An improved intelligent water drops feature selection for finger vein recognition

Abstract

In recent years, finger vein recognition has gained a lot of attention and been considered as a possible biometric feature. Various feature selection techniques were investigated for intrinsic finger vein recognition on single feature extraction, but their computational cost remains undesirable. However, the retrieved features from the finger vein pattern are massive and include a lot of redundancy. By using fusion methods on feature extraction approaches involving weighted averages, the error rate is minimized to produce an ideal weight. In this research, a novel combinational model of intelligent water droplets is proposed along with hybrid PCA LDA feature extraction for improved finger vein pattern recognition. Initially, finger vein images are pre-processed to remove noise and improve image quality. For feature extraction, Linear Discriminant Analysis (LDA) and Principle Component Analysis (PCA) are employed to identify the most relevant characteristics. The PCA and LDA algorithms combine features to accomplish feature fusion. A global best selection method using intelligent water drops (GBS-IWD) is employed to find the ideal characteristics for vein recognition. The K Nearest Neighbour Classifier was used to recognize finger veins based on the selected optimum features. Based on empirical data, the proposed method decreases the equal error rate by 0.13% in comparison to existing CNN, 3DFM, and JAFVNet techniques. The overall accuracy of the proposed GBSPSO-KNN is 3.89% and 0.85% better than FFF and GWO, whereas, the proposed GBSIWD-KNN is 4.37% and 1.35% better than FFF and GWO respectively.

Keywords

Principle component analysis finger vein recognition linear discriminant analysis k-nearest neighbor intelligent water drops

1 Introduction

Several biometric technologies are available, such as fingerprints, DNA, palm prints, iris scans, ear scans, facial scans, voice scans, and finger vein scans. Blood that contains deoxygenated haemoglobin is more sensitive to near-infrared (NIR) light than blood from other tissues or from other skin types [1]. In the acquired picture, finger-vein patterns are visible as darker regions due to NIR illumination and the camera [2]. Due to low blood pressure, a vein’s shape is non-uniform, presenting distinguishing features for identification [3]. Finger veins are easy to get but difficult to fabricate due to their inside location [4]. However, shades and variations in light impair the visual observation of finger-vein patterns, resulting in a reduction in identification accuracy [5]. If noise induced by shades and variations in lighting in pictures is not appropriately adjusted during pre-processing, existing feature-based approaches for finger vein detection, like Discriminative Binary Descriptor and local binary pattern, have lower recognition accuracy [6 –8]. Feature extraction methods have been used to investigate intrinsic finger characteristics; however, they are still unsuitable for extracting big features because of high redundancy and computational expense. This research proposes a novel combinational model based on the aforementioned concerns. The major contribution of the proposed method is given as follows;

An authentication system based on finger veins is designed to improve the recognition rate of the system, and the fusion principle has been applied to feature selection and feature fusion.

Initially, the finger vein images are pre-processed to eliminate the noise and enhance image quality.

Consequently, principle component analysis and linear discriminant analysis techniques are used to identify high value features.

These selected features are concatenated using the PCA and LDA techniques for feature fusion.

The proposed system uses Global Best Selection Intelligent Water Drops [22] (GBSIWD) to determine the relevant features for vein recognition.

K Nearest Neighbour Classifier is used for Finger vein recognition of authorized person.

The proposed method’s performance is assessed using accuracy, sensitivity, specificity, and Equal Error Rate (EER).

There is no research paper previously available for the range of best features. Minimum features degrade the accuracy. Maximum features will take more time. The main novelty is finding the best k for accurate classification with minimum time.

The remaining paper is ordered in a resulting manner. literature survey is summarized in Section 2, Section 3 discuss the proposed model and related algorithms, Section 4 examines the suggested method’s performance and compares it to existing techniques. Finally, Section 5 brings the paper to a conclusion and future work.

2 Related works

H. Liu, [9] LBP or Gabor descriptors are used to extract the local feature, and it is then mapped to improve its discriminative performance by utilizing a discriminative feature learning approach. Multi-directional pixel difference vectors allow for the discrimination of this feature by comparing each pixel with its neighbouring pixel. The classification of finger veins is performed using a histogram of the characteristics of the veins.

A finger-vein extraction strategy was designed by Y.H. Liu and C. Liu Kim [10] using a random forest training method, a regression method, and an efficient local binary pattern feature. They accomplished cutting-edge finger-vein recognition by combining it with a vein pattern matching approach that is resistant to finger misalignment. Extensive tests were carried out on two prominent databases to demonstrate the efficacy and resilience of the suggested strategy.

Yang, J. et al. [11] used adaptive vector field estimation to establish an effective solution to finger-vein feature encoding. Because vein networks are built up of vein curve segments, an initial collection of spatial curve filters (SCFs) with varied curvature and orientation is formed. The experimental findings suggest that the proposed technique improves finger-vein matching accuracy significantly.

Darwish, S.M., [12] proposed a personal identification system based on finger veins. The vein pattern was created using the Gabor transformation by combining the vein images with local and global features. Comparing the performance of the suggested model with some of the most complex finger vein recognition systems, the results show an improvement of 6% in accuracy.

