Optimization driven deep belief network using chronological monarch butterfly optimization for iris recognition at-a-distance

Abstract

Identification of eye considering biometric traits is an essential field to recognize persons. Biometrics using iris images seems to be an effective identification of individuals. Various Iris Recognition at-Distance (IAAD) systems are used for extracting features of iris and improve image quality using the biometric model. Even though the quality of the iris is better, accuracy is a challenging question for the research community. Thus, an effective IAAD, namely Chronological Monarch Butterfly Optimization-Deep Belief Network (Chronological MBO-DBN) is devised to detect iris. The detection of iris using DBN is trained with Chronological MBO, which is the integration of Chronological theory and Monarch Butterfly Optimization (MBO). The features of iris are extracted with ScatT-Loop descriptor and Local Gradient Pattern (LGP) and subjected to Chronological MBO-DBN for the recognition of iris which improved accuracy. The implementation of proposed Chronological MBO-based DBN is performed using the dataset, CASIA Iris, and efficiency is evaluated by the accuracy of 96.078%, False Rejection Rate (FRR) of 0.4745% False Acceptance Rate (FAR) of 0.4847%, and F-Measure of 98.658%.

Keywords

Deep belief network ScatT-loop local gradient pattern iris recognition hough transform

1. Introduction

The outstanding act of the biometric system is obtained by focussing on the features of the image, which is captured from the sensors [1]. Biometric technology is a new recognition system for identifying the individuals both behaviourally, like palm prints, signature, fingerprints, face, voice, and veins [2, 30]. A Biometric model is an emerging and continually developing expertise utilized in several automatic models for detecting the identity of a person powerfully and distinctively without the requirement of remembering or carrying something, like user ID and password [3]. Also, it is used in the defect prediction systems [33, 34]. Several studies proved that iris trait contains several merits than other biometric is devised using features like face and fingerprint, this makes the iris system to be commonly accepted in many applications [3]. The biometric is broadly classified into two types, namely multimodal and unimodal biometrics. The unimodal establishes the person identity using a single source of information, like the left part of the iris, right part of the iris, and face [3]. In multimodal, whenever the model function on type of identification, the classifier output is inspected as ranks generated by candidates that symbolize feasible matches [3]. Development and execution of multimodal needs several factors that influence overall system performance [3].

Among various biometric modalities, the recognition of iris is termed as the most reliable biometric modalities considering minimal rates of error [3]. For identifying persons, acknowledgment of iris is an important research field termed as iris recognition [4]. The determination of iris is an automated authentication approach used for recognizing the pattern and arithmetical features [1]. Recognition of iris is considered as an imperative task in determining the person using the texture of iris [5]. Recognition of iris is emerged as the most promising technology to offer reliable human identification [6]. Recognition of iris is an automatic process to identify the individuals using the patterns of iris [7]. The system for recognizing iris is flexible, precise, and reliable to identify the person considering other biometric models. Hence, the detection of the iris is extensively used in many national ID programs, like UAEs border security and India’s Aadhaar to process the biometric data for millions of peoples [5]. The recognition of iris requires an intellectual model with elevated consistency and fewer coefficient errors [2]. Mostly the system based on Iris utilizes three components for processing, which includes acquisition, pre-processing, and extraction of features [8]. Iris utilizes Gabor filters for extracting imperative proofs of the image as a feature vector [9].

Most of the methods are very useful, but iris is the best recognition system [2]. Iris biometric system provides a reliable method for identifying individuals in the most significant functions [10]. The various methods, such as the triple-A method, fuzzy logic assessment, active contour method, deep learning approach, and other classical recognition strategies are used for the iris recognition. The Triple-A method is adopted for the alignment process to establish alignment between the iris codes in an effective manner [8]. The fuzzy logic edge estimation method is a discrete approach for localizing the iris efficiently for improved acknowledgment [4, 31, 32]. The active contour method consists of a snifter of joysticks to shift across images to obtain symmetry for identifying the edge of eyes [11]. The correctness of remarkable iris recognition is reported using the conventional recognition approach utilized in eyes images using NIR collected through controlled milieu [6]. The optimization algorithms [35, 36] are used to train the deep learning approach, which offers a marvellous achievement in the CAD design area and achieves improved performance over classical image classification methodologies [12, 29]. The recognition of iris is used in several applications, which includes nationwide ID cards, dataset access, and economic services [3].

The goal of research work is to devise a technique considering the machine learning technique. At first, the image of the iris is subjected to the pre-processing step for the mining iris region by adapting the HT. After extracting the region, the segmentation and normalization are performed considering Daugman’s rubber sheet. Then, features are successfully mined considering the combination of ScatT-Loop descriptor and LGP, where ScatT-Loop is devised considering TT, Loop descriptor, and ST. After the extraction of features, the detection is carried out considering DBN using the proposed Chronological MBO.

Contribution of the work

Chronological MBO-DBN: Here, optimal weights of DBN are determined using a chronological-MBO algorithm to perform effective iris recognition.

Chronological MBO: It is developed by integrating the MBO algorithm with the chronological concept to train the DBN.

Other sections of the paper are orchestrated as: literature survey of existing methods is discussed in Section 2. Section 3 demonstrates the proposed recognition strategy, and the results along with the examination are discussed in Section 4, and the conclusion is presented in Section 5.

2. Motivation

Here, the motivation of the proposed approach is elaborated, which involves various existing iris recognition methods.

2.1 Literature survey

Some existing methods of iris recognition considering the issues and benefits are surveyed and illustrated below.

Ramaiah and Kumar [5] devised the NBNN model for evaluating the patterns of iris using the iris images. The method utilized a bi-spectral recognition model for acquiring visible and infra-red images. This model efficiently matched iris using a different set of domains. However, an efficient method was needed for recovering discriminant features. Tan and Kumar [13] devised the iris encoding method for providing recognition ability for iris images. In the encoding method, the information related to textures was used considering pixels of iris regions. The matching of iris offered precise matching ability and was valuable for making better decisions. The feature-based on encoded iris was modelled in binary that permitted matching of the template with hamming distance. Othman and Dorizzi [1] devised a quality fusion mechanism for computing quality using iris image. This measure was based on the Gaussian model for evaluating texture distribution using iris images. The method effectively eliminated inadequately segmented pixels but failed to test efficiency with other fusion mechanism using iris codes. Tan and Kumar [6] devised the Zernike moments based encoding approach for extracting and integrating the global and local iris features. The method concurrently performs local constancy in iris bit. The phase features are generated using regions and hold region variation using iris images but, the matching of the iris was not effectively done. Nguyen et al. [7] devised CNNs for expressing features of images. The textural nuances using the patterns of iris were mined and encoded successfully adapting Gabor wavelets and alter the response of phasor with binary code. Moreover, the iris templates capacity was modelled by integrating CNN with other methods. Ahmadi and Akbarizadeh [2] devised a human iris recognition approach by integrating MLP NN with the PSO algorithm for improving performance. The method extracted features with Gabor using iris images. For maximizing the efficiency, the PSO is integrated using the fuzzy model. Liu et al. [10] devised the code-level approach for recognizing iris using iris images. The method utilized binary feature codes for transforming the iris templates. However, the method failed to use other biometric modalities for improving the performance. Al Waisy et al. [3] devised deep learning strategy known as IrisConvNet which was the incorporation of Softmax classifier on CNN for retrieving discriminative features from the iris images. This method utilized training methods for controlling overfitting issues and improves the generalization ability of NN.

Figure 1.

Architecture of proposed deep learning approach using optimization.

2.2 Challenges

The issues confronted in the classical iris recognition strategies are enlisted below.

