J. Vis. Commun. Image R.

103 (2024) 104262

Contents lists available at ScienceDirect

Journal of Visual Communication and Image Representation

journal homepage: www.elsevier.com/locate/jvci

Full Length Article

Performance evaluation of efficient segmentation and classification based

iris recognition using sheaf attention network☆
Sushilkumar S. Salve a,* , Sandipann P Narote b
a
Research Scholar, Department of Electronics and Telecommunication, Shri J.J.T. University, Jhunjhunu, Rajasthan 333001, India
b
Head of Department, Department of Electronics and Telecommunication, Government Polytechnic, Pune, Maharashtra 411016, India

A R T I C L E I N F O A B S T R A C T

Keywords: Iris recognition, a precise biometric identification technique, relies on the distinct epigenetic patterns within the
Iris recognition iris. Existing methods often face challenges related to segmentation accuracy and classification efficiency. To
Sheaf attention network improve the accuracy and efficiency of iris recognition systems, this research proposes an innovative approach
Grayscale transformation
for iris recognition, focusing on efficient segmentation and classification using Convolutional neural networks
Wavelet transform
with Sheaf Attention Networks (CSAN). Main objective is to develop an integrated framework that optimizes iris
Up sampling network
segmentation and classification. Subsequently, dense extreme inception multipath guided up sampling network is
employed for accurate segmentation. Finally, classifiers including convolutional neural network with sheaf
attention networks are evaluated. The findings indicate that the proposed method achieves superior iris recog­
nition accuracy and robustness, making it suitable for applications such as secure authentication and access
control. By comparing with existing approaches CSAN obtains 99.98%, 99.35%, 99.45% and 99.65% accuracy
for the four different proposed datasets respectively.

1. Introduction individuals even in challenging conditions such as varying lighting or

occlusions [10].
Iris recognition stands out as a cutting-edge biometric technology As technology continues to evolve, iris recognition is expected to
due to its unparalleled accuracy and reliability in identifying individuals play an increasingly important role in enhancing security and identity
based on the unique patterns within the iris [1,2]. The iris is an ideal verification across various domains, including law enforcement,
biometric trait because its patterns are established early in life and healthcare, finance, and public services [26]. Its exceptional accuracy
remain stable over time, ensuring consistency and permanence for and resistance to fraud make it a trusted biometric modality for ensuring
identity verification purposes [3]. One of the key advantages of iris secure and reliable identification. The success and reliability of iris
recognition is its resistance to fraud and spoofing attempts. Unlike some recognition systems hinge on several critical components, each serving a
biometric modalities such as fingerprints, which can be replicated or specific role in the authentication process. Beginning with preprocess­
altered, the intricate and complex patterns of the iris are extremely ing, iris images undergo enhancement techniques to optimize their
difficult to duplicate [4,5]. This makes iris recognition a highly secure quality and improve the visibility of intricate details [27]. This step
method for authentication, particularly in applications where robust often includes contrast adjustments, normalization, and noise reduction,
security measures are essential, such as border control, secure access crucial for preparing iris images for subsequent stages. Iris segmentation
management, and national identification systems [6,7]. In practice, iris is a pivotal phase aimed at isolating the iris region from the overall eye
recognition systems capture high-resolution images of the iris, extract­ image. Accurate segmentation requires precise detection of iris bound­
ing unique features such as crypts, furrows, and freckles [8,9]. These aries and proper exclusion of irrelevant areas [28]. Central to this pro­
features are then converted into mathematical templates that can be cess is the accurate localization of the pupil, which establishes the center
compared with stored templates in a database. The matching process is of the iris and aids in its delineation.
typically fast and accurate, providing reliable identification of Feature extraction follows segmentation and involves transforming

☆
This paper has been recommended for acceptance by Kuo-Liang Chung.
* Corresponding author.
E-mail addresses: [email protected] (S.S. Salve), [email protected] (S.P. Narote).

https://doi.org/10.1016/j.jvcir.2024.104262
Received 18 October 2023; Received in revised form 29 April 2024; Accepted 16 August 2024
Available online 17 August 2024
1047-3203/© 2024 Elsevier Inc. All rights are reserved, including those for text and data mining, AI training, and similar technologies.
S.S. Salve and S.P. Narote Journal of Visual Communication and Image Representation 103 (2024) 104262