F. Saadat and M. Nasri [13] suggested a multi-biometric approach for human authentication. The suggested approach fused three different finger vein patterns that used a score-level fusion mechanism. The gravitational search Algorithm is employed in the suggested system to optimize the weights of the sum fusion technique. The experimental findings show that the suggested technique outperforms the standard fusion strategy in human detection.

Xiong et al. [14] provide a modified chaotic binary particle swarm optimization (MCBPSO) technique and it is used for feature selection so that the fusion of features at the face and iris level is reduced in complexity. It starts particle swarms using chaotic binary sequences and uses the kernel extreme learning machine (KELM) as a fitness function.

Garg et al. [15] have developed an automatic method for iris identification based on 2-D iris images. In this, for selecting a subset of iris features without compromising sensitive information, the 2DPCA (two-dimensional Principal Component Analysis) and GA (Genetic Algorithm) are being used as feature extraction and feature selection methods.

Xie, C., and Kumar, A., [16] proposed an approach based on supervised discrete hashing and CNN for fingerprint vein authentication. In this experiment, finger-vein data from two sessions is available for public use. When supervised discrete hashing is added to a CNN trained with a triplet-based loss function, the accuracy is higher.

Kang, W et al. [17] suggested a 3D reconstruction method for generating a full-view 3D finger vein image. In addition, a lightweight CNN with depth wise separable convolution is used for 3D finger vein matching and feature extraction. Although the suggested approach generates promising results for finger vein verification, much work has to be done to reduce time consumption and increase verification accuracy.

A joint attention (JA) module was proposed by Huang, J., et al. [18] that focuses on fine-grained features through dynamic adjustment and information aggregation of the channel and spatial dimensions of feature maps. In this manner, vein patterns can contribute to extracting identifying characteristics more effectively. According to the experimental results, the suggested technique performs better on numerous self-built and public datasets: EER of SDUMLA-HMT is 0.35% and FV-SCUT is 0.49%.

According to K. Kapoor et al. [19], a robust finger vein identification method was developed by combining hybrid Local Phase Quantization (LPQ) for robust feature extraction with Grey Wolf Optimization-based SVMs (GWO-SVMs) to optimize the SVM parameter combination. SVM and kNCN-SRC two-stage algorithms were outperformed by this technique, with a recognition accuracy of 98%.

According to the literature study, Various feature selection methods were investigated for intrinsic finger vein recognition on single feature extraction, but their computational cost remains undesirable. In this work, for Finger vein recognition, the optimization technique Intelligent Water Drops (IWD) is implemented with the proposed Global Best Solution technique to select the subset of features which achieves the best accuracy for KNN classification show in Fig. 1.

Fig. 1

Finger vein recognition system.

3 Proposed GBSIWD method

In this section, a novel combinational method of intelligent water drops with hybrid PCA_LDA feature extraction has been proposed to enhance the recognition of finger vein patterns. The Proposed model for finger vein recognition consist of following phases, which includes hybrid subspace feature extraction techniques such as PCA and LDA, feature fusion, Global Best Selection Intelligent water Drops optimization technique for feature selection and finally feature classification on the optimal set of features. Figure 2 depicts the entire workflow of the proposed finger vein recognition system.

Fig. 2

Overall workflow of the proposed model.

Initially, the pre-processed image will be subjected to feature extraction (PCA LDA), with two subspace approaches used to choose significant features. Those extracted features will be concatenated in order to perform feature fusion. To obtain the ideal set of features for the finger vein classifier, concatenated features are used in feature selection. In order to find the most useful subset of attributes for vein detection, Global Best Selection Intelligent Water Drops (GBS-IWD) was used. Finally, the K-NN classifier is utilised to classify or recognise features.

3.1 Principle component analysis feature extraction

For feature extraction, principal component analysis (PCA) is utilised to create the most distinctive features. PCA is used to build a collection of training data that may be segregated as much as feasible and compressed as closely as possible. It is used to analyse visual objects, find patterns, and express the object in order to emphasise distinctions. It characterises the image’s subject in terms of variance. Each primary component of the picture describes the largest degree of variation. Because patterns might be difficult to identify in high-dimensional data, PCA can find these patterns by lowering the number of dimensions without considerable information loss.

3.2 Linear discriminant analysis feature extraction

Fisher’s criterion function is used to choose a solution vector for linear discriminant analysis, which will ensure that after projection in this direction, the homogeneous samples in the original data set will be as close as possible to each other and as far apart as possible from each other. As the name implies, the LDA method is merely concerned with finding the optimal transformation vector. In this case, a scatter matrix is used to reduce the feature space before determining the ideal solution vector.

The ideal feature set was generated by utilizing linear discriminant analysis. It also normalized correlated features from the variance of the object. When two or more vectors are evaluated linearly, dimensions are reduced.

3.3 Feature level fusion

For recognition, feature fusion aims to combine information from two or more feature vectors into a single feature vector with higher discriminating power than any of its inputs. Using linear discriminant analysis models in low dimension space, concatenated features are constructed by combining perceptually salient picture characteristics. Correlation analysis has been demonstrated to be effective in identifying vector relationships in decreased feature sets [20]. A fused vector is generated by combining these two homogeneous vectors.