•
The main issue faced by classical iris recognition strategies is to extract features of iris using iris images as features obtained through the constrained environment are contaminated by the noisy environment [14].
•
The global and local quality measures employed for estimating the texture distribution of iris by choosing optimal images using fusion is considered to be a major issue [13].
•
Sensors that capture iris patterns endure several noteworthy transformations. These transformations contain many issues in the recognition model. When the model utilizes many users, then enrolment becomes lengthy and costly in the iris recognition models. Hence, this model was infeasible for re-enrolling the user when a new user arrives [15].
•
Acquisition of elevated accuracy for recognition of patterns using iris is a main issue and is complex for determining evident feature from iris images and is complex to set its capacity [16].
•
The recognition of iris needed controlled environments and iris images and poses many issues mainly when images are captured with visible imaging and vibrant environments and may degrade the system performance [6].

3. Proposed chronological-MBO based DBN for iris recognition

Recognition of iris is imperative in the research field in identifying a person using iris texture features. Nowadays, recognition of iris is extensively used in several nationwide ID programs, like UAEs border security and India’s Aadhaar to process the biometric data for millions of peoples. However, the recognition of iris requires an intellectual model using high consistency and fewer coefficient errors. Thus, an optimized deep learning method is proposed for recognizing iris. Initially, input iris image undergoes pre-processing to acquire an exact region of iris and after pre-processing, the iris extraction is carried out using the Hough transform. Once the region is extracted from iris images, the segmentation and normalization are performed with Daugman’s rubber sheet model. The texture features are mined with ScatT-Loop and LGP, which are processed using the Chronological-MBO-DBN classifier for recognizing iris. Assume $M$ as database with $n$ number of images, and let $X_{i}$ indicate $i^{\rm th}$ iris image passed into the pre-processing stage in the iris recognition. Figure 1 depicts the structural design of the proposed method for the recognition of iris.

3.1 Pre-processing

Let us consider the input image as $X_{i}$ , which is fed to the pre-processing stage to improve the contrasts of the image and to model it appropriately for processing using IAAD. In this stage, the accuracy and the robustness of the input image are enhanced by eliminating the uninterested regions, and the pre-processed image is denoted as $X$ .

3.2 Hough transform for extraction of iris region

It is imperative to extract region of iris or region of interest using a pre-processed image $X$ for performing extraction of feature efficiently. HT is applied to the pre-processed image to attain the iris region using the pre-processed image $X$ . HT [17, 18] enables vigorous recognition using noisy and biased occlusion by determining circles, lines or other parametric curves that exist in the image. HT detects the boundary of the outer pupil and inner pupil and the parameters are used in the below equation,

$\displaystyle(m-m_{0})^{2}+(p-p_{0})^{2}=a^{2}$ (1)

where, $(m_{0},p_{0},a)$ signifies circle with radius $a$ and center coordinates $(m_{0},p_{0})$ . The results contain the papillary boundary, even if the prior center is not positioned at the center of the pupil and the circular HT computes the exact pupil center.

3.3 Normalization and segmentation of iris using Daugman’s rubber sheet model

The extracted region of the iris is processed to execute normalization and segmentation with the Daugman’s rubber sheet model [19], which is used to improve the performance of the system for recognizing iris. This model considers the size and dilation of the pupil, which do not utilize the interference of the user regarding earlier information of the user. Segmentation is carried out for retrieving engrossed region and normalization is performed for eliminating noise from iris to enhance recognition efficiency.

3.3.1 Segmentation

After extracting the iris image, the extremely significant region is segmented using segmentation for further processing the iris. The segmentation is carried out using the localization of the pupil and its eyelids, eyelashes, and boundaries. If it fails to segment the iris, there is a possibility to generate false data, which will affect the recognition rates. Thus, segmentation is essential using HT through effective iris localization.

3.3.2 Iris normalization

After segmenting regions of iris, the segmented iris is normalized with block considering equivalent size respect for block width $x$ and angular displacement $\theta$ and mapping pursue normalization. The Daugman’s rubber sheet indicates a model which are allocated with individual pixel using size, dilation, and real coordinate $(x,\theta)$ , in which $x$ represent unit interval and $\theta$ relies 0 to $2\pi$ . The images of iris are remapped from Cartesian coordinated $(r,s)$ to the polar coordinate system $(x,\theta)$ in such a way that $(r,s)$ is the normal coordinates and $(x,\theta)$ is the normalized polar coordinates. Assume iris boundaries and the coordinates of the pupil are indicated as, $(r_{e},s_{e})$ and $(r_{b},s_{b})$ in $\theta$ direction.

3.4 Extraction of noteworthy features with ScatT-LOOP

The regularized iris is fed to the extraction of features phase considering LGP and ScatT-LOOP descriptor. The ScatT-Loop descriptor is combination of the Loop descriptor [20], ST [21] and TT [22]. The ScatT-LOOP produces texture features for precise recognition of iris to exclusively detect individuals. Let $A$ indicate normalized iris image generated from Daugman’s rubber sheet model wherein features are obtained considering the ScatT-Loop descriptor. Initially, image $A$ is fed to LOOP descriptor-1 $A_{1}$ . Simultaneously, the iris image $A$ is fed to TT for generating image $A_{2}$ , which is distributed using LOOP descriptor-2 for producing image $A_{3}$ . The image $A_{1}$ is EXOR using image $A_{3}$ and obtained EXOR image is subjected to LOOP descriptor-3 for evaluating image $A_{6}$ . The output image $A_{2}$ generated from TT is subjected to ST for obtaining image $A_{4}$ , which is subjected to LOOP descriptor-4 for generating $A_{5}$ . At last, image $A_{5}$ is EXOR-ed using image $A_{6}$ for obtaining image $A_{7}$ , which is fed to histogram depiction. The image $A_{7}$ is input image with iris recognition phase, which is produced using optimized DBN.

3.4.1 LOOP descriptor

The LOOP descriptor [20] gains the advantages from LDP and LBP, which enhances the dependency on orientation and the drawback associated with the empirical assignment values are resolved. The intensity of the image is represented as, $G$ with, $G^{*}$ centered at $(r^{*},s^{*})$ and the neighbourhood pixel intensity is denoted as, $G^{k}$ and $k$ ranges from 0 to7. The Kirsch mask application yields the output in a particular direction that enables the probability sign in the presence of an edge in that direction. The output of eight Kirsch masks is denoted as $f_{k}$ and the individual pixels are assigned with an exponential weight $\upsilon$ based on the magnitude rank $f_{k}$ for the eight kirsch masks. The LOOP value of the equivalent pixel $(r^{*},s^{*})$ is formulated as,

$\displaystyle\textit{LOOP}(r^{*},s^{*})=\sum_{k=0}^{7}h(G^{k}-G^{*}).2^{\upsilon}$ (2)

where, $h(d)=\begin{cases}1;&\text{if}∼{}d\geqslant 0\\ 0;&\text{Otherwise}\end{cases}$ .

3.4.2 Tetrolet transform

The adaptive Haar wavelet transform is referred to as the tetrolet descriptor [22], to support the tetrominoes that are formed by joining the four squares using the same size. The input low pass image is partitioned as blocks and the local Tetrolet basis is carried out using the geometry of the image. The steps involved in the TT are explained below.

a) Formation of image blocks.

The input image undergoes segmentation and is partitioned as blocks of size 4 $\times$ 4.

b) Demonstration of blocks of image using sparsest tetrolet.

Each block of the image is fed to sparsest tetrolet depiction and for each block, 117 tetromino coverings were confessed, wherein each individual is subjected to Haar wavelet transform to generate 12 coefficients with the four low pass coefficients. The sparse image is obtained by performing the tetrolet decomposition for each block.

c) Illustration of low pass and high pass coefficients.

The steps contained in Tetrolet decomposition strategy utilize 2 $\times$ 2 matrix using the reshape function.

d) Tetrolet coefficients.

Once the sparse matrix is represented for each block, high pass and low pass matrix are kept secure for future use.

e) End.

The procedure is repetitive for low pass image and result generated from the binary image. The output generated from the Tetrolet transform is represented as, $A_{2}$ .