the complex patterns of the iris into a compact and unique representa­ integration of traditional computer vision techniques with cutting-edge
tion. This process distills the iris features into an iris template or code, deep learning models to enhance recognition accuracy. In 2023,
which serves as a mathematical representation of the iris’s distinguish­ Abdulhasan et al. [15] focused on achieving instant learning in iris
ing characteristics [29]. The iris template is securely stored for subse­ recognition by combining Deep Neural Networks (DNN) with Linear
quent comparison and verification tasks. In the classification stage, an Discriminant Analysis (LDA) to enhance accuracy. Instant learning aims
individual’s iris template is compared against previously stored tem­ to make iris recognition systems more user-friendly and adaptable.
plates in a database [30,31]. Matching algorithms are employed to Abdellatef et al. [16], in 2022, introduced an iris recognition system
assess the similarity between templates and determine potential that employs deep learning techniques and allows users to cancel and re-
matches. This final stage of the process enables secure and reliable enroll their biometric data, addressing security concerns and improving
identification based on the unique iris patterns. These sequential stages system robustness.
collectively contribute to the accuracy, reliability, and robustness of iris The classification methods used here are CNN and Support Vector
recognition technology, making it a valuable tool in various applications Machine (SVM) In 2022, Kagawade and Angadi [17] proposed a novel
requiring precise and secure biometric authentication.The major scheme for iris recognition using the Polar Fast Fourier Transform
contribution of the work are, (PFFT) code, leveraging symbolic modeling to enhance accuracy and
efficiency. Sun et al. [18] in 2022, presented an IR system based on Local
• The hybrid preprocessing method combines the strengths of both Circular Gabor Filters (LCGF) and a Multi-scale Convolution Feature
Hybrid multi-scale Retinex Adaptive grayscale Transformation Fusion Network (MCFFN), demonstrating the use of advanced image
(HRAT) and Quantized Haar Wavelet Transform (QHWT). This en­ processing techniques to extract relevant features from iris images. In
hances image quality by improving contrast, reducing noise, and 2022, Sun et al. [19] explored open-set IR using deep learning and the
preserving texture and fine details. Open-Class Features Outlier Network (OCFON), offering a solution for
• Dense extreme Inception Multipath guided up sampling Network classifying iris images, even when encountering unknown or unautho­
(DIMNet) is capable of accurately delineate the iris boundaries from rized users. In 2022, Lei et al. [20] introduced an attention Meta-
the surrounding ocular structures, thereby improving the precision Transfer Learning (MTL) approach for few-shot IR adapting the recog­
of the segmentation process. nition system to new, unseen iris samples. Table 1represents the sum­
• Double layer angle Multi kernel Extreme Learning Analysis (DMELA) mary of reviewed works.
is designed to efficiently extract features from segmented images. It Table 1 provides a concised summary of various methods employed
reduces feature space dimensions while preserving relevant infor­ in iris recognition, along with their specific objectives and associated
mation. This leads to faster and more efficient processing during limitations. Each method offers unique advantages and faces specific
classification. limitations, highlighting the diverse challenges and opportunities within
• The use of Convolutional Neural Networks (CNN) in conjunction the field of iris recognition.
with sheaf attention networks for classification is a powerful com­
bination. It allows for high accuracy in recognizing and classifying • Problem Statement
iris patterns, which are critical in biometric authentication.
• This classification approach is adaptable across diverse image clas­ Iris recognition using deep learning is to develop an accurate and
sification tasks, making it versatile and suitable for various appli­ reliable biometric authentication system that automatically identify and
cations beyond iris recognition. verify individuals based on the unique patterns present in the iris of the
eye. This system aims to address challenges related to identity verifi­
The remaining part of this work is organized as: Section 2 summaries cation and access control by leveraging deep neural networks to extract
various existing methods related to biometric iris recognition, Section 3 and learn intricate iris features from high-resolution images. In this
describes various proposed methods, Section 4 includes results and research work, an efficient segmentation and classification-based iris
performance evaluation and Section 5 involves conclusion and future recognition using CSAN is proposed. The primary objectives include
work. achieving high recognition accuracy, robustness to variations in lighting
and occlusions. The ability to distinguish genuine iris samples from
2. Literature survey impostors while ensuring data privacy and security in deployment sce­
narios such as border control, secure authentication, and access
Iris Recognition (IR) technology has made significant advancements management.
in recent years, with researchers continuously exploring innovative
techniques to enhance its accuracy and applicability. The selected arti­ 3. Proposed method
cles encompass various aspects of iris recognition, including biometric
techniques, edge detection methods, deep learning neural networks, The proposed methodology for iris recognition represents a
feature extraction, and instant learning. comprehensive and sophisticated approach that combines advance
In 2022, Farouk RH et al. [11] proposed a Hamming Distance (HD) techniques to improve the efficiency and accuracy of the recognition
like biometric technique to enhance iris recognition, reflecting the process. This method revolves around four key components: pre­
ongoing efforts to expand the reliability and accuracy of IR systems. It processing, segmentation, feature extraction, and classification.
underscores the significance of robust biometric methods for secure From Fig. 1, the preprocessing stage, powered by HRAT and QHWT,
authentication in various applications. In 2023, Khan SN et al. [12] optimizes image quality, setting the foundation for precise segmenta­
focused on iris recognition using edge detection methods, particularly tion. DIMNet, employed for segmentation, efficiently detects iris
for user identification in flight simulators. This highlights the practical boundaries and edges, crucial for accurate isolation. DMELA serves as a
applications of iris recognition beyond traditional security measures. In powerful feature extraction tool, reducing classifier runtime while
2023, Ali EH et al. [13] presented a method for IR using deep learning capturing essential iris features., K-Nearest Neighbors (KNN), SVM,
neural networks, such as CNN, with Semi-Discrete Matrix Decomposi­ CSAN and Random Forest (RF) are the classification method used to
tion (SDD) as the feature extraction method. Deep learning has emerged classify the extracted features for iris recognition.
as a game-changer in biometrics, offering increased accuracy and
robustness. In 2022, Babu G and Khayum PA [14] explored a hybrid 3.1. Preprocessing to enhance image quality
approach that combines Elephant Herding with Whale Optimization
Algorithm (EH-WOA) and CNN for IR. This fusion approach signifies the Initially, iris images undergo a critical preprocessing stage, which