$X_{PCA} = [m_{{PCA}_{1}}, m_{{PCA}_{2}}, m_{{PCA}_{3}} . . . . m_{{PCA}_{n}}]$ (1)

Where X_PCA represents vein feature Vector extracted by the PCA

$X_{LDA} = [m_{{LDA}_{1}}, m_{{LDA}_{2}}, m_{{LDA}_{3}} . . . . m_{{LDA}_{n}}]$ (2)

Where X_LDA represents Vector feature of vein extracted by the LDA, n denotes the training samples.

The distance vectors of the PCA X_PCA extracted features and the LDA extracted features X_LDA were combined after computing the distance vectors of d^PDA and d^LDA of each vector. By maximizing pairwise correlation, these two vectors are joined using the combination rule. According to this study, the vector combination rule is the sum of all the feature vectors. Equation (4) displays the resulting vectors after adding the distance vector to the mean vector (equation 3).

$d = {\frac{d 1 PCA + d 1 LDA}{2} \dots \frac{dNPCA + dNLDA}{2}}$ (3)

${d = {d}_{1}^{PCA}, . . . . . . {, d}_{N}^{PCA} {, d}_{1}^{LDA}, . . . . . d_{N}^{LDA}}$ (4)

3.4 Proposed GBS intelligent water drops feature selection model

Feature selection maximizes classification accuracy and reduces computation costs by removing the least important features from a feature set. Figure 3 depicts the work flow of IWD employing a unique combinational model for feature selection using Global Best Selection-Intelligent Water Drops model to create the ideal features for vein detection.

Fig. 3

Flow chart for IWD algorithm with HUD search.

3.5 Overview of IWD

Shah-Hossein was the first to develop intelligent water droplets as an optimization method that was inspired by nature. Shah-Hossein overcame the Traveling Salesman Problem (TSP), the Knapsack Problem, and the n-Queen Problem with this approach. The optimization technique has recently been used in a variety of domains, including shop scheduling, task scheduling in a Grid environment, practical multi-echelon supply chain, and sentiment analysis feature selection [21].

Intelligent water drops construct the optimal best features by collaborating on various heuristics, particularly scatter search and iterated local search. Probabilistic Selection using a roulette wheel based on velocity and F score computation are two-step processes. A set of population vectors is chosen based on probability distribution in the roulette wheel selection process. To determine the trial vector, the f score is computed for each parameter vector and instance. Tuba Parlar and Esra Sarac explore IWD optimization for feature selection in sentiment analysis.

The basic nature of rivers is that they flow along their path, and the flow of water is referred to as water drops here. The characteristics considered for the optimization problem are the optimal path distance and flow speed. Soil and velocity are the two most important characteristics of flowing water. The flow of water carries small amount of soil and its velocity (speed) is based on the soil the water carries.

3.6 Mathematic modelling of IWD

Intelligent water drops optimization technique [23] is constructed on the natural flow of water. The flow of water drops owns some properties and the main two properties considered in this technique is Soil (Soil_IWD) and Velocity (Velocity_IWD). Here the problem is to find the optimal feature subset. The graph is constructed and the flow of water drops traverse to all the features (nodes) based on the shortest path (edge) and find out the best subset features with minimal time.

The probability function of water drop is shown in below equation

$P_{i} IWT (t) = \frac{f (soil (i, j))}{\sum_{t = v_{c}} {IDW}^{(soil (i, j))}}$ (5)

The ΔVelocity_IWD of the time depends upon the Soil (i, j), the flow of the water in the path between i and j and is derived from the equation (6)

$Δ {Velocity}_{(IWD (t))} = \frac{a_{v}}{b_{v} + c_{v} . {Soil}^{2} (i, j)}$ (6)

a_v, b_v, c_v, are positive static parameters. Soil (i, j), amount of the soil carried from i to j. It is the amount of soil that is added to the water drops that determines the flow of the water drops. As a result of the water drop removing soil from the local path, it is represented by

$Δ soil (i, j) = \frac{a_{s}}{b_{s} + c_{s} . {time}^{2} (i, j : {vel}^{IWD})}$ (7)

a_s, b_s, c_s, are positive static parameters. In equation (8), the duration of time taken by the IWD to move from location i to j is defined as the velocity of the IWD is proportional to the distance between the two locations.

${time}^{.} (i, j : {vel}^{IWD} (t + 1) = \frac{HUD (i, j)}{{vel}^{IWD} (t + 1)}$ (8)

HUD(I, j) is a heuristic function used to calculate the objectionable move of water drops. In addition to soil delivered by a water drop, soil on the path between nodes ‘i’ and ‘j’ can be updated using equations (9 and 10).

$soil (i, j) = (1 - ρ_{0}) . soil (i, j) - ρ_{n} . Δ soil (i, j)$ (9)

${soil}^{IWD} = {soil}^{IWD} + Δ soil (i, j)$ (10)

The best solution for each iteration, based on the collection of solutions obtained by each IWD, the FIA can be calculated as follows:

$F^{IA} = arg (\min) g (F^{IA}) \forall F^{IWD}$ (11)

The optimum solution for all IWDs in equation (11), F^IA, has the fewest features. Following receipt of this solution from F^IA, the soil of the path that best solves the problem of the existing way is updated as indicated below.