3.4.3 Scattering transform

ST [21] computes the texture features for the image $A_{2}$ to render the highly robust features and locally invariant features that are generated using the Morlet wavelet application using orientation and multiple scales. The steps contained in ST are deliberated as.

a) Local affine transformation to protect image.

The transformation is performed using the convolution of individual image segments with the low pass filter, which operates based on the scaling factor to prevent the deformation in the input image. By the local affine transformation, the high-frequency components are eliminated.

b) Capture the components with high-frequency using Morlet filters.

The high-frequency components is generated using the coefficients of wavelet, which becomes the result of the Morlet or bandpass and average filters.

c) Acquisition of scattering coefficients.

The transformation of wavelet modulus is responsible to generate scattering coefficients, where high order scattering coefficients are generated by convoluting wavelet modulus coefficients. The result obtained using ST is expressed as, $A_{4}$ .

3.5 Feature extraction with LGP

LGP [23] is face representation method to generate constant patterns and is irrespective of variation in intensities considering edges. The operator contained in LGP uses values of gradient pixel and is discovered as an intensity value. The least value of gradient amongst eight neighbouring pixels is termed as the value of the threshold. When the value of gradient and neighbouring pixel becomes high against threshold, then the allocated values of the pixel becomes ‘one’ else value becomes ‘zero’. Consider circle having radius $m$ positioned on pixel considering $b$ sample points on the circle. LGP uses bilinear interpolation to compute neighbourhood values of the pixel for $m$ and $b$ . The LGP utilizes $2\times m+1$ and $2\times m+1$ kernel for summarizing structures of iris. From center of pixel position $x_{a},s_{a}$ , LGP adapts $2\times m+1$ by $2\times m+1$ neighbouring pixels. The value of gradient between neighbouring pixel $n_{i}$ and center pixel $n_{a}$ is denoted as, $y_{c}=|n_{i}-n_{a}|$ and described the average gradient value of $b$ as, $\bar{y}=\frac{1}{b}\sum_{i=0}^{b-1}y_{i}$ . Thus, $\textit{LGP}_{b,m}(r_{a},s_{a})$ is represented as,

$\displaystyle\textit{LGP}_{b,m}(r_{a},s_{a})=\sum_{i=0}^{b-1}p(y_{i}-\bar{y})2% ^{n}$ (3)

where, $p(r)=\begin{cases}0,&\text{if}∼{}r<0,\\ 1,&\text{otherwise}.\end{cases}$ .

Feature vector: The feature vector is expressed as,

$\displaystyle\gamma=\{\gamma_{1},\gamma_{2},\ldots,\gamma_{w}\};∼{}\text{for}∼% {}(1<\tau<w),$ (4)

where, $\gamma$ indicate features valuated with the input image $X_{i}$ and is expressed as, $[1\times w]$ , and $w$ indicate total features.

3.5.1 Dimensional reduction with PCA

PCA [24] is a technique utilized to reduce dimensionality and for performing multivariate analysis. This method is an optimal scheme for compressing huge dimensional vectors into least dimensional vectors and evaluates attributes from data. PCA extracted fewer components to achieve parsimony to reduce the dimensionality. PCA uses either the covariance matrix or multivariate set to extract Principal Components. PCA easily performs the compression and decompression operations using matrix multiplication. PCA is utilized in most of the applications, like time series prediction, pattern recognition, exploratory data analysis, data compression, visualization, and image processing. The PCA model is represented as,

$\displaystyle z_{oc1}=D_{ocw}c_{wc1}$ (5)

where, $z$ is $o$ dimensional vector using projection $c$ , and $w$ indicate feature vector size as $\left(o<w\right)$ . The covariance matrix $E$ is expressed as,

$\displaystyle E=\frac{1}{\rho-1}\sum_{\tau=1}^{\rho}(c-\kappa)(c-\kappa)^{G}$ (6)

where, $\kappa$ indicate mean vector of $c$ . The eigenvectors $\beta_{\tau}$ are formulated as,

$\displaystyle(E-\phi_{\tau}\vartheta)\beta_{\tau}=0;\tau=1,2,\ldots,w$ (7)

where, $\phi_{\tau}$ represent eigenvectors of $E$ . The projection matrix is expressed as,

$\displaystyle D=J^{G}$ (8)

where, $J$ poses $o$ eigenvectors and $D$ indicates $o c w$ matrix.

The feature vector with reduced dimension is expressed as,

$\displaystyle\gamma=\{\gamma_{1},\gamma_{2},\ldots,\gamma_{\tau}\ldots\gamma_{% o}\};∼{}\text{for}∼{}(1<\tau<o).$ (9)

where, $o$ indicate features with reduced dimension and $o<w$ .

Figure 2.

Structural design of DBN.

3.6 Iris recognition using the proposed chronological MBO-based DBN

The reduced features with low dimensions generated with PCA are subjected as an input to DBN for recognizing each sample. The effectual recognition is generated with optimal DBN tuning using Chronological MBO. DBN learns non-linear complex relationship that subsists in real-life, and chronological MBO trained DBN classifier. The iris recognition is carried out using chronological MBO-based DBN, which is a combination of chronological concept and MBO algorithm [25] for training DBN which is based on migration features of each butterfly. The benchmark MBO is an effective tool for several types of optimization problems. It involves fine-tuning of attributes and complication-free evaluation. It is applicable for the parallel processing and has the capability to develop the trade-off among intensification and diversification. Also, issues with high dimensions are efficiently handled with MBO. The main problem of this method is the premature convergence. Nevertheless, the speed of searching and speed of convergence are improved by combining a chronological concept that describes weights from previous iterations for revising new weights and biases. The benefit of the DBN classifier includes the effective generation of solutions for the optimization problems. In addition, the DBN classifier can be used in various engineering applications to provide numerous benefits to the industries. Hence, we use MBO with the chronological concept to enhance the performance of the DBN classifier.

3.6.1 Structural design of DBN

DBN is developed using a single MLP layer and two RBM layers, as portrayed in Fig. 2. In DBN, no connection persists amongst hidden and visible neurons, as the connection is associated between hidden and visible neurons. In first RBM, the feature vector $\gamma$ is fed as an input to the visible layer. The hidden layer output of the first RBM is subjected to second RBM, and the second RBM output is fed to MLP.

The input of the visible layer in first RBM and output from the hidden layer is expressed as follows,

$\displaystyle g^{1}=\{g_{1}^{1},g_{2}^{1},\ldots,g_{b}^{1},\ldots,g_{10}^{1}\}% ;1\leqslant b\leqslant r$ (10) $\displaystyle a^{1}=\{a_{1}^{1},a_{2}^{1},\ldots,a_{k}^{1},\ldots,a_{d}^{1}\};% 1\leqslant k\leqslant d$ (11)

where, $r$ indicates the total number of neurons $g_{b}^{1}$ denotes $b^{\rm th}$ visible neuron in first RBM, $a_{k}^{1}$ represent $k^{\rm th}$ hidden neuron and $d$ represents total hidden neurons. Each neuron contained in hidden layer and visible layer poses bias. Let $h$ and $z$ represents bias in hidden as well as the visible layer. The two biases, linked to neurons for first RBM is represented as,

$\displaystyle z^{1}=\{z_{1}^{1},z_{2}^{1},\ldots,z_{b}^{1},\ldots,z_{r}^{1}\}$ (12) $\displaystyle h^{1}=\{h_{1}^{1},h_{2}^{1},\ldots,h_{k}^{1},\ldots,h_{d}^{1}\}$ (13)

where, $z_{b}^{1}$ represents the bias related to $b^{\rm th}$ visible neuron and $h_{k}^{1}$ represent the bias related to $k^{\rm th}$ hidden neuron. The weight of the first RBM is denoted as,

$\displaystyle P^{1}=\left\{P_{bk}^{1}\right\};1\leqslant b\leqslant r;1% \leqslant k\leqslant d$ (14)

where, $P_{bk}^{1}$ denotes the weight amid $b^{\rm th}$ visible neuron and $k^{\rm th}$ hidden neuron and the size of the weight vector is $r\times d$ . Thus, in RBM1, the hidden layer output is evaluated with bias and weight for each visible neuron and is represented as,

$\displaystyle a_{k}^{1}=\chi\left[h_{l}^{1}+\sum_{b}g_{b}^{1}P_{bk}^{1}\right]$ (15)

where, $\chi$ is an activation function.

