the acquired feature is given into the proposed FLBSO algorithm for the biometric recognition system.

4.4 Proposed FLBSO algorithm

Once we extract the feature, the proposed FLBSO algorithm is used to perform the person recognition. The extracted features of hand, palm and finger vein images get fused by the proposed entropy Euclidean distance (EED) based score level fusion. In this fusion, we exploit weight score which is determined iteratively by the proposed FLBSO algorithm.

4.4.1 Score level fusion by proposed EED

This section presents the score level fusion using entropy Euclidean distance for the recognition system.The score level fusion is used to fuse the features of hand, palm and finger vein image based on the matching score. It is also called matching score level. The fusion at matching level poses two main aspects: 1) While in the classification approach, the feature vectors are trained and tested with the testing image which leads to classifying the legitimate and illegitimate person and 2) individual matching score of the feature is combined, which is fed into the decision making module.

EED: Before determining the score value for the different modalities, the matching is an important aspect for the feature level fusion. Due to the matching results different from each other, the entropy based Euclidean distance (EED) is proposed in this work. Thus, the matching results for vein based recognition are defined as the score value. The Euclidean distance measure degrades the performance by the high sensitivity and even contains a small deformation in the vein image. In order to solve this issue, we propose the EED based score level fusion for the vein recognition since the entropy measure has been efficiently histogram equalized and robust against the noisy pixels.

Matching: The proposed EED is used for matching by the distance measure between the user query image and the input image in the database. Consider that is the extracted feature of the query image for three modalities and represents the extracted feature of the hand, palm and finger vein images. The EED is derived as follows:

where w is the entropy measure; l is the number of training images; z represents the total number of pixels in the image and p_i represents the probability associated with the pixel value. The entropy value is evaluated by the summation of each pixel in the l number of training images. Then, the matching is done through the entropy based Euclidean distance (EED) for finger, hand and palm vein images.

where A_F, A_H and A_P are the matching score values for three modalities. Then, based on the matching score, the features are fused by the score level fusion with constant weight score α, β and γ. It is determined as

4.4.2 FLBSO algorithm

The novel FLBSO algorithm is proposed for vein based recognition where the LMS algorithm [19] is integrated with the fuzzy brain storm optimization (FBSO). The main aim of this algorithm is used to determine the optimal weight score which varies from zero to one. Initially, the BSO [20] optimization algorithm is inspired by the behaviour of human beings, which is performed by the clustering operation. It has the advantage to resolve the exploitation and exploration problem. The cluster is formed by the grouping of the features where the best value is considered the cluster centre. Then, the new solution is generated assist by LMS and FBSO. Thus, the proposed FLBSO algorithm is apparently explained as follows.

Solution encoding: The solution encoding is an important aspect for the proposed FLBSO algorithm. In this algorithm, the c number of ideas is randomly generated for the first iteration. It is used to encode the constant weight values with c number of ideas. Thus, each solution is stored in the form of three dimensions, which ranges from 0 to 1.

Fitness function: Each dimension contains one variable and it is determined by the fitness function. After the fitness value is obtained, the searching algorithm is performed by the human ideas and LMS algorithm. Thus, the fitness value is determined by the false acceptance rate and false rejection rate.

Algorithmic elucidation:

1) In brain storm optimization [20] algorithm, the c number of ideas is generated randomly with d dimensions, which is used to determine the optimal weight score value for the vein based recognition. It is represented by

where is the number of ideas and d represents the dimension.

2) During each generation, the k-means clustering is utilized to group the c number of ideas. Thus, the old value is replaced by the new solution, which is determined by the proposed FLBSO algorithm where the LMS [19] algorithm is integrated with the FBSO optimization. The standard equation for the LMS algorithm is expressed as

The above equation is written as

where is the newly generated cluster centre using LMS algorithm and R_t denotes the current solution; μ is the constant value and the error e which is estimated by the