$soil (i, j) = (1 + ρ_{IWD}) . soil (i, j) - ρ_{IWD} \frac{1}{g (F^{IWD})}$ (12)

The procedure is repeated until the maximum number of iterations has been achieved. It is depicted below-

$F^{IA} = {\begin{matrix} F^{IA} {ifgF}^{IA} ⩾ g (F^{IA}) \\ F^{IA}, otherwise \end{matrix}$ (13)

3.7 Parameter setting

The selection of parameters may influence the quality of the computational results.

Here, the traditional Practical Swarm Optimization (PSO) approach is utilized to calculate the effectiveness of the proposed GBSIWD feature selection. PSO is a swarm intelligence based on global evolutionary search techniques.

The algorithms in the wrapper technique, which implies they all, require a learning algorithm to measure the classifier accuracy of the selected feature subset during the evolutionary training phase. Here KNN learning algorithm is used. We set the maximum number of runs to 30, and set the maximum number of iterations to 3000 to terminate the algorithm. The following table shows the initial parameter settings for the feature selection model utilising GBS PSO and GBS-IWD.

The training settings utilised for GBS-PSO feature selection are shown in Table 1. It creates the best feature from a given fusion vector using a fitness calculation based on an objective function.

Table 1
Parameter setting for feature selection using Global Best Selection PSO

Parameters Size

Population Size(NP) 50

Iterations 100

C₁ 0.7(0.1)

C₂ 0.5(0.1)

Inertia Weight 1

Parameters	Size
Population Size(NP)	50
Iterations	100
C₁	0.7(0.1)
C₂	0.5(0.1)
Inertia Weight	1

Table 2 displays the training settings utilised for GBS-IWD feature selection. GBS-IWD generates the feature vector on 30 runs, for each run 100 iteration. IWD based feature selection algorithm selects a predefined number of features for each water drop. They update soil and velocity values, as well as feature selection probabilities, based on the performance of the selected features. This process is repeated until only the best traits remain. After selecting the best features, the dataset is classified using the specified feature set.

Table 2

Parameter Setting for Feature Selection using Intelligent Water Drops

Parameters	Size
initSoil	1000
Initvel	100
a_v, b_v,Cv	1000, 0.01, 1
bs, cs	0.01,1
ρ	0.01
Number of iterations	100
Number of IWD runs	30

3.8 GBS_IWD Algorithm for feature Selection

GBSIWD is used to construct the graph for the finger vein features. The parameters have been set to their default values. Initially the water drops spread over randomly to construct the graph and then starting node is initiated with the probability of selecting node i as in (14)

$p (i) = \frac{f (soil (i))}{\sum_{new \notin_{L}} f (soil (new))}$ (14)

Here p(i) is the probability of the selection node i and f (soil (i)) is the fitness function. And is inversely proportional to the entire soil value as in equation (15)

$f (soil (i)) = \frac{1}{ɛ_{s} + g (soil (i))}$ (15)

Where

$\begin{matrix} g (soil (i, j)) \\ = {\begin{matrix} soil (i) \underset{new \notin_{L}}{ifmin} (soil (new)) ⩾ 0 \\ else (soil (i)) - \min_{new \notin_{L}} (soil (new)) \end{matrix} \end{matrix}$ (16)

Soil(i) is quantity of the soil in the node (feature) and soil(new) refers to the node unvisited and the new node is added to the list L. After selection the possible nodes using equation (14) and training dataset is classified according to the selected features.

As a fitness function, classification accuracy is employed to assess the quality of the features. The high accuracy features are chosen, and the velocity and soil values are updated based on the accuracy.

For each water drops the velocity is updated as in (17) follows

${vel}^{IWD} (t + 1) = \frac{{Velocity}_{(IWD (t)) +} a_{v}}{b_{v} + c_{v} . \sum_{i \in L} Soil (i)}$ (17)

The soil quantity of selected features are constructed using

$Δ soil (L) = \frac{a_{s}}{b_{s} + c_{s} . {time}^{2} (L, {vel}^{IWD} (t + 1))}$ (18)

Where

$\begin{matrix} {time}^{2} (L, {vel}^{IWD} (t + 1)) \\ = \frac{1}{Accuracy (L) \times {vel}^{IWD} (t + 1)} \end{matrix}$ (19)

Here L refers to the selected subset features and Accuracy (L) refers to the accuracy of the selected subset features. The soil content of the selected feature L is removed and updates using (20)

${soil}^{IWD} = {soil}^{IWD} + Δ soil (L)$ (20)

Until the termination, the above steps are computed. Here we use 100 iteration as termination for each run and totally 30 runs are experimented.

Initialization parameter s

Representing the features in the form of graph

Initializing values for static parameters

Initialization of population NP

Max iteration

Soil updating parameters a _s, c _s , b _s

Velocity updating parameters a _v, c _v , b _v

Local soil updating parameter ρ = 0.09

Global Soil updating parameter ρ_IWD = 0.09

Initial soil edge of the graph as Initsoil

Initial velocity of the graph as InitVel

Initialization of Dynamic parameters

Visited Node V_c(IWD)={}, empty

InitVel

Soil={}, empty

Pesudocode of proposed GBSIWD Algorithm

While (termination criteria not reached)

Do{

Initial placement of IWD positions on random features

Global Best Feature Selection by IWDs

Select an edge between two features considered as nodes

Update local soil content

Obtain the iteration-best feature Subset

Update Global soil on edges for adjacent features in the iteation-bset feature subset

Update the global-best feature subset

} End while

Return the global-best feature subset

3.9 KNN classification

The KNN classifier was used to classify the optimum feature subset generated through feature selection. It is a non-parametric approach. The approach is termed instance-based since it finds the k nearest points to forecast the class label for the feature sample. Using the distance weighting estimate, the closest k points are predicted as classes for training feature samples.