The first RBM output is represented as,

$\displaystyle a^{1}=\left\{h_{k}^{1}\right\};∼{}1\leqslant k\leqslant d$ (16)

The input to the second RBM is based on output obtained from the hidden layer of the first RBM layer. The output from the first RBM is passed as an input of the visible layer in the next RBM. Thus, visible neurons in the second RBM are equal to count of hidden neurons exist in first RBM, thus it is expressed as,

$\displaystyle g^{2}=\{g_{1}^{2},g_{2}^{2},\ldots,g_{d}^{2}\}=\{a_{k}^{1}\};∼{}% 1\leqslant k\leqslant d$ (17)

where, $\{a_{k}^{1}\}$ is first RBM output.

The hidden layer of RBM2 is represented as,

$\displaystyle a^{2}=\{a_{1}^{2},a_{2}^{2},\ldots,a_{k}^{2},\ldots,a_{d}^{2}\};% ∼{}1\leqslant k\leqslant d$ (18)

The biases in the visible as well as the hidden layer are expressed as $u^{2}$ and $v^{2}$ . The weight vector of the second RBM is formulated as,

$\displaystyle P^{2}=\{P_{kk}^{2}\};∼{}1\leqslant k\leqslant d$ (19)

where, $P_{kk}^{2}$ is the weight amidst $k^{\rm th}$ visible neuron and $k^{\rm th}$ hidden neuron in RBM2 and dimension of the weight vector is $d\times d$ . The output of $k^{\rm th}$ hidden neuron is expressed as,

$\displaystyle a_{l}^{2}=\chi[h_{k}^{2}+\sum_{b}g_{b}^{2}P_{kk}^{2}]∼{}\forall g% _{b}^{2}=a_{k}^{1}$ (20)

where, $h_{k}^{2}$ indicate bias connected with $k^{\rm th}$ hidden neuron. Thus, hidden layer output is expressed as,

$\displaystyle a^{2}=\{a_{k}^{2}\};∼{}1\leqslant k\leqslant d$ (21)

The above equation is passed as input to the MLP and is represented as,

$\displaystyle y=\{y_{1},y_{2},\ldots,y_{k},\ldots,y_{d}\}=\{a_{k}^{2}\};1% \leqslant k\leqslant d$ (22)

where, $d$ denotes count of neurons present between input layer and hidden layer of the MLP is expressed as,

$\displaystyle p=\{p_{1},p_{2},\ldots,p_{r},\ldots,p_{L}\};∼{}1\leqslant T\leqslant L$ (23)

where, $L$ denotes total hidden neurons. Consider, $F_{T}$ as the bias of $T^{\rm th}$ hidden neuron, where $T=1,2,\ldots,L$ . The MLP output is denoted as,

$\displaystyle q=\{q_{1},q_{2},\ldots,q_{m},\ldots,q_{h}\};∼{}1\leqslant m\leqslant h$ (24)

where, $h$ represents the number of neurons and the MLP consists of two weight vectors. Let $P^{X}$ denotes the weight amidst input and hidden layer and is given below,

$\displaystyle P^{X}=\{P_{kT}^{X}\};∼{}1\leqslant k\leqslant d;∼{}1\leqslant T\leqslant L$ (25)

where, $P_{kT}^{X}$ denotes the weight amidst $k^{\rm th}$ input neuron and $T^{\rm th}$ hidden neuron and size of $P^{X}$ is $d\times L$ . The output of the hidden layer is computed using weight and bias of neurons and is calculated as,

$\displaystyle p_{T}=\left[\sum_{k=1}^{d}P_{kT}^{X}*y_{k}\right]F_{T}∼{}\forall y% _{k}=a_{k}^{2}$ (26)

where, $F_{T}$ indicate bias of hidden neuron and $y_{k}=a_{k}^{2}$ , as the output of second RBM and is fed to MLP. The weight $P^{S}$ is represented as,

$\displaystyle P^{S}=\{P_{Tm}^{S}\};1\leqslant T\leqslant L;1\leqslant m\leqslant h$ (27)

The output vector is calculated using the weight $P^{S}$ and the hidden layer output is represented as,

$\displaystyle q_{m}=\sum_{T=1}^{L}P_{Tm}^{S}*p_{T}$ (28)

where, $P_{Tm}^{S}$ indicate weight amidst $T^{\rm th}$ hidden neuron and $m^{\rm th}$ output neuron and $p_{T}$ is hidden layer output.

3.6.2 Training of chronological MBO-DBN

The training process of chronological MBO-based DBN is discussed here. The effective detection is generated by performing optimal tuning using proposed Chronological MBO-DBN. The benchmark MBO utilizes fine-tuning of attributes and complex-free calculation for improving the efficiency of proposed Chronological MBO-DBN. The pseudo-code of the proposed chronological MBO is shown in Algorithm 1.

a) Training of RBM layer 1.

The training sample $\gamma$ is passed as input to the RBM first layer. RBM1 calculates the probability distribution for each hidden neurons and computes the positive gradient using the visible vector and the hidden layer probability. The weight is evaluated with positive and negative gradient and weight is updated at the end of the iteration.

b) Training of RBM layer 2.

The output acquired from RBM layer1 is subjected as an input to the visible layer of RBM2 and discovers probability distribution in DBN. The energy is evaluated for weights $P^{2}$ and weight with limited energy is chosen for the subsequent iteration.

c) Training of MLP.

The obtained output using RBM layer 2 is passed as an input to the MLP layer. The network is adjusted by analyzing data iteratively to attain the best weight. The chronological MBO is adopted to determine the optimal weight and enhances the performance of the DBN classifier.

The training procedure of MLP layer is shown as follows:

viii) i)
Initialize the weights $P^{X}$ and $P^{S}$ randomly.
ii)
Interpret the input sample $\left\{a_{k}^{2}\right\}$ obtained using the output of the previous layer.
iii)
Compute $p_{r}$ and $q_{m}$ using the Eqs (26) and (28).
iv)
Compute the average error, the desired output and the output obtained as,

$\displaystyle H_{\textit{avg}}=\frac{1}{i}\sum_{n=1}^{i}(q_{m}^{n}-t_{m}^{n})^% {2};∼{}1\leqslant m\leqslant h$

where, $q_{m}^{n}$ is the output obtained and $t_{m}^{n}$ specifies the desired output.
v)
Evaluate weights in the hidden layer and visible layer by considering partial derivative of the Eq. (24), as

$\displaystyle\Delta P_{kT}^{X}=-\eta\frac{\partial H_{\textit{avg}}}{\partial P% _{kT}^{X}}$ (30) $\displaystyle\Delta P_{Tm}^{S}=-\eta\frac{\partial H_{\textit{avg}}}{\partial P% _{Tm}^{S}}$ (31)

where, $\eta$ denotes the learning rate.
vi)
The gradient descent is applied to compute the new weights of the hidden and visible layer as,

$\displaystyle P_{kT(X)}^{X}(l+1)=P_{kT}^{X}(l)+\Delta P_{kT}^{X}$ (32) $\displaystyle P_{Tm(X)}^{S}(l+1)=P_{Tm}^{S}(l)+\Delta P_{Tm}^{S}$ (33)

where, $P_{Tm}^{S}(l)$ and $P_{kT}^{X}(l)$ are weights in the hidden and visible layer at the present iteration $l$ .
vii)
Evaluate new weight of hidden and visible layer using the chronological MBO-based DBN and is represented as,