2
S.S. Salve and S.P. Narote Journal of Visual Communication and Image Representation 103 (2024) 104262

Table 1 where X(u, v) represents the grayscale value of the input image, Xʹ(u, v)
Summary of existing methods. signifies the grayscale value of the resulting output image, u, v represents
References Methods Objectives Limitations pixel coordinates of image, T stands for parameters related to contrast
correction, L denotes the offset for the overall transformation, typically
[11] CNN and HD Improve Iris Specific details of the
Recognition. Potential technique are not set to a value of 127. The parameter B is equally important, as it is
for increased security verified. associated with brightness correction parameters. Its value is propor­
[12] Edge detection Use Edge Detection for Lack of information on tional to the average brightness of the image, as expressed in Eqn. (2).
methods User Identification. scalability
Improve recognition 1
B∝ (2)
speed mean
[13] SDD and CNN Utilizes deep learning Potential
for iris localization computational The adaptive nature of the parameters ensures that the preprocessing
complexity method effectively addresses the specific challenges posed by different
[14] EH-WOA Fusion of traditional Required more conditions within the datasets, ultimately leading to improved image
CV and deep learning computational
quality and readiness for subsequent analysis and recognition tasks.
techniques. Improved resources
feature extraction for HRAT enhances iris image clarity by optimizing contrast and
iris recognition brightness, revealing fine details crucial for feature extraction. It reduces
[15] LDA Iris recognition This model doesn’t noise without compromising textural integrity, crucial for precise bio­
employs a hybrid necessitate an metric recognition. Additionally, HRAT adjusts to different lighting
LDADNN model, which extensive training
combines 1D deep dataset or specialized
conditions, ensuring consistent image quality essential for accurate iris
learning for the hardware for recognition.
classification process deployment.
and delivers precise • Quantized Haar Wavelet Transform
outcomes.
[16] CNN and SVM To enhance security by The demands for
allowing users to computational QHWT is a powerful mathematical technique used for data analysis,
cancel and re-enroll resources could be particularly in the context of eye image preprocessing for the collected
biometric data substantial. image. QHWT is designed to efficiently organize data by their fre­
[17] PFFT Symbolic modeling Lack of information quencies. This transformation shifts data from the spatial domain into
approach improves regarding scalability.
accuracy and
the frequency domain, preserving each component at its respective
efficiency. Potential for resolution scale. Consider a wavelet as a foundational set for a vector
enhanced recognition space. The forward Haar transform explained as a combination of
in non-standard averaging and differencing operations. Eqn. (3) represents the scaling
conditions
function ϕ(s) for low pass and the partial Haar function ψ (s) for high
[18] LCGF+MCFFN The image processing No attention is given to
techniques improve distinguishing pass.
feature extraction. unknown classes. ⎧
Enhanced recognition ⎨1
⎪ s ∈ [0, 1/2)
accuracy and ψ (s) = − 1 s ∈ [1/2, 1)
reliability
⎪
⎩
0 s ∈ [0, 1)
[19] OCFON To address the Potential for overfitting √̅̅̅̅ ( i )
scenarios where the ψ j (s) = 2 × ψ 2 s − j ,
i i (3)
system encounters ( )
unknown irises i = 0, 1, .... and j = 0, 1, ...., 2i − 1 ,
{
[20] MTL To address the few- Need a more in-depth 1 0≤s<1
shot recognition examination of the ϕ(s) =
0 otherwise
scenarios with limited integration of attention
training samples. mechanisms and few-
Utilizes meta-transfer shot learning. where i , s and j represents the variables of Haar function. When dealing
learning and attention with an input signified a list of 2n numbers, QHWT is a straightforward
mechanisms process of pairing up input values, storing the differences, and then
adding up the sums. This recursive process continues by pairing up the
plays a pivotal role in optimizing image quality [21,22]. This phase sums, ultimately yielding 2n− 1 transformations and a concluding sum.
employs a HRAT and QHWT. HRAT enhances image quality by adjusting Consider a vector y with N values, represented as Y = [y1 , y2 , ..., yN ]
contrast, reducing noise, and ensuring uniform lighting conditions. where N must be a power of 2. The total number of recursive steps Ω in a
QHWT, on the other hand, allows for the transformation of image data wavelet transform is determined by a certain parameter like log2 N.
into a more analyzable frequency domain, preserving intricate iris Then, calculate the directional distances between pairs of numbers pa for
details. the high pass and the average of pairs of numbers in the vector ha for the
low pass, as described in Eqn. (4).
• Multi-Scale Retinex Adaptive Grayscale Transformation (y2a + ysa+1 )
ha =
2
The HRAT model plays a pivotal role in enhancing eye images as part ( ) (4)
(y2a − y2a+1 ) N
of the preprocessing stage. It essentially converts the gray values pa = for a = 0, ..., − 1
2 2
attained from real sampling into corresponding gray values using a
linear function. It achieves several important objectives including the where a represents the variable QHWT functions; N represents the
extraction of edge and texture information from the image and the number of values. The next data list p is obtained, in a way that allows to
enhancement of image contrast. Eqn. (1) shows the mathematical rep­ reconstruct the original vector Y from h and p as Y→[h|p]. Eqn. (5)
resentation of grayscale function. represents the inverse operation function.
Xʹ(u, v) = [X(u, v) + B − L]⋅tan(T⋅π) + B + L (1) [y1 , y2 , ...yN ]→[h|p] = [h1 , ....hN/2 |p1 , ..., pN/2 ] (5)