3) The FBSO algorithm is utilized where the fuzzy theory is incorporated with the brain storm optimization. In order to enhance the correlation between the images, fuzzy rules are used in this work. The optimization problem of BSO algorithm is resolved by the cluster formation of the individual ideas. Thus, for the further iteration, the cluster centre is newly generated by the old value with the Gaussian function of random values. The standard equation of the BSO [20] algorithm is defined by

where R_t+1 represents the new solution by the BSO algorithm. Then, the advantage of fuzzy theory [10] is simplic and flexibile, which also improves the recognition performance. Due to the non feasible of optimum solution, we employ the fuzzy weighted parameter for the generation of new solution. Thus, we utilized the FBSO which is integrated with the least mean square algorithm. We named the fuzzy least brain storm optimization (FLBSO). The LMS algorithm provides the robustness against the fraudulent users, reduces the computational complexity and has well convergence behaviour for the recognition. Finally, the new solution is generated for every iteration with the aid of new cluster centre of LMS and random values of individuals. Thus, each term is multiplied by the fuzzy weighted parameter to enhance the recognition rate.

where an old individual in the FBSO algorithm is replaced by the new solution of the LMS algorithm and error value which ensures the better recognition performance. Then, the fuzzy weighted parameters are

where d denotes the dimension and is the Gaussian random function which is expressed by

where p and q define the maximum number of iteration and current iteration; lg sig is the logarithmic sigmoid function which is performed by the r altering slope value and represents the random individual value ranging between 0 and 1.

4) The above process is repeated until the c number of ideas is updated. Thus, the optimal weight score value is obtained which is used to fuse the features of the hand, palm and finger vein images. Thus, the vein based recognition is performed using the proposed FLBSO algorithm.

FLBSO algorithm

1  Input: c number of ideas

2  Output: Optimal matching score

3  Procedure

4  Begin

5  Randomly initialize c ideas

6  Find the fitness of the randomly initialised ideas

7  Find the best idea

8  While

9  Cluster c ideas into g clusters

10 Generate new ideas based on randomly select cluster centre and replace an old individual by the generation of new cluster centre using LMS

11 Randomly select another one cluster centre and replace an individual by adding the random vector with this cluster centre

12 For (i=1 to MaxGen)

13 Create using the formula

14 Generate if is better than

15 End for

16 Update the best idea

17 End while

18 End

5 Results and discussion

This section presents the experimental results for the proposed FLBSO algorithm based recognition system. Also, the performance is analyzed by the evaluation metrics which are compared with the existing systems.

5.1 Experimental setup

1) Dataset description. For the experimentation, three benchmark databases, a) SDUMLA-HMT finger vein database [21], which is used to capture finger vein images is designed by Joint Lab for Intelligent Computing and Intelligent Systems of Wuhan University. Every image is stored in "bmp" format with 320×240 pixels in size and thus, the finger vein database takes up around 0.85G Bytes in total. b) Bosphorus hand vein database [22] is acquired using NIR imaging technology with a monochrome NIR CCD camera equipped with an infrared lens. The back of the hand is irradiated by two IR light sources. The images have 300×240 pixel size with a gray-scale resolution of 8-bit. c) CASIA multi-spectral palmprint image database [23] where the palm images are 8 bit gray-level JPEG files. Each sample contains six palm images which are captured at the same time with six different electromagnetic spectrums. Overall, we have taken the 100 persons and every person contains left and right hand images of 6 samples.

2) Evaluation parameters. The performance for the vein based recognition system is analyzed by the metrics: false acceptance rate (FAR), false rejection rate (FRR) and accuracy. The evaluation metrics are calculated by the true and false positive (TP and FP), true and false negative (TN and FN). The evaluation metrics are described below.

3) Experimental databases. We create the two bases for our experimentation to recognize the individual. Let db1 be the database which consists of fused feature of left hand vein images of finger, palm and hand. And also, let db2 be the database of the fused features of right hand vein images.

4) Algorithms taken for comparison. Product and sum rule: It is based on the assumptions of statistical independent variable. This rule is used where the fused score is computed by multiplying the scores of different modalities of the hand vein images. Similarly, the sum rule based recognition [21] is performed by the normalization scores of its corresponding person. Here, the fused score is obtained by the summation of the normalized scores with its weights. It is used to compute the posterior probability of individuals which is used to classify the features.

BSO: It is performed by the number of individual ideas to determine the optimal solution. The BSO algorithm resolves the optimization problem while recognizing the person identification. For every iteration, the solution is newly generated by the old individual value and Gaussian function with mean and variance.

FBSO: The FBSO algorithm is employed for the vein recognition system where the fuzzy theory is integrated with the BSO algorithm. The BSO is inspired by the individual idea which is used to generate the optimal matching score value with the aid of fuzzy weight parameters. Once the solution is obtained, the matching is done by the testing image to provide the recognition rate.