$\displaystyle l_{r,y}^{l+1}=\frac{z_{r,y}^{l}+z_{r,y}^{l-1}+3\beta\times(tz_{y% }-0.5)}{2}$
viii)
Calculate the error function $H_{\textit{avg}(T)}$ using Eq. (iv)) for the updated weight using the gradient descent.
ix)
Calculate the error function $H_{\textit{avg}(P)}$ using Eq. (iv)) for the updated weight using the chronological MBO.
x)
Select new weights $W_{lK}^{I}(t+1)$ and $W_{Kr}^{H}(t+1)$ using error minimization as,

$\displaystyle P_{kT}^{X}(l+1)$ (35) $\displaystyle=\begin{cases}P_{kT(C)}^{X}(l+1);&\text{if}∼{}H_{\textit{avg}(P)}% <H_{\textit{avg}(X)}\\ P_{kT(X)}^{X}(l+1);&\text{otherwise}\end{cases}$ $\displaystyle P_{Tm}^{S}(l+1)$ (36) $\displaystyle=\begin{cases}P_{Tm(P)}^{S}(l+1);&\text{if}∼{}E_{\textit{avg}(P)}% <E_{\textit{avg}(X)}\\ P_{Tm(X)}^{S}(l+1);&\text{otherwise}\end{cases}$
xi)
Repeat the steps ii to x, till the optimal weight is identified.

Algorithm 1 explains the pseudo-code of the proposed chronological-MBO.

Algorithm 1
Pseudocode of the proposed chronological-MBO algorithm

Chronological-MBO algorithm for optimal weight calculation

1 Input: $z_{r,y}^{l}$ ; $r\leqslant b$

2 Output: Optimum solution, $z^{*}$ .

3 Begin

4 Initialize related parameters

5 Maximal step size $Q$ , adjusting rate $C$ , migrating rate $L$ , migration ratio $T$

6 Fitness calculation $\delta_{l}=\sum_{v=1}^{a}(B_{v}^{l}-X_{v})$

7 While $l<b$

8 Sort the butterflies based on fitness

9 Sub-divide the population as $K_{1}$ and $K_{2}$

10 For $\varphi<K_{1}$ (Sub-population 1)

11 Update the population based on the migration

operator

12 End For $\varphi$

13 For $r<K_{2}$ (Sub-population 2)

14 Update the population based on the adjusting

operator using Eq. (vii))

15 End For $r$

16 Combine $K_{1}$ and $K_{2}$

17 Evaluate the solutions

18 $l=l+1$

19 End while

20 Derive the best solution

21 End

4. Results and discussion

Chronological-MBO algorithm for optimal weight calculation
1	Input: $z_{r,y}^{l}$ ; $r\leqslant b$
2	Output: Optimum solution, $z^{*}$ .
3	Begin
4	Initialize related parameters
5	Maximal step size $Q$ , adjusting rate $C$ , migrating rate $L$ , migration ratio $T$
6	Fitness calculation $\delta_{l}=\sum_{v=1}^{a}(B_{v}^{l}-X_{v})$
7	While $l<b$
8	Sort the butterflies based on fitness
9	Sub-divide the population as $K_{1}$ and $K_{2}$
10	For $\varphi<K_{1}$ (Sub-population 1)
11	Update the population based on the migration
	operator
12	End For $\varphi$
13	For $r<K_{2}$ (Sub-population 2)
14	Update the population based on the adjusting
	operator using Eq. (vii))
15	End For $r$
16	Combine $K_{1}$ and $K_{2}$
17	Evaluate the solutions
18	$l=l+1$
19	End while
20	Derive the best solution
21	End

Here, the efficiency of the proposed IAAD method based on the Chronological MBO-DBN classifier is evaluated along with other classical strategies.

4.1 Experimental setup

The analysis is carried out using the CASIA-IrisV4 dataset [26], and execution is done in MATLAB. CASIA-IrisV4 consists of 54,601 iris images with 1,800 genuine subjects and 1,000 virtual subjects. The images of iris employ 8 bit gray-level JPEG files, accumulated from NII. Also, in this dataset, the images are taken by the advanced biometric sensor, which can recognize users from 3 meters away by actively searching iris in the visual field via an intelligent multi-camera imaging system. The features of collected images are extracted and the analysis is carried out using the metrics, namely FRR, FAR, accuracy, F-Measure, and ROC.

4.2 Performance metrics

The metrics, such as FRR, FAR, accuracy, F-measure, and ROC explained below.

4.2.1 Accuracy

The accuracy indicates the precision of recognizing iris using the modality of the iris and is expressed as,

$\displaystyle\textit{Accuracy}=\frac{\textit{Jc}+\textit{Ju}}{\textit{Jc}+% \textit{Rc}+\textit{Ju}+\textit{Ru}}$ (37)

where, Ju represent true positives, Jc indicate true negatives. Ru is false positives and Rc stands for false negatives.

4.2.2 FRR

FRR indicates the proportion of false rejection with respect to indisputable tries, and is formulated as,

$\displaystyle\textit{FRR}=\frac{\textit{Rc}}{\textit{Ju}+\textit{Rc}}=1-% \textit{TPR}$ (38)

4.2.3 FAR

FAR is the proportion of false attempt to pretend attempts, and is formulated as,

$\displaystyle\textit{FAR}=\frac{\textit{Ru}}{\textit{Ru}+\textit{Jc}}=1-% \textit{TNR}$ (39)

4.2.4 F-measure

F-Measure a single measure, which is calculated by using precision and recall and is calculated as,

$\displaystyle\textit{F-Measure}=2*\left[\frac{\textit{Precision}*\textit{% recall}}{\textit{Precision}+\textit{recall}}\right]$ (40)

4.2.5 ROC

ROC indicates the relation between TNR and TPR, which is adapted for evaluating system performance.

4.3 Experimental results

Here, the results of the analysis using the proposed Chronological MBO based DBN approach by considering the input image 1 and the input image 2. The input image with resolution 640*480 is shown in Fig. 3. Figure 3a portrays input image 1, Fig. 3b output of the pre-processing image 1 and Fig. 3c The output of the HT with respect to the input image 1. Figure 3d portrays input image 2, Fig. 3e output of the pre-processing image 2, and the output of the HT with respect to the input image 2 is shown in Fig. 3f.

Figure 3.

Sample results (a) input image 1, (b) Output of the Pre-processing image 1, (c) Output of Hough transform for image 1, (d) input image 2, (e) Output of the Pre-processing image 2, (f) Output of Hough transform for image 2.

The output generated from HT is subjected to a rubber sheet model for normalizing image, and the outcome is portrayed in Fig. 4. The resolution of the normalized image is similar to the resolution of the original image, i.e., resolution-640*480. The normalized output of image 1 is shown in Fig. 4a. The normalized output for image 2 is shown in Fig. 4b. The output computed from the descriptor is shown in Fig. 4c and d. The histogram representation for image 1 is depicted in Fig. 4e, and Fig. 4f denotes histogram representation for image 2.

Figure 4.

Sample results (a) Normalized image for the input image 1, (b) Normalized image for the input image 2, (c) Output of descriptor for image 1, (d) Output of descriptor for image 2, (e) Histogram representation for image 1, (f) Histogram representation for image 2.

4.4 Comparative methods

This section deliberates the methods used for the comparison, as deep learning [3], BPNN [27], DBN [28], and spectral matching [5]. The proposed Chronological MBO establishes effectiveness in recognizing iris by comparing classical techniques.

4.5 Comparative analysis

The analysis using performance measures of the proposed Chronological MBO-based DBN approach is discussed and evaluation is done by altering training data.