3
S.S. Salve and S.P. Narote Journal of Visual Communication and Image Representation 103 (2024) 104262

Fig. 1. Proposed architecture.

Finally, in the reverse operation, the vector is decomposed and rest of the eye. DIMNet is tailored to excel in this specific task, effectively
mapped as part of the eye image preprocessing procedure for datasets. detecting and delineating the boundaries of the iris.
QHWT shifts iris image data to the frequency domain, capturing
detailed features at multiple resolutions to enhance pattern differenti­ • Dense Extreme Inception Multipath guided up sampling
ation. It segments data into frequency components, boosting processing Network for Edge Detection (DIMNet)
efficiency, crucial for managing large datasets and ensuring rapid real-
time analysis. DIMNet inspired from the architecture of xception but introduces a
unique enhancement: two parallel skip-connections. These skip-
• Hybrid Multi-Scale Retinex Adaptive Grayscale Transformation connections play a pivotal role in preserving vital edge information
with Quantized Haar Wavelet Transform across various network layers [23]. Fig. 2 represents the architecture of
DIMNet.
Apply the grayscale transformation approach to the initial eye The architecture is structured around six blocks, each serving as an
image. This initial step serves to enhance image contrast, refine edge and encoder. Each block encompasses sub-blocks that consist of convolu­
texture details, and adjust brightness and contrast settings. Following tional layers. Skip-connections establish connections not only between
this initial enhancement, proceed to apply the QHWT to the output blocks but also among sub-blocks. Feature-maps generated at each block
obtained in the previous step. The QHWT analysis operates in the fre­ are then dispatched to an independent upsampling network, where they
quency domain, offering further insights into the transformed image. undergo processing to yield intermediate edge-maps. These intermedi­
Given that the multi-scale retinex transformation has already improved ate edge-maps are then combined by stacking them, creating a collection
image contrast and texture, the QHWT is employed to extract frequency- of learned filters, which are ultimately merged to generate a unified
based features from these enhanced images. Apply the QHWT at various edge-map.
scales to capture both fine and coarse image details. Now, combine the Each sub-block comprises two convolutional layers, both featuring
outcomes of both methods. Consider either a weighted or linear com­ 3×3 kernels. Following each convolution operation, a batch normali­
bination of the enhanced image generated by the multi-scale retinex zation and rectified linear unit activation function are applied. Notably,
transformation and the wavelet coefficients obtained through the starting from block 3 onward, the last convolutional layer within the
QHWT. Eqn. (6) provides the formula for the integrated and recon­ final sub-block ignores the rectifier linear unit function. Red rectangles
structed image, expressing this combination process. denote the presence of max-pooling operators characterized by 3×3
kernels and a stride of 2.
PI = α ∗ X[u, v] + β ∗ [h|p] (6)
To combat the issue of significant edge features vanishing during the
where α and β are weighting coefficients that determine the influence of processing pipeline, the architecture introduces parallel skip-
each method on the result. Adjust these coefficients according to the connections. Starting from block-3 and onwards, the output of each
unique characteristics of the dataset and the desired outcome. This sub-block undergoes averaging with an additional skip-connection
adjustment (P1 ) represents the combined image that incorporates both referred to as “second skip-connections” (SSC). This averaged output
the contrast-enhanced features from HRAT and the frequency-based is then merged with the output of the first skip-connection (FSC).
information from QHWT. The collected data undergoes preprocessing Concurrently, the output from max-pooling layers is directly passed on
using the hybrid method and is subsequently segmented using the to the following sub-blocks.
DIMNet approach. The functions of the segmentation process are
described as follows. • Upsampling Network (USNet)