Assume that X represents an ideal feature vector for which the nearest neighbour must choose the class that it represents; the distance measure can be used to determine which class T_ij. Distance (d) is calculated using Euclidean distance.

$d (x, y) = \sqrt{\sum_{i = 1}^{n} {(x_{i} - y_{i})}^{2}}$ (21)

Where n is a number of dimensions, x and y are data points.

4 Results and discussion

Finger vein identification was tested using a novel combinational model in this study. Based on the results of the experiment and the K-fold cross validation test described in the sections below, its performance has been evaluated in various settings.

4.1 Experimental environment and dataset description

The experiment is run on a Windows 7 Intel Core i7 CPU with an operating frequency of 2.53 GHz and an useable memory of 8GB. The THU-FVFDT1 finger vein dataset utilised in this study was collected from 40 people, with 5 fingers veins from each person, yielding 200 samples. The photos were taken in three sessions, each with four photographs, with at least one week between each session. The database image has a resolution of 720x576.

4.2 Evaluation criteria

The different classifiers are introduced and evaluated based on the evaluation criteria to show which one performs better in the proposed work. Performance measures like accuracy, FAR, FAR, and EER are used to calculate the FKP recognition’s results. Equations (22), (23), (24), and (25) include general formulas for calculating precision, FAR, and FRR.

$\begin{matrix} FalseRejectionRate (FRR) \\ = \frac{FalsePositve}{TruePositive + FalseNegative} \times 100 % \end{matrix}$ (22)

$\begin{matrix} FalseAcceptanceRate (FAR) \\ = \frac{FalsePositve}{FalsePositve + TrueNegative} \times 100 % \end{matrix}$ (23)

$EqualErrorRate (EER) = 1 - FRR$ (24)

$Accuracy = \frac{TruePositve + TrueNegative}{TruePositve + TrueNegative + FalsePostive + FalseNegative} \times 100 %$ (25)

4.3 Pre-processing

The preprocessing is done to enhance the quality of the finger vein image. Typically, the finger vein image has low contrast, noise, and shades. This is due to changes in the light, the finger’s rotation and translation, and the functionality of the capturing device. To alleviate these issues, the preprocessing step is used. In this work, the ROI (region of interest) is extracted and histogram equalization is used in image enhancement. The preprocessing finger vein images are given in Fig. 4a) input finger vein image, b) ROI of finger vein and c) enhanced FV image using Histogram equalization.

Fig. 4

a) Input Finger vein b) ROI FV image c) After Histogram Equalization.

4.4 Feature extraction: PCA_LDA

The pre-processed finger vein images are used to extract the more relevant features using hybrid feature PCA_LDA feature extraction technique. For each image, the decreased dimension of fusion vector length is 72. Photographs of three images from each of the 40 patients are utilised for training in the first fold validation. The vein patterns were extracted as single dimension feature vectors using principle Component Analysis and Linear Discriminant Analysis during the training of the finger vein detection model. In addition, Feature fusion is employed using Concatenation process of hybrid PCA_LDA feature extraction.

4.5 Feature selection with GBSPSO and GBSIWD optimization technique

Comparisons of the proposed techniques’ recognition performance with existing PCA LDA Euclidean and PCA LDA KNN methods are shown in Tables 3 and 4. The cutting-edge approaches are contrasted by categorising them as non-training-based or training-based.

Table 3
Performance evaluation of existing models using PCA-LDA euclidean

Feature extraction Performance metrics (%)

FAR FRR ERR Accuracy

PCA 9.6 7.2 8.4 91.6

LDA 6.7 4.5 5.6 94.4

PCA_LDA 5.8 3.6 4.7 95.3

PCA_LDA Features with GBSPSO 3.1 1.8 2.45 97.55

PCA_LDA Features with Proposed GBSIWD 1.2 2.4 1.8 98.2

Feature extraction	Performance metrics (%)
PCA	9.6	7.2	8.4	91.6
LDA	6.7	4.5	5.6	94.4
PCA_LDA	5.8	3.6	4.7	95.3
PCA_LDA Features with GBSPSO	3.1	1.8	2.45	97.55
PCA_LDA Features with Proposed GBSIWD	1.2	2.4	1.8	98.2

Table 4

Performance evaluation of existing models using PCA-LDA-KNN

Feature extraction	Performance metrics (%)
	FAR	FRR	ERR	Accuracy
PCA	9.6	8.1	8.85	91.15
LDA	7.8	6.2	7	93
PCA_LDA	5.2	2.1	3.65	96.35
PCA_LDA Features with GBSPSO	1.4	0.9	1.15	98.85
PCA_LDA Features with Proposed GBSIWD	0.8	0.5	0.65	99.35

In this part, we describe experiments that demonstrate the efficacy of the proposed strategy in merging feature sets generated from the hybrid PCA LDA model on finger vein pictures. Because neighbouring pixels in a picture are frequently associated [24], down sampling the feature images can decrease information redundancy.