4.5.1 Comparative analysis using image 1

Figure 5 depicts the analysis of methods using FAR, accuracy, and FRR using k-fold validation. Figure 5a portrays the analysis of FRR using k-fold validation with image 1. When the value for k-fold is 8, the FRR obtained by spectral matching is 1.8897%, deep learning is 1.5322%, BPNN is 1.0055%, Chronological-MBO-NN is 0.5%, DBN is 0.4852%, and Chronological-MBO DBN without PCA is 0.4851, whereas the proposed Chronological-MBO DBN attained improved FRR rate as 0.4812%. Figure 5b depicts the analysis of FAR using image 1. When the value for k-fold is 8, FAR rate obtained by spectral matching is 0.5495%, deep learning is 0.5413%, BPNN is 0.5164%, Chronological-MBO NN is 0.5%, DBN is 0.4933%, and Chronological-MBO DBN without PCA is 0.4932, while proposed Chronological-MBO DBN is 0.4850%, respectively. Figure 5c portrays comparative analysis for accuracy using image 1. If the k-fold value is 8, then the accuracy attained by the methods, spectral matching is 94.3116%, deep learning is 95.0821%, BPNN is 96.2302%, Chronological-MBO NN is 97.0000%, DBN is 97.9610% and Chronological-MBO DBN without PCA is 97.9708% while proposed Chronological-MBO DBN is 98.0571%. Figure 5d shows the comparative analysis for F-Measure using image 1. For k-fold is 8, the F-Measure of the methods, such as spectral matching, deep learning, BPNN, Chronological-MBO NN, DBN, Chronological-MBO DBN without PCA, and the proposed Chronological-MBO DBN is 94.7966%, 95.5754%, 96.7234%, 97.5%, 98.4774%, 98.5121%, and 98.6066%, respectively.

Figure 5.

Comparative analysis using k-fold with image 1, (a) FRR, (b) FAR, (c) Accuracy, and (d) F-Measure.

Figure 6 shows the analysis of methods by altering training data with image 1. Figure 6a portrays the analysis of FRR using training data with image 1. When training data is 70%, FRR obtained by spectral matching is 14.0897%, deep learning is 4.3571%, BPNN is 4.2024%, Chronological-MBO NN is 1.9286%, DBN is 1.8646, and Chronological-MBO DBN without PCA is 1.8644. Moreover, the proposed Chronological-MBO DBN attained FRR of 1.8476%. Figure 6b depicts the analysis of FAR using image 1. If the training data is 70%, the FAR rate attained by existing spectral matching is 0.6239%, deep learning is 0.5320%, BPNN is 0.5300%, Chronological-MBO NN is 0.5100%, DBN is 0.5031, and Chronological-MBO DBN without PCA is 0.503, whereas proposed Chronological-MBO DBN obtained better FAR of 0.4946%. Figure 6c portrays the analysis of methods with accuracy using image 1. When training percentage is 70, the accuracy achieved by spectral matching is 85.0911%, deep learning is 93.9267%, BPNN is 94.1188%, Chronological-MBO NN is 96.0396, DBN is 96.9782%, Chronological-MBO DBN without PCA is 96.9879, and proposed Chronological-MBO DBN attained improved accuracy of 97.0721%. Figure 6d shows the comparative analysis for F-Measure using image 1. For 70% training data, the F-Measure of the methods, such as spectral matching, deep learning, BPNN, Chronological-MBO NN, DBN, Chronological-MBO DBN without PCA, and the proposed Chronological-MBO DBN is 85.5856%, 94.4298%, 94.6219%, 96.5496%, 97.5082%, 97.5199%, and 97.696%, respectively.

Figure 6.

Comparative analysis using training percentage with image 1, (a) FRR, (b) FAR, (c) Accuracy, and (d) F-Measure.

4.5.2 Comparative analysis using image 2

The analysis of methods using k-fold validation with image 2 is shown in Fig. 7. Figure 7a portrays the analysis of FRR using k-fold with image 2. If the k-fold value is 8, the FRR rate evaluated by spectral matching is 2.9446%, deep learning is 1.3673%, BPNN is 0.9954%, Chronological-MBO NN is 0.5000%, DBN is 0.4830%, and Chronological-MBO DBN without PCA is 0.4829, while FAR of proposed Chronological-MBO DBN is 0.4784%, respectively. Figure 7b depicts the analysis of FAR using image 2. When the value for k-fold is 8, the FAR rate measured by spectral matching is 0.5944%, deep learning is 0.5321%, BPNN is 0.5183%, Chronological-MBO NN is 0.5%, DBN is 0.4932%, and Chronological-MBO DBN without PCA is 0.4931, whereas the proposed Chronological-MBO DBN attained better FAR rate with the value of 0.4848%. Figure 7c portrays comparative analysis for accuracy using image 2. If the k-fold value is 8, the accuracy attained by spectral matching is 92.1637%, deep learning is 95.4572%, BPNN is 96.2240%, Chronological-MBO NN is 97.00%, DBN is 97.9581%, and Chronological-MBO DBN without PCA is 97.9678%, however, proposed Chronological-MBO DBN attained improved accuracy of 98.0539%. The comparative analysis for F-Measure is depicted in Fig. 4d. For the k-fold is 8, the F-Measure of the existing methods, such as spectral matching, deep learning, BPNN, Chronological-MBO NN, DBN, and Chronological-MBO DBN without PCA is 92.6485%, 95.9504%, 96.7172%, 97.5%, 98.4764%, and 98.5%, respectively. For the same k-fold, the F-Measure of the proposed Chronological-MBO DBN is 98.6482%, which is maximum than the all other existing methods.

Figure 7.

Comparative analysis using k-fold with image 2, (a) FRR, (b) FAR, (c) Accuracy, and (d) F-Measure.

The evaluation of methods based on training percentage using image 2 is portrayed in Fig. 8. Figure 8a portrays the analysis of FRR using training percentage with image 2. When the training data is 70%, the FRR acquired by spectral matching is 11.0548%, deep learning is 3.0429%, BPNN is 1.0714%, Chronological-MBO NN is 0.9286%, DBN is 0.8858%, and Chronological-MBO DBN without PCA is 0.8857, but proposed Chronological-MBO DBN attained FRR rate of 0.8744%. Figure 8b depicts the analysis of FAR using image 2. If the training data is 70%, the FAR of existing spectral matching is 0.5984%, deep learning is 0.5282%, BPNN is 0.5081%, Chronological-MBO NN is 0.5060%, DBN is 0.4992%, Chronological-MBO DBN without PCA is 0.4992, and proposed Chronological-MBO DBN obtained FAR of 0.4909%. Figure 8c portrays the analysis of methods using accuracy with image 2. When training percentage is 70, the accuracy achieved by the methods, spectral matching is 87.5319%, deep learning is 94.2948%, BPNN is 96.2271%, Chronological-MBO NN is 96.4203%, DBN is 97.3678%, and Chronological-MBO DBN without PCA is 97.3775, whereas proposed Chronological-MBO DBN attained an accuracy of 97.4625%. The comparative analysis for F-Measure is depicted in Fig. 4d. For the 70% training data, the F-Measure of the existing methods, such as spectral matching, deep learning, BPNN, Chronological-MBO NN, DBN, and Chronological-MBO DBN without PCA is 88.0227%, 94.794%, 96.7263%, 96.9264%, 97.8758%, and 97.9056%, respectively. For the same k-fold, the F-Measure of the proposed Chronological-MBO DBN is 98.0609%, which is maximum than the all other existing methods.

Table 1

Comparative discussion

Methods		Spectral matching	Deep learning	BPNN	Chronological- MBO-NN	DBN	Chronological-MBO DBN without PCA	Proposed Chronological- MBO DBN
FRR (%)	k-fold	1.3055	1.0372	0.8577	0.5000	0.4798	0.4798	0.4745
	Training percentage	10.9120	2.9000	0.7860	0.7860	0.7450	0.7450	0.7340
FAR (%)	k-fold	0.5289	0.5198	0.5131	0.5000	0.4932	0.4931	0.4847
	Training percentage	0.5964	0.5261	0.5040	0.5040	0.4972	0.4972	0.4889
Accuracy (%)	k-fold	95.642	96.034	96.227	97.000	97.964	97.975	98.061
	Training percentage	87.725	94.488	96.614	96.614	97.563	97.572	97.658
F-Measure (%)	k-fold	91.494	95.948	96.531	97.500	98.476	98.500	98.658
	Training percentage	88.214	94.985	97.111	97.118	98.067	98.098	98.253

Figure 8.