USNet is composed of two conditional blocks, each featuring a

3.2. Segmenting to separate the iris region sequence comprising a convolutional layer and a deconvolutional layer
designed for up-sampling features. Block-2 exclusively activates to scale
Accurate segmentation is vital for isolating the iris region from the

4
S.S. Salve and S.P. Narote Journal of Visual Communication and Image Representation 103 (2024) 104262

Fig. 2. DIMNet architecture.

input feature-maps derived from the Dense inception network, and this to extract features comprehensively from the segmented iris image. The
process continues iteratively until the feature-map size ranges twice that input layer takes the raw pixel values of an image. The first hidden layer
of the ground truth. Upon reaching this condition, the feature-map is of DMELA is responsible for processing the input data and extracting
directed to block-1. Block-1 commences processing with a 1x1 kernel, low-level features. It employs a set of kernels or filters that convolve
followed by a rectified linear unit activation and subsequently un­ with the input to capture various patterns and edges. The second hidden
dertakes a transpose convolution operation with a kernel size of s × s. s layer builds upon the features extracted in the first hidden layer. It
corresponds to the scale level of the input feature-map. It is significant to combines the low-level features to form higher-level representations.
note that the last convolutional layer within block-1 does not employ an The output layer synthesizes the features obtained from the second
activation function. kernal layer into a final feature representation.
DIMNet able to detect edges and highlight boundaries ensures that DMELA excels at reducing data dimensionality while capturing
the iris region is precisely isolated from the rest of the eye. Iris images distinctive iris features, streamlining the computational process and
vary significantly in terms of lighting conditions, occlusions and image enhancing system runtime efficiency. In essence, after segmentation,
quality. DIMNets multi-scale processing and edge detection capabilities DMELA contributes to the iris recognition process by extracting essential
make it robust to these variations, ensuring reliable segmentation results iris features. These features serve as the basis for subsequent classifi­
across diverse image conditions. The segmented iris regions are sub­ cation, allowing the system to accurately identify and verify individuals
jected to the feature extraction process. based on the unique patterns present in the iris of the eye.

3.3. Feature extraction to optimize classifier runtime 3.4. Image interpretability classification

Feature extraction is to reduce computational complexity and Finally, the features that have been extracted are employed for the
enhance classifier runtime. DMELA specializes in extracting essential iris purpose of classification.. The methodology evaluates classifiers,
features, capture unique characteristics while optimizing runtime including CSAN, KNN, RF and SVM [25]. This combination of neural
efficiency. network architectures is renowned for its accuracy and interpretability
DMELA operates as a feature extraction method specifically tailored in image classification tasks. Fig. 4 represents the architecture of CSAN
to expand the efficiency of the IR system [24]. The main objective is to classification method.
reduce the dimensionality of the iris data while capturing meaningful The CNN component of the model processes input images through
and distinctive features. The feature extraction transforms the complex multiple convolutional layers. These layers automatically extract hier­
iris patterns into a more compact and informative representation that archical features from the images, creating a feature map hierarchy.
efficiently utilized by subsequent stages, such as classification. Fig. 3 Sheaf attention network operate on the feature maps produced by the
represents the process of DMELA approach. CNN. For each feature map, Sheaf attention network compute attention
This method leverages a double-layer architecture, which enables it scores that highlight the importance of different regions and channels.

Fig. 3. Process of DMELA.

5
S.S. Salve and S.P. Narote Journal of Visual Communication and Image Representation 103 (2024) 104262

Fig. 4. CSAN architecture.