A hybrid PCA LDA feature vector is formed by combining the N-dimensional feature vectors generated at each point. Moreover, the Fused Feature vector increases recognition, which is then subjected to feature selection using the suggested IWD. For the comparison of the proposed GBSIWD feature selection the traditional PSO optimization technique is implemented in the same procedure.

4.6 Comparative results of GBSIWD with GBSPSO based on no of generations/users

In a research of finger vein recognition, the optimization strategies for feature selection are applied. Therefore, the comparison the Practical Swarm Optimization is implemented in the same procedure and the results compared with the proposed GBSIWD. Both IWD and PSO, which are optimization algorithms, produce better results for the feature summary. When comparing PSO and IWD in terms of classification accuracy, IWD outperforms PSO with fewer features and faster computation. As a result, the error case for each method appears differently. As a result, the error rate of each approach differs in the first and second folds. Because of the misalignment, the erroneous rejection case varies greatly [25].

In this experiment, we use hybrid PCA_LDA features to select the optimal subset features based on proposed IWD-based feature selection techniques. The features of the Global Best Solution are chosen based on the relevant features. Table 5 depict the number of features selected form various feature vectors reliable to counts from 1 to 5 using GBSIWD and GBSPSO-based feature selection algorithms.

Table 5
Comparative analysis of GBSIWD and GBS PSO features for number of runs

No of runs Count 1 Count 2 Count 3 Count 4 Count 5

GPSIWD GPSPSO GPSIWD GPSPSO GPSIWD GPSPSO GPSIWD GPSPSO GPSIWD GPSPSO

10 55 57 28 26 15 11 4 25 1 2

15 61 66 47 44 26 21 9 12 2 2

20 69 72 59 60 40 36 19 26 8 11

25 73 72 62 65 53 53 34 36 18 21

30 73 73 69 67 60 60 46 50 26 30

No of runs	Count 1	Count 2	Count 3	Count 4	Count 5
10	55	57	28	26	15	11	4	25	1	2
15	61	66	47	44	26	21	9	12	2	2
20	69	72	59	60	40	36	19	26	8	11
25	73	72	62	65	53	53	34	36	18	21
30	73	73	69	67	60	60	46	50	26	30

For feature selection, the hybrid PCA LDA features are extracted and fed into the IWD and PSO algorithms. For runs 10, 15, 20, 25, and 30, the global best solution of features is selected by eliminating repeated feature values and here the increasing count in optimization algorithms such as IWD and PSO results in subset of optimal features. In count 5, for 30 runs it shows 26 selected features for GPSIWD and 30 features for GPSPSO respectively. Finally, the feature count 5 reveals that, as compared to PSO, the proposed GBSIWD performs well with a reduced number of features.

According to Table 6, when IWD and PSO-based feature selection are used, fewer features are selected, the subset of features is optimal, and recognition accuracy is improved. As the number of generations increases, both feature selection algorithms improve their classification accuracy. In every evaluated feature, the accuracy of GBSIWD and GBSPSO is comparable, but with a reduced number of features. GBSIWD and GBSPSO feature selection algorithms have accuracy of 99.4 percent and 99.6 percent, respectively. Besides that, IWD has 26 selected features, while PSO has 30 for 30 runs.

Table 6

Performance accuracy of the proposed GBSIWD with GBSPSO

No of	Reduced No of features (A-B)		Optimized selected features (B)		Accuracy (%)
Generations	PSO-GBS	IWD-GBS	PSO-GBS	IWD-GBS	PSO-GBS	IWD-GBS
10	78	79	1	0	62.16	56.65
15	78	78	1	1	70.27	68.2
20	69	72	10	7	94.2	95.6
25	59	62	20	17	96.2	97.8
30	49	53	30	26	98.8	99.3

Table 7 shows the number of features reduced and selection of relevant features based on users. Each class contain 5 finger vein images 70% for training and 30% for testing. 10, 15, 20, 25, 30, 35, classes contain 50, 75, 100, 125, 150, 175, 200 users respectively. The hybrid PCA_LDA feature vector is performed with count 5 for different class levels. The GBSIWD selected optimal feature vectors are 19, 28, 32, 29, 25, 22, 20 for 10, 15, 20, 25, 30, 35, classes, which is comparatively less than GBSPSO. This shows that the number of features selected for the GBSIWD feature selection algorithm is less than the GBSPSO. The GBSIWD-based feature selection algorithm, takes more time to execute than the GBSPSO feature selection algorithm. However, GBSIWD is good in finding the optimal number of features, which compensate for computational inefficiency.