Comparative analysis using training percentage with image 2, (a) FRR, (b) FAR, (c) Accuracy, and (d) F-Measure.

Table 2

Pair-wise statistical test of algorithms on FRR

	Spectral matching	Deep learning	BPNN	Chronological-MBO-NN	DBN	Chronological-MBO based DBN without PCA	Proposed Chronological-MBO based DBN
Spectral matching	–	0.231	0.432	0.134	0.267	0.654	0.0315
Deep learning	0.675	–	0.234	0.543	0.124	0.432	0.0315
BPNN	0.432	0.234	–	0.563	0.234	0.456	0.2330
Chronological-MBO-NN	0.245	0.156	0.564	–	0.345	0.321	0.0315
DBN	0.365	0.254	0.178	0.419	–	0.390	0.0315
Chronological-MBO based DBN without PCA	0.187	0.687	0.456	0.0315	0.124	–	0.0315
Proposed chronological-MBO based DBN	0.576	0.198	0.234	0.456	0.218	0.0315	–

Table 3

Pair-wise statistical test of algorithms on FAR

	Spectral matching	Deep learning	BPNN	Chronological-MBO-NN	DBN	Chronological-MBO based DBN without PCA	Proposed Chronological-MBO based DBN
Spectral matching	–	0.231	0.432	0.134	0.267	0.654	0.0315
Deep learning	0.675	–	0.234	0.543	0.124	0.432	0.0315
BPNN	0.432	0.0315	–	0.563	0.234	0.456	0.2330
Chronological-MBO-NN	0.245	0.156	0.564	–	0.345	0.321	0.2740
DBN	0.365	0.254	0.178	0.419	–	0.0315	0.0315
Chronological-MBO based DBN without PCA	0.187	0.687	0.456	0.0315	0.124	–	0.0315
Proposed chronological-MBO based DBN	0.576	0.198	0.234	0.456	0.218	0.378	–

Table 4

Pair-wise statistical test of algorithms on accuracy

	Spectral matching	Deep learning	BPNN	Chronological-MBO-NN	DBN	Chronological-MBO based DBN without PCA	Proposed Chronological-MBO based DBN
Spectral matching	–	0.231	0.432	0.134	0.267	0.654	0.0315
Deep learning	0.675	–	0.234	0.543	0.124	0.432	0.0315
BPNN	0.432	0.265	–	0.563	0.234	0.456	0.2330
Chronological-MBO-NN	0.245	0.156	0.564	–	0.345	0.321	0.0315
DBN	0.365	0.254	0.178	0.419	–	0.0315	0.0315
Chronological-MBO based DBN without PCA	0.187	0.0315	0.456	0.234	0.124	–	0.0315
Proposed chronological-MBO based DBN	0.576	0.198	0.234	0.0315	0.218	0.378	–

Figure 9 depicts the analysis based on ROC using image 1 and image 2. Figure 9a shows ROC analysis made using image 1. If FPR value is 0.8%, the TPR obtained using spectral matching is 86.6425%, deep learning is 95.9405%, BPNN is 97.6357%, Chronological-MBO NN is 98.0714%, DBN is 98.1282%, Chronological-MBO DBN without PCA is 98.1283%, and proposed Chronological-MBO DBN is 98.1433%. The ROC analysis with image 2 is portrayed in Fig. 9b. When FPR value is 0.8%, then TPR attained by spectral matching is 89.088%, deep learning is 97.1%, BPNN is 99.048%, Chronological-MBO NN is 99.214%, DBN is 99.256%, Chronological-MBO DBN without PCA is 98.1283%, 99.2557%and proposed Chronological-MBO DBN attained TPR of 99.267%.

4.6 Comparative discussion

Table 1 deliberates the evaluation of strategies using FPR, FAR, accuracy, and F-Measure. From the table, while considering k-fold and the training percentage the proposed Chronological-MBO DBN is better than the existing methods, such as Spectral Matching deep learning, BPNN, Chronological-MBO NN, DBN, and Chronological-MBO DBN without PCA, for the metrics FRR, FAR, accuracy and F-Measure. Thus, it depicts that the proposed Chronological MBO-DBN attained minimum FRR, minimum FAR, maximum accuracy, and maximum F-Measure. Hence, the proposed classifier offers high performance.

Table 5
Pair-wise statistical test of algorithms on F-measure

	Spectral matching	Deep learning	BPNN	Chronological-MBO-NN	DBN	Chronological-MBO based DBN without PCA	Proposed Chronological-MBO based DBN
Spectral matching	–	0.231	0.432	0.134	0.267	0.654	0.0315
Deep learning	0.675	–	0.234	0.543	0.124	0.432	0.0315
BPNN	0.432	0.265	–	0.563	0.234	0.456	0.2330
Chronological-MBO-NN	0.245	0.156	0.564	–	0.345	0.0315	0.1533
DBN	0.365	0.254	0.178	0.419	–	0.0315	0.0315
Chronological-MBO based DBN without PCA	0.187	0.0315	0.456	0.234	0.124	–	0.0315
Proposed Chronological-MBO based DBN	0.576	0.198	0.234	0.0315	0.218	0.378	–

Figure 9.

ROC analysis (a) with image 1, (b) with image 2.

4.7 Statistical test

This section evaluates the statistical test based on different combinations of algorithms. Table 2 depicts the statistical analysis based on the metrics FRR by using the different combinations of algorithms. The table shows the $p$ values of different algorithms, and in this hypothesis testing, $p$ values should be less than 0.1 to reject the null hypothesis. From this table, the proposed method has rejected the maximum null hypothesis, with a minimum value of 0.0315. Table 3 shows the statistical analysis of different algorithms on the performance metrics FAR. The proposed method has almost always rejected the null hypothesis, with the $p$ value of 0.0315. Similarly, Tables 4 and 5 show the statistical test based on the metrics, such as accuracy and F-Measure, in both cases, the proposed method avoids the null hypothesis, with lower $p$ values of 0.0315, as compared with the existing methods.

4.8 Threats to validity

This section discusses the threats to validity of the proposed method.

Construct validity threats: In iris recognition, fewer coefficient errors are possible. This kind of threat is solved by using optimized deep learning methods.

Internal validity threats: In segmentation, the localization of the pupil eyelids, eyelashes, and boundaries are required. If it fails to segment the iris, there is a possibility to generate false data, which will affect the recognition rates. In the proposed method, the results contain the papillary boundary, even if the prior center is not positioned at center of the pupil and the circular HT computes the exact pupil center.

External validity threats: The quality of the normalized image is a threat, but in the proposed work both the normalized and the original image have the same resolution.

Conclusion validity threats: It deals with the relation between the theory and observation of the results. Normally, the expected output and the real output of the system has a lot of variations. This threat has overcome by the proposed method by calculating the average error, in which the weights are updated based on the minimum error.

5. Conclusion

An IAAD method is devised for recognizing iris using proposed Chronological MBO-DBN. Here, the optimal weights are tuned with chronological MBO for training DBN. The iris image is pre-processed and the iris regions are extracted by applying the Hough Transform and the segmentation and normalization of iris are performed with rubber sheet model. The ScatT-Loop along with the LGP descriptor is utilized for extracting features of iris, which enables the recognition of iris. The recognition of iris using biometric features is carried out with proposed Chronological MBO and is trained with a deep learning model named DBN. The proposed Chronological based DBN performs the detection of iris by undergoing segmentation and normalization of images using the Daughman’s rubber sheet model. The features are mined by combining LGP and the ScatT-Loop descriptor. The experimentation is performed using the CASIA iris dataset and the performances are evaluated using the measures, like FAR, accuracy, and FRR. The proposed Chronological-MBO-DBN approach provides superior performance with least FRR of 0.4745%, least FAR of 0.4847%, highest accuracy of 96.078%, and maximum F-Measure of 98.658%. In future, the work will be extended by current optimization techniques to improve efficiency for recognizing iris.