This attention mechanism helps the model focus on relevant features SANs to refine these features before making classification decisions. The
and suppresses distractions. The feature maps refined by Sheaf attention process is summarized as follows:
network are then combined or fused to create a comprehensive repre­
sentation of the image that captures both spatial and channel-wise re­ • The CNN extracts feature from the input image, producing a set of
lationships among features. The combined features are used as input to a feature maps Xl at different layers.
classification layer, softmax layer or typically a fully connected layer. • These feature maps Xl are used as the values for the sheaf attention
This final layer allocates the image to a specific class or category based mechanism. Query vectors Qi and key vectors Kj are computed for
on the extracted and refined features. each feature map.
• The sheaf attention mechanism computes attention scores Sij be­
• Convolutional neural network and sheaf attention-based tween feature maps, highlighting their relationships.
network: • The weighted sum of feature maps based on attention scores is used
as the refined representation of the input image.
CNN automatically learning meaningful features from raw image
data through a series of convolutional layers. CNNs are known for their The integration of CNNs and sheaf attention networks enhances the
ability to capture hierarchical patterns, making them well-suited for ability of method to detect complex relationships and patterns in the
image recognition tasks. Eqn. (7) expresses the output of a convolutional image, resulting in improved image classification accuracy.
layer Xl+1 :
• K-Nearest Neighbors:
Xl+1 = σ (Wl ∗ Xl + bl ) (7)

where Xl denotes the input feature map of layer l; Wl signifies the weight KNN is an intuitive and simple classification procedure. It sorts data
matrix; bl signifies the bias vector and σ denotes the activation function. points by equating them to the K closest data points within the training
Sheaf attention networks enhance spatial and channel-wise relation­ set, employing a chosen distance measurement. The KNN algorithm
ships by modeling interactions between feature maps. Eqn. (8) is the comprises two primary phases. During the training phase, it stores
mathematical expression for sheaf attention. feature vectors and assigns class labels to the training samples. During
the classification phase, when a test sample is provided, it assigns a label
eQi ⋅Kj based on the most frequent label among the K training samples that are
Sij = ∑ Qi ⋅Kj (8)
je nearest to the test sample. Eqn. (9) is employed to calculate the distance
parameter used in the KNN algorithm.
where Sij denotes the attention score between the ith and jth feature map; m
∑
Qi denotes the query vector; Kj denotes the key vector for the feature d(u, v) = |vj − uj| (9)
map; eQi ⋅Kj represents the exponential of query and key vectors. The j=1

attention scores are used to compute weighted sums of the values of

feature maps, forming context-aware representations. where u and v represents the samples; j represents the iterations and m
represents the maximum limit of training samples. KNN is easy to un­
• Integration of CNN with sheaf attention network: derstand, requires minimal parameter tuning and effective for both
small and large datasets. KNN offers simplicity and effectiveness, espe­
The integration of CNNs with SANs for feature classification involves cially in scenarios where real-time decisions are required, as it classifies
using CNNs to extract features from input images and then applying based on the nearest previously analyzed features.

6
S.S. Salve and S.P. Narote Journal of Visual Communication and Image Representation 103 (2024) 104262

• Random Forest: set.