Table 7

Analysis of Recognition time of the GBSIWD and GBSPSO

No of classes	Total Eifi	Reduced No of features (A-B)		Optimized selected features (B)		Time (s)
	features	GBSPSO	GBSIWD	GBSPSO	GBSIWD	GBSPSO	GBSIWD
10	20	1	1	19	19	180	210
15	30	1	2	29	28	270	300
20	40	7	8	33	32	330	390
25	50	11	21	39	29	390	480
30	60	26	35	34	25	450	600
35	70	41	48	29	22	570	720
40	80	49	60	31	20	630	720

Table 9

Performance comparison of existing Finger vein verification techniques

Authour	Techniques	EER
Xie, C. and Kumar, A [16]	CNN with triplet Similarity loss	13.16%
Supervised Discrete hashing with CNN	9.77%
Kang, W., et al. [17]	3DFM with non-multipose data	1.35%
3DFM with multipose data	2.13%
Huang, J., et al. [18]	JAFVNet for SDUMLA-HMT	0.35%
JAFVNet for FV-USM	0.49
Proposed method	GBSPSO_KNN	0.31%
GBSIWD_KNN	0.15%

However, the feature selection of the PSO has been carried out on the same iteration of the IWD for regeneration of optimal features. Optimal feature generates on the parameter setting which is in table. With KNN, a feature identified with GBS-IWD is used to recognize the finger vein in the query image. As described above, KNN uses a number of neighbours (K = 3) to measures the performance accuracy of the described methods: it increases the probability that neighbours will be close so that noise can be reduced (increasing the chance of a close neighbor).

Compared to existing finger vein verification techniques, the proposed technique performs better shown in Fig. 9. From the table the equal error rate (EER) of the proposed technique is 0.15% and 0.31% is less than the existing JAFVNet, 3DFM and CNN respectively.

4.7 Discussion and comparison with other cutting-edge methods

The proposed feature selection technique for finger vein recognition achieves good accuracy with 99.35% for the IWD and 98.85% for the PSO, respectively. Furthermore, the processing time for the proposed approach takes 0.2s, whereas existing methods GWO and FFF takes 1.23s and 5.1s, respectively [26]. As a result, the suggested feature selection algorithm consumes less time and reduces computing costs. Further, the proposed feature selection algorithm works well with the hybrid PCA LDA features extraction technique for finger vein recognition. Therefore, in comparison to other existing approaches shown in Table 8, the proposed approach improves accuracy and reduces the computational cost.

Table 8
Performance Comparison with existing methods

Feature Selection Techniques Accuracy Sensitivity Specificity Recognition rate EER Time (s)

Grey Wolf Optimization 98 97.5 94.6 92 0.18 1.23

Feature-level fusion using fractional firefly (FFF) 95 98 99.2 93.5 0.35 5.1

Proposed GBSPSO_KNN 98.85 97.6 98.2 97 0.31 0.21/image

Proposed GBSIWD_KNN 99.35 98.4 96.8 98.3 0.15 0.24/image

Feature Selection Techniques	Accuracy	Sensitivity	Specificity	Recognition rate	EER	Time (s)
Grey Wolf Optimization	98	97.5	94.6	92	0.18	1.23
Feature-level fusion using fractional firefly (FFF)	95	98	99.2	93.5	0.35	5.1
Proposed GBSPSO_KNN	98.85	97.6	98.2	97	0.31	0.21/image
Proposed GBSIWD_KNN	99.35	98.4	96.8	98.3	0.15	0.24/image

5 Conclusion

For finger vein recognition, we developed and implemented a new combinational model. The use of feature extraction techniques such as PCA & LDA, which yielded high relevant features, resulted in an improved finger vein detection system in this study. Furthermore, those features are combined to form feature fusion. Further, population-based optimization techniques such as Practical Swarm Optimization based global Best Selection (PSOGBS) and Intelligent Water Drops (IWDGBS) have been developed to provide a technique of categorization of features besides that, IWD has 26 selected features, while PSO has 30 for 30 runs. The IWDGBS selected optimal feature vectors are 19, 28, 32, 29, 25, 22, 20 for 10, 15, 20, 25, 30, 35, classes, which is comparatively less than PSO-GBS. The proposed feature selection technique for finger vein recognition achieves good accuracy with 99.35 % for IWD and 98.85 % for PSO respectively. The processing time for the proposed takes 0.2 s whereas existing methods GWO and FFF take 1.23s and 5.1s. Therefore, the proposed feature selection algorithm takes only lesser time consuming and reduces the computational cost. For finger vein recognition, hybrid PCA_LDA features extraction technique, which supports well for the proposed feature selection algorithm. In comparison to existing approaches, the proposed approach improves accuracy and reduces the computational cost. The proposed technique will be tested on large number of datasets in the future. Also, the proposed techniques intend to conduct experiments by combining more robust meta-heuristic optimization tools for further inquiry in this domain.

References

Liu

, Yin

, Yang

, Dong

, Xi

Finger vein recognition with superpixel-based features, In IEEE International Joint Conference on Biometrics, Clearwater, FL, USA, 2014, pp. 1–8.

Yang

, Shi

and Jia

, Finger-vein image matching based on adaptive curve transformation, Pattern Recognition 66 (2017), 34–43.

Noh

K.J.

, Choi

, Hong

J.S.

and Park

K.R.

, Finger-vein recognition based on densely connected convolutional network using score-level fusion with shape and texture images, IEEE Access 8 (2020), 96748–96766.

SyedMeeran

M.M.

, Kathiravan

M.G.

, SenthurBalaji

, Saravanan

M.T.

Finger Vein Recognition System-A Review.

Wimmer

, Prommegger

, Uhl

Finger vein recognition and intra-subject similarity evaluation of finger veins using the cnn triplet loss, In 2020 25th International Conference on Pattern Recognition (ICPR) (pp. 400–406). IEEE, 2021, January.