Footnotes

Nomenclature

References

Othman

Dorizzi

. Impact of quality-based fusion techniques for video-based iris recognition at a distance. IEEE Transactions on Information Forensics and Security. 2015; 10(8): 1590-1602.

Ahmadi

Akbarizadeh

. Hybrid robust iris recognition approach using iris image pre-processing, two-dimensional gabor features and multi-layer perceptron neural network/PSO. IET Biometrics. 2017; 7(2): 153-162.

Al Waisy

Qahwaji

Ipson

Al Fahdawi

Nagem

TAM

. A multi-biometric iris recognition system based on a deep learning approach. Pattern Analysis and Applications. August 2018; 21(3): 783-802.

Kaudki

Bhurchandi

. A robust iris recognition approach using fuzzy edge processing technique. in: Proceeding of IEEE 9th International Conference on Computing, Communication and Networking Technologies (ICCCNT). Bangalore, India. July 2018; 1-6.

Ramaiah

Kumar

. Toward more accurate iris recognition using cross-spectral matching. IEEE Transactions on Image Processing. 2017; 26(1): 208-221.

Tan

Kumar

. Accurate iris recognition at a distance using stabilized iris encoding and zernike moments phase features. IEEE Transactions on Image Processing. 2014; 23(9): 3962-3974.

Nguyen

Fookes

Ross

Sridharan

. Iris recognition with off-the-shelf CNN features: A deep learning perspective. IEEE Access. 2018; 6: 18848-18855.

Rathgeb

Buchmann

Hofbauer

Baier

Uhl

Busch

. Methods for accuracy-preserving acceleration of large-scale comparisons in CPU-based iris recognition systems. IET Biometrics. 2017; 7(4): 356-364.

Wang

Tan

. Iris recognition based on multichannel gabor filtering. in: Proceedings of Fifth Asian Conference on Computer Vision. Melbourne, Australia. January 2002; 1: 279-283.

10.

Liu

Sun

Tan

. A code-level approach to heterogeneous iris recognition. IEEE Transactions on Information Forensics and Security. 2017; 12(10): 2373-2386.

11.

Bouzouina

Hamami

. Multimodal biometric: Iris and face recognition based on feature selection of iris with GA and scores level fusion with SVM. in: Bio-engineering for Smart Technologies (BioSMART), in Proceedings of IEEE 2nd International Conference. Paris, France. August 2017; 1-7.

12.

Zhao

Kumar

. Towards more accurate iris recognition using deeply learned spatially corresponding features. in: Proceedings of the IEEE International Conference on Computer Vision. Venice, Italy. October 2017; 22-29.

13.

Tan

Kumar

. Efficient and accurate at-a-distance iris recognition using geometric key-based iris encoding. IEEE Transactions on Information Forensics and Security. 2014; 9(9): 1518-1526.

14.

Tan

Kumar

. Efficient and accurate at-a-distance iris recognition using geometric key-based iris encoding. IEEE Transactions on Information Forensics and Security. 2014; 9(9): 1518-1526.

15.

Pillai

Puertas

Chellappa

. Cross-sensor iris recognition through kernel learning. IEEE Transactions on Pattern Analysis and Machine Intelligence. 2014; 36(1): 73-85.

16.

Lim

Lee

Byeon

Kim

. Efficient iris recognition through improvement of feature vector and classifier. ETRI Journal. 2001; 23(2): 61-70.

17.

Koh

Govindaraju

Chaudhary

. A robust iris localization method using an active contour model and hough transform. in: Proceedings of the 20th International Conference on Pattern Recognition. Istanbul, Turkey. 2010; 2852-2856.

18.

Popplewell

Roy

Ahmad

Shelton

. Multispectral iris recognition utilizing hough transform and modified LBP. in: Proceedings of IEEE International Conference on Systems, Man, and Cybernetics (SMC). San Diego, CA, USA. 2014; 1396-1399.

19.

Chawla

Oberoi

. A robust algorithm for iris segmentation and normalization using hough transform. in: Proceedings of the International Conference on Advanced Computing and Communication Technologies. Panipat, Haryana, India. November 2011.

20.

Chakraborti

McCane

Mills

Pal

. LOOP descriptor: Encoding repeated local patterns for fine-grained visual identification of lepidoptera. 2017.

21.

Pandey

Singh

Vatsa

. Face recognition using scatteringwavelet under illicit drug abuse variations. in: Proceedings of International Conference on Biometrics (ICB), Halmstad, Sweden. August 2016.

22.

Krommweh

. Tetrolet transform: A new adaptive haar wavelet algorithm for sparse image representation. Journal of Visual Communication and Image Representation. May 2010; 21(4): 364-374.

23.

Jun

Choi

Kim

. Local transform features and hybridization for accurate face and human detection. IEEE Transactions on Pattern Analysis and Machine Intelligence. June 2013; 35(6): 1423-1436.

24.

Labib

Vemuri

. An application of principal component analysis to the detection and visualization of computer network attacks. Annales Des Telecommunications. February 2006; 61(1–2): 218-234.

25.

Wang Deb

Cui

. Monarch butterfly optimization. Neural Computing and Applications. May 2015; 31: 1995-2014.

26.

CASIA-Iris dataset taken from http://biometrics.idealtest.org/dbDetailForUser.do?id=4, accessed on August 2018.

27.

Nithya

Lakshmi

. Enhancing iris recognition framework using feature selection and BPNN. Cluster Computing. 2019; 22: 12363-12372.

28.

Jan Vojt

. Deep neural networks and their implementation. 2016.

29.

Gomathi

Karlekar

. Ontology and hybrid optimization based SVNN for privacy preserved medical data classification in cloud. International Journal on Artificial Intelligence Tools. 2019; 28(3).

30.

Waykar

. Emerging trends in biometrics. ICWET2010 At TECT Mumbai. 2010.

31.

Veeraiah

Krishna

. Dr., Intrusion detection based on piecewise fuzzy c-means clustering and fuzzy naïve bayes rule. Multimedia Research (MR). 2018; 1(1): 27-32.

32.

George

Rajakumar

. Fuzzy aided ant colony optimization algorithm to solve optimization problem, intelligent informatics. Advances in Intelligent Systems and Computing. 2013; 182: 207-215.

33.

Ghosh

Rana

Kansal

. A nonlinear manifold detection based model for software defect prediction. Procedia Computer Science. 2018; 132: 581-594.

34.

Ghosh

Rana

Kansal

. Statistical assessment of nonlinear manifold detection-based software defect prediction techniques. International Journal of Intelligent Systems Technologies and Applications. 2019; 18(6).

35.

Naik

Satapathy

Ashour

Dey

. Social group optimization for global optimization of multimodal functions and data clustering problems. 2018; 30: 271-287.

36.

Satapathy

Naik

. Social group optimization (SGO): A new population evolutionary optimization technique. Complex & Intelligent Systems. 2016; 3.

Optimization driven deep belief network using chronological monarch butterfly optimization for iris recognition at-a-distance

Abstract

Keywords

1. Introduction

Contribution of the work

2. Motivation

2.1 Literature survey

3.1 Pre-processing

3.2 Hough transform for extraction of iris region

3.3.1 Segmentation

3.3.2 Iris normalization

3.4 Extraction of noteworthy features with ScatT-LOOP

3.4.1 LOOP descriptor

3.4.3 Scattering transform

3.5 Feature extraction with LGP

3.6.1 Structural design of DBN

4.1 Experimental setup

4.2 Performance metrics

4.2.1 Accuracy

4.3 Experimental results

4.5 Comparative analysis

4.5.1 Comparative analysis using image 1

Table 5 Pair-wise statistical test of algorithms on F-measure

4.8 Threats to validity

5. Conclusion

Footnotes

Nomenclature

References

Table 5
Pair-wise statistical test of algorithms on F-measure