RF is a machine learning technique that utilizes ensemble learning. It 4.1. Dataset description
combines multiple decision trees to make predictions. Each separate
decision tree is trained using a randomly chosen subset of both the data The MMU V1 dataset, sourced from Multimedia University, com­
and the available features. Their outcomes are then averaged or voted prises 450 images collected from 45 individuals, resulting in 90 different
upon to arrive at the final prediction. In RF, a training set with M cases classes. These images were captured using a specialized iris scanning
are randomly sampled to create a new training set. This set is used to sensor. The MRL Eye dataset includes images collected from 37 in­
grow decision trees. If there are N variables, a subset with n << N dividuals, predominantly men, with 33 men and 4 women represented.
variables are selected for node splitting. The trees are grown without The CASIA dataset emphasizes iris samples captured under infrared
pruning. Each tree provides its classification for a given input vector and illumination, featuring 108 unique iris categories. Each category in­
this is considered as voting. The forest selects the class with the most cludes 7 images of the same iris taken during two distinct imaging ses­
votes. During the training process, estimating the classification error sions, with each image having a resolution of 280 × 320 pixels. Table 2
involves sampling feature vectors and leaving some out, typically in a provides specifications for the datasets utilized in the study.
subset of about M/3 size. The calculation of the classification error
follows these steps:
A prediction is generated for each vector with respect to the tree. The 4.2. Performance evaluation of proposed approach
class with the most votes among the vectors is compared to the actual
ground-truth response. The classification error is determined by the For experimentation, several datasets were utilized with specific
ratio of misclassified vectors to all vectors in the original dataset. The image counts and characteristics. These datasets include 600 JPEG iris
overall forest error depends on two crucial factors: images from the CASIA V4 database, 500 JPEG iris images from the
CASIA V3 database, 150 BMP iris images from the MRL dataset, and 45
• The level of correlation between any two trees in the forest. An in­ BMP format images from the MMU V1 database. The images from CASIA
crease in correlation results in a higher error rate. databases have a resolution of 320 × 280 pixels, while those from the
• The individual predictive power of each tree in the forest. Higher MRL dataset have a resolution of 320 × 240 pixels. The MMU V1 dataset
predictive power of each tree tends to decrease the error rate. consists of images with a resolution of 320 × 280 pixels. To evaluate
precision, false negative and true positive matches were calculated by
RF provides robustness and accuracy, employing multiple decision dividing the number of matches by the total number of picture samples
trees to reduce the risk of overfitting and ensuring reliable predictions in each dataset. Table 3 presents the classification accuracy results
even with substantial input data variability. achieved using various methods.
The methods listed include RF, KNN, SVM and CSAN. These methods
• Support Vector Machine Classifier: achieved high accuracy percentages in their respective classification
tasks, with these CSAN performs more better than other methods,
SVM is a robust and versatile classification algorithm. It operates by achieving an accuracy of 99.89 %. These accuracy values indicate the
identifying a hyperplane that effectively separates data points belonging effectiveness of these methods in appropriately classifying and recog­
to distinct classes, with an emphasis on maximizing the margin between nizing iris patterns, a crucial aspect of iris recognition systems. Table 4
these classes. SVM takes a dataset as input and predicts the class to represents the performance evaluation metrics for iris recognition
which each input belongs. The input data is considered as an x-dimen­ without segmentation using proposed classification methods.
sional vector, and SVM builds a model by creating a set of hyperplanes in The RF method achieves an accuracy of 95.12 %, KNN reaches 96.34
a high-dimensional space. For an x-dimensional vector, several hyper­ %, SVM also attains 96.34 %, and the CSAN method leads with an ac­
planes are produced. Effective separation is attained by determining the curacy of 97.89 %. These metrics collectively provide a comprehensive
hyperplane that maximizes the distance to the closest training data assessment of the iris recognition performance for each method. Table 5
points from any class, a measure referred to as the functional margin. A provides a comprehensive evaluation of iris recognition performance
larger margin corresponds to improved classification performance. SVM with segmentation using proposed classification methods. Segmentation
is effective in high-dimensional spaces, capable of generating optimal significantly improves iris recognition methods by isolating the iris,
hyperplanes that distinctly categorize different classes with a high de­ enhancing feature extraction and classification. CSAN achieves 99.89 %
gree of accuracy. accuracy, demonstrating effective iris pattern recognition with superior
CNN with sheaf attention networks excels in complex image classi­ precision and recall. Methods show strong ROC values, indicating reli­
fication tasks, while KNN is simple and suitable for small datasets. SVM able discrimination between positive and negative cases, but FAR and
is versatile and effective in high-dimensional spaces, and RF is known for FRR should be minimized for better performance in authentication.
its accuracy and robustness. Segmentation significantly improves iris recognition methods by
isolating the iris, enhancing feature extraction and classification. CSAN
achieves 99.89 % accuracy, demonstrating effective iris pattern
4. Results

The performance of the proposed systems is estimated by comparing Table 2

them to preceding literature based on recognition accuracy. The iris Specification of the used dataset.
image serves as the input for feature extraction, and the classification Dataset MMU iris MRL eye CASIA V3 iris CASIA V4 iris
process is carried out using CSAN. Testing is conducted on four publicly dataset dataset dataset dataset

available datasets: CASIA-Iris-Interval V3, MMU, CASIA-Iris-Interval Number of 90 37 249 60

V4and MRL eye dataset. The entire Iris Recognition scheme proposed subjects
Samples per 45 7 10 7
is implemented using MATLAB. The implementation is takes place on a
subject
PC featuring a 3.40 GHz Intel Core i7-6700 CPU and 16 GB of RAM. The Number of 450 150 500 600
iris classification process is implemented using a Matlab script within images
the Matlab R2021b environment. For dataset preparation, the databases Image size (320 x 480) (320 x 240) (320 x 280) (320 x 280)
are randomly split into two groups: 80 % of the database was allocated (pixels)
Image format JPEG JPEG JPEG JPEG
as the training set, 20 % of the entire database was set aside as the test