FLBSO: The features of hand, palm and finger vein get fused by the score level fusion. Here, the EED is proposed to fuse the features along with the matching weight score value. To enhance the performance, the matching score is computed optimally by the newly proposed FLBSO algorithm. Thus, LMS algorithm is incorporated with the FBSO algorithm to ensure the better recognition performance.

5) Experimental results. Figure 2 depicts the experimental results for the recognition system using different modalities. Figure 2(a) shows the input image of finger, hand and palm images. Then, the input image is fed into ROI extraction which is represented in Fig. 2(b). Consequently, the resultant image is given as input to the feature extraction step where the features are obtained by the holoentropy based thresholding mechanism. Finally, these features are recognized significantly by the proposed EED and FLBSO algorithm.

5.2 Performance analysis

The performance is analyzed by FAR, FRR and accuracy and is then compared with the existing systems like product rule, sum rule, BSO and FBSO.

1) Comparative performance by filter size

Figure 3 shows the comparative performance analysis for the recognition system using filter size. FAR is defined as the error measure which should be low to ensure the better recognition. Figure 3(a) depicts the performance analysis for the database db1. When the filter size is 5, the existing system like FBSO algorithm attains 18.9% FAR, BSO achieves 19%, and sum rule 18.9% and product rule 19.3%. But, the proposed FLBSO algorithm attains the minimum value of 11.6%. While the filter size is seven, the proposed FLBSO method achieves 11.1% rate of false acceptance which proves the better recognition performance. Subsequently, Fig. 3(b) represents the comparative performance analysis for the database db2. The existing product rule method acquires 19.5%, 19.1% and 19.2% while varying the filter size. Comparing to the existing systems, the proposed method obtains 12.4% when filter size is three, 11.9% when filter size is five and 11.4% FAR when filter size is seven. Thus, the proposed FLBSO algorithm achieves the low error for the vein based recognition system.

Figure 4 depicts the performance analysis by the false rejection rate. The FRR is the measure in which the system wrongly rejects an authorized user. The comparative performance analysis for the database db1 is shown in Fig. 4(a). The existing sum and product rule achieve 10.1% and 10% when the filter size is three, five and seven. Compared to the existing systems, the proposed FLBSO algorithm attains 0.94%, 0.9% and 0.94% for the filter size is three, five and seven, respectively. Finally, the low value of FRR is attained which is represented in Fig. 4(a). Similarly, Fig. 4(b) shows the comparative performance analysis for the right hand databases. By varying the filter size for the vein recognition system, product rule acquires 10.1%, sum rule attains 10.09%, brain storm optimization obtains 10.09% and then fuzzy BSO algorithm achieves 10.09%. Compared with the existing systems, the FLBSO algorithm achieves 0.88% error rate which leads to attaining the higher accuracy value for the person identification as shown in Fig. 4(b).

Fig. 2 Experimental results:

Fig. 3 FAR performance:

Fig. 4 FRR performance:

The accuracy performance analysis for the vein based recognition system is shown in Fig. 5. The accuracy is an efficient measure for the recognition system where the higher value indicates to recognize the authorized person significantly. Figure 5(a) depicts the comparative performance for the left hand databases. Due to the minimum value of error, the proposed system achieves the higher accuracy value. When exploiting the filter size three in the circular averaging filter, the proposed FLBSO algorithm attains the higher accuracy value of 87.9% which is compared to the sum rule, product rule, brain storm optimization and fuzzy based BSO algorithm since it achieves 80.7% accuracy value. Consequently, Fig. 5(b) represents the comparative performance analysis for the database db2. Based on the filter size, 81.3% of accuracy is obtained by the FBSO, the BSO algorithm attains 81% accuracy, then, 80.9% and 80.8% accuracy for the sum rule and product rule method. But, the proposed FLBSO algorithm achieves the maximum accuracy 88.6%.