Liu

, Yang

and Yin

, Discriminative binary descriptor for finger vein recognition, IEEE Access 6 (2017), 5795–5804.

Fang

Z.M.

and Lu

Z.M.

, Deep belief network based finger vein recognition using histograms of uniform local binary patterns of curvature gray images, International Journal of Innovative Computing, Information and Control 15(5) (2019), 1701–1715.

Shaheed

, Mao

, Qureshi

, Kumar

, Hussain

and Zhang

, Recent advancements in finger vein recognition technology: methodology, challenges and opportunities, Information Fusion 79 (2022), 84–109.

Liu

, Yang

and Yin

, Discriminative binary descriptor for finger vein recognition, IEEE Access 6 (2017), 5795–5804.

10.

Liu

, Kim

Y.H.

An efficient finger-vein extraction algorithm based on random forest regression with efficient ocal binary patterns, In 2016 IEEE International Conference on Image Processing (ICIP) (pp. 3141–3145), IEEE, 2016, September

11.

Yang

, Shi

and Jia

, Finger-vein image matching based on adaptive curve transformation, Pattern Recognition 66 (2017), 34–43.

12.

Darwish

S.M.

, Feature extraction of finger-vein patterns based on boosting evolutionary algorithm and its application for loT identity and access management, Multimedia Tools and Applications 80(10) (2021), 14829–14851.

13.

Saadat

, Nasri

A GSA-based method in human identification using finger vein patterns, In 2016 1st Confer ence on swarm Intelligence and Evolutionary Computation (CSIEC), (pp. 102–106), IEEE, 2016, March.

14.

Xiong

, Zhang

, Xu

and He

, A modified chaotic binary particle swarm optimization scheme and its application in face-iris multimodal biometric identification, Electronics 10(2) (2021), 217.

15.

Garg

, Arora

and Gupta

, An efficient human identification through iris recognition system, Journal of Signal Processing Systems 93(6) (2021), 701–708.

16.

Xie

and Kumar

, Finger vein identification using convolutional neural network and supervised discrete hashing, Pattern Recognition Letters 119 (2019), 148–156.

17.

Kang

, Liu

, Luo

and Deng

, Study of a full-view 3D finger vein verification technique, IEEE Transactions on Information Forensics and Security 15 (2019), 1175–1189.

18.

Huang

, Tu

, Yang

and Kang

, Joint attention network for finger vein authentication, IEEE Transactions on Instrumentation and Measurement 70 (2021), 1–11.

19.

Kapoor

, Rani

, Kumar

, Chopra

and Brar

G.S.

, Hybrid local phase quantization and grey wolf optimization based SVM for finger vein recognition, Multimedia Tools and Applications 80(10) (2021), 15233–15271.

20.

Mokhtari

, A nature inspired intelligent water drops evolutionary algorithm for parallel processor scheduling with rejection, Applied Soft Computing 26 (2015), 166–179.

21.

Ahmadi

, Mehdizadeh

, Mohammadi

, Pham

Q.B.

, Doan

T.N.C.

and Vo

N.D.

, Application of an artificial intelligence technique enhanced with intelligent water drops for monthly reference evapotranspiration estimation, Agricultural Water Management 244 (2021), 106622.

22.

Kayvanfar

, Husseini

S.M.

, Karimi

and Sajadieh

M.S.

, Bi-objective intelligent water drops algorithm to a practical multi-echelon supply chain optimization problem, Journal of Manufacturing Systems 44 (2017), 93–114.

23.

Kalita

D.J.

, Singh

V.P.

and Kumar

, Two-way threshold-based intelligent water drops feature selection algorithm for accurate detection of breast cancer, Soft Computing 26(5) (2022), 2277–2305.

24.

Farmanbar

and Toygar

Ö.

, Feature selection for the fusion of face and palmprint biometrics, Signal, Image and Video Processing 10(5) (2016), 951–958.

25.

Kapoor

, Rani

, Kumar

, Chopra

and Brar

G.S.

, Hybrid local phase quantization and grey wolf optimization based SVM for finger vein recognition, Multimedia Tools and Applications 80(10) (2021), 15233–15271.

26.

Veluchamy

and Karlmarx

L.R.

, System for multimodal biometric recognition based on finger knuckle and finger vein using feature-level fusion and k-support vector machine classifier, IET Biometrics 6(3) (2017), 232–242.

An improved intelligent water drops feature selection for finger vein recognition

Abstract

Keywords

1 Introduction

2 Related works

3.2 Linear discriminant analysis feature extraction

3.3 Feature level fusion

3.6 Mathematic modelling of IWD

Table 1 Parameter setting for feature selection using Global Best Selection PSO Parameters Size Population Size(NP) 50 Iterations 100 C1 0.7(0.1) C2 0.5(0.1) Inertia Weight 1

4.1 Experimental environment and dataset description

4.2 Evaluation criteria

4.5 Feature selection with GBSPSO and GBSIWD optimization technique

References

Table 1
Parameter setting for feature selection using Global Best Selection PSO

Parameters Size

Population Size(NP) 50

Iterations 100

C₁ 0.7(0.1)

C₂ 0.5(0.1)

Inertia Weight 1