Fig. 5 Accuracy performance for vein based recognition system:

2) Comparative analysis by training data

Figure 6 demonstrates the comparative performance analysis for the two databases based on the percentage of training data samples. The existing brain storm optimization algorithm achieves the false acceptance rate of 19.3% while using 70% of training data samples. Then, the FAR is decreased to 18.7% for the vein recognition. But, the proposed system initially attains 12.1% of false rate which is gradually decreased to the lower value of 10.1% while increasing the percentage of training data as is shown in Fig. 6(a). Similarly, Fig. 6(b) shows the performance analysis for the right hand databases. When the percentage of training data sample is 80, the FBSO attains 18.9% FAR, BSO existing algorithm acquires 19.4% FAR, another existing sum and product rule obtains 19.5% and 19.6% false rates. When compared to the existing algorithms, the proposed FLBSO algorithm obtains 11.4% FAR value as shown in Fig. 6(b). Thus, the proposed method acquires the lower error of 10.4% when compared to the FBSO, BSO, sum rule and product rule.

Fig. 6 Comparative analysis for FAR:

Figure 7 represents the FRR comparative performance based on the percentage of training data samples. While the percentage of training samples is 85%, the all existing algorithms achieve the same error value of 10.09%. But, the proposed FLBSO algorithm obtains the lower error of .94% till 80% of training data and also 88% of FRR value is achieved for 85% and 90% of data samples as shown in Fig. 7(a). Similarly, Fig. 7(b) depicts the comparative analysis for the right hand databases. When the percentage of training data is increased from 70% to 90%, the existing FBSO, BSO, sum rule and product rule acquires 10.1% FRR whereas the proposed algorithm obtains 0.87% of false rejection rate which is represented in Fig. 7(b). Thus, we infer from Fig. 7 that the minimum error value is obtained for the left and right hand vein images. Thus, the proposed system ensures to attain the higher accuracy value when compared with the FBSO, BSO, sum rule and product rule.

Figure 8 depicts the accuracy performance analysis for both database db1 and db2. The accuracy is a significant measure to analyze the vein based recognition system. Figure 8(a) depicts the accuracy comparative performance for the left hand vein images. When the percentage of training data sample is 70%, the existing BSO algorithm obtains 80.7% accuracy value. Then, it is gradually increased to 81% while increasing the percentage of training data. But, the proposed FLBSO algorithm achieves the higher accuracy 89.9% based on the size of training data samples which is demonstrated in Fig. 8(a). Simultaneously, Fig. 8(b) shows the comparative performance for the database db2. While using 75% of training data for recognition, the FBSO algorithm acquires 81.1% accuracy, the BSO obtains 80.6%, and then 80.5% and 80.4% are attained by the sum rule and product rule, respectively. However, the proposed vein recognition method achieves the higher accuracy 88.1% rather than the existing systems. Thus, the maximum accuracy of 89.6% is obtained while using the right hand vein images, which proves the better recognition system.

Fig. 7 Comparative analysis for FRR:

Fig. 8 Comparative analysis for accuracy:

6 Conclusions

We have presented the FLBSO algorithm and EED based score level fusion for the vein recognition system. Here, the different modalities of hand, palm and finger vein images were considered. The input images were given into the preprocessing step where ROI extraction was utilized. Thus, the image was obtained concisely for the further steps. Then, the significant part of vein patterns was extracted from the image. The extraction was done by the three steps namely enlightened image, circular averaging filter and holoentropy. Once we extract the features, the EED based score level fusion was proposed to fuse the feature using matching. Then, the LMS algorithm was integrated with the FBSO algorithm to estimate the matching weight score value. Finally, the optimal matching score was obtained by the proposed FLBSO algorithm. The experimental results were validated and the performance was analyzed by the evaluation metrics and then compared with FBSO, BSO, and sum and product rule. The proposed FLBSO algorithm achieves the higher accuracy of 89.9% value for the vein based recognition system.

References

[1] LEE J C, LO T M, CHANG C P. Dorsal hand vein recognition based on directional filter bank [J]. Signal, Image and Video Processing, 2016, 10(1): 145-152.

[2] SHEKHA S, PATEL V M, NASRABADI N M, CHELLAPPA R. IEEE Transactions on Pattern Analysis and machine Intelligence, 2014, 36:(1): 113-126.

[3] WANG Y D, ZHANG K, SHARK L K. Personal identification based on multiple keypoint sets of dorsal hand vein images [J]. IET Biometrics, 3(4): 234-245.

[4] MURUKESH C, THANUSHKODI K, PREETHI P, MOHAMED F N. Multimodal biometric authentication system based on score-level fusion of palmprint and finger vein [J]. Artificial Intelligence and Evolutionary Algorithms in Engineering Systems, 2014, 324: 551-560.

[5] GUPTA P, GUPTA P. Multi-modal fusion of palm-dorsa vein pattern for accurate personal authentication [J]. Knowledge based systems, 2015, 81: 117-130.

[6] YU Cheng-bo, QIN Hua-feng, CUI Yan-zhe, HU Xiao-qian. Finger-vein image recognition combining modified Hausdorff distance with minutiae feature matching [J]. Interdisciplinary Science Life, 2014, 4(1): 280-289.

[7] YAN Xue-kui, KANG Wen-xiong, WU Qiu-xia, DENG Fei-qi. Palm vein recognition based on multisampling and feature-level fusion [J]. Neurocomputing, 2015, 151(2): 798-807.

[8] GUPTA P, GUPTA P. An accurate finger vein based verification system [J]. Digital Signal Processing, 2015, 38: 43-52.

[9] LEE J C. A novel biometric system based on palm vein image [J]. Pattern Recognition, 2012, 33(12): 1520-1528.

[10] MIRON R, LETIA T. Fuzzy logic decision in partial fingerprint recognition [C]// IEEE International Conference on Automatic Quality and Testing Robotics: IEEE, 2010, 3: 1-6.

[11] WANG J G, YAU W Y, SUWANDY A, SUNG E. person recognition by fusing palmprint and palm vein images based on Laplacianpalm representation [J]. Pattern Recognition, 2008, 41(5): 1514-1527.

[12] JOARDAR S, CHATTERJEE A, RAKSHIT A. A real-time palm dorsa subcutaneous vein pattern recognition system using collaborative representation-based classification [J]. IEEE Transactions on Instrumentation and Measurement, 2015, 64(4): 959-966.

[13] LAJEVARDI S M, ARAKALA A, DAVIS S, HORADAM K J. Hand vein authentication using biometric graph matching [J]. IET Biometrics, 2014, 3(4): 302-313.

[14] LIN C L, FAN K C. Biometric Verification Using thermal images of palm-dorsa vein patterns [J]. IEEE Transactions on Circuits and Systems for Video Technology, 2004, 12(2): 199-213.

[15] LEE H C, KANG B J, LEE E C, PARK K R, Finger vein recognition using weighted local binary pattern code based on a support vector machine [J]. Journal of Zhejiang University, 2010, 11(7): 514-524.

[16] HEENAVE M, KHAN M. A multimodal hand vein biometric based on score level fusion [C]// Proceedings of International Symposium on Robotics and Intelligent Sensors, 2012: 897-903.

[17] Circular average filter. http://in.mathworks.com/help/images/ ref/fspecial.html?requestedDomain=www.mathworks.com.

[18] WU Shu, WANG Sheng-ru. Information-theoretic outlier detection for large scale categorical data [J]. IEEE Transactions on Knowledge and Data Engineering, 2013, 25: 3.

[19] NGUYEN D D, LEE J S J. A new LMS-based algorithm for rapid adaptive classification in dynamic environments [J]. Neural Networks, 1989, 2(3): 215-228.

[20] DUAN Hai-bin, LI Shuang-tian, SHI Yu-hui. Predator-prey brain storm optimisation for DC brushless motor [J]. IEEE Transactions on Magnetic, 2013, 49(10): 5336-5340.

[21] SDUMLA-HMT finger vein database [2016-08-25]. http://mla.sdu. edu.cn/sdumla- hmt.html.

[22] Bosplhorus hand vein database [2016-08-25]. http://bosphorus.ee. boun.edu.tr/hand/Home.aspx.

[23] CASIA multi-spectral palmprint database [2016-08-25]. http://biometrics.Idealtest.org/dbDetailForUser.do?id=6.

(Edited by YANG Hua)

Cite this article as: Dipti Verma, Sipi Dubey. Fuzzy least brain storm optimization and entropy-based Euclidean distance for multimodal vein-based recognition system [J]. Journal of Central South University, 2017, 24(10): 2360–2371. DOI:https://doi.org/10.1007/s11771-017-3648-9.

Received date: 2016-05-05; Accepted date: 2016-12-27

Corresponding author: Dipti Verma, Assistant Professor; E-mail: diptiverma2701@gmail.com