7
S.S. Salve and S.P. Narote Journal of Visual Communication and Image Representation 103 (2024) 104262

Table 3 dataset, the obtained ROC value is 99.43 %, reflecting the high accuracy
Accuracy for proposed classification methods. and discrimination ability of the classification model. A ROC value close
Methods Accuracy (%) to 100 % suggests an excellent ability to distinguish between negative
and positive cases. In the case of the MRL eye dataset, the obtained ROC
RF 99.12
KNN 99.34 value is 98.91 %, indicating strong discriminatory power and model
SVM 99.34 performance. Similarly, for the CASIA V3 and V4 dataset, the obtained
CSAN 99.89 ROC value is 99.12 % and 99.34 % signifying a robust and accurate
classification model.
Figs. 8–11 represents the accuracy (a) and loss (b) obtained by the
recognition with superior precision and recall. Methods show strong
trained CSAN method for proposed datasets. It provides valuable in­
ROC values, indicating reliable discrimination between positive and
sights into the performance and optimization of the CSAN method on
negative cases, but FAR and FRR should be minimized for better per­
each dataset, helps to measure its effectiveness and fine-tune the model
formance in authentication.
as needed. Accuracy indicates the model’s ability to make correct pre­
Table 6 presents the success rates for various classification methods
dictions, while loss reflects the convergence and training process, with
across different numbers of features per image. All methods achieve
exceptionally high accuracy, with RF at 99.12 %, KNN at 99.34 %, SVM
also at 99.34 %, and the CSAN method leading with an accuracy of
Table 6
99.89 %. The CSAN method stands out with the highest accuracy, pre­
Success rate for proposed classification methods.
cision, and ROC value while maintaining low FAR and FRR rates.
As the number of features per image decreases, there is a variation in Number of features per image Success Rate

success rates across the classification methods. Notably, when 50 fea­ RF (%) SVM (%) KNN (%) CSAN (%)
tures per image are used, RF, KNN, and CSAN achieve similar high 50 99.15 97.34 99.12 99.15
success rates, around 99.12 % to 99.15 %. As the number of features 40 98.24 97.65 99.45 99.56
decreases, some methods, like KNN, RF and SVM, maintain high success 30 95.98 97.12 98.67 92.65
rates like CSAN experience a slight drop-in success rate. 20 87.96 95.56 94.27 90.36
10 82.81 80.34 80.98 87.89
Fig. 5 represents the performance of proposed classification models
using two key metrics: The Genuine Accept Rate (GAR) and the False
Non-Match Rate (FNMR). The GAR indicates the system’s ability to
accurately identify legitimate users, while the FNMR reflects the rate at
which the system incorrectly fails to recognize genuine users. The CSAN
model showcases exceptional performance with a consistent GAR sur­
passing 99 %. This outstanding GAR highlights the CSAN model’s pro­
ficiency in accurately identifying genuine users, thereby ensuring a
robust and dependable authentication process. The FNMR remains
remarkably low, reaffirming the CSAN model’s ability to minimize in­
stances of failing to recognize legitimate users.
Fig. 6 represents the relationship between the False Match Rate
(FMR) and FNMR for proposed classification models. CSAN model
achieves an exceptionally low FMR of 0.001 %. This outcome un­
derscores the model’s remarkable precision in distinguishing genuine
users from imitators, underscoring its efficacy in ensuring a secure and
dependable authentication process.
Fig. 7 illustrates Receiver Operating Characteristic (ROC) curves for
four distinct datasets: (a) MMU iris dataset, (b) MRL eye dataset (c)
CASIA V3 iris database and (d) CASIA V4 iris database. For the MMU iris Fig. 5. Genuine accept rate of proposed classification methods.

Table 4
Performance evaluation of iris recognition without segmentation.
Methods Performance evaluation

Accuracy(%) Precision(%) Recall (%) F1-Score (%) ROC (%) FAR (%) FRR (%)

RF 95.12 94.42 93.12 95.67 0.8712 73.82 2.145

KNN 96.34 95.69 95.35 94.35 0.7812 74.12 2.167
SVM 96.34 94.55 94.12 96.23 0.7841 73.87 1.623
CSAN 97.89 97.12 94.87 96.35 0.8943 74.56 1.512

Table 5
Performance evaluation of iris recognition with segmentation.
Methods Performance evaluation

Accuracy (%) Precision (%) Recall (%) F1-Score (%) ROC (%) FAR (%) FRR
(%)

RF 99.12 98.42 98.12 97.67 0.9712 74.82 2.1

KNN 99.34 98.69 97.35 98.35 0.9812 74.12 2.1
SVM 99.34 98.55 98.12 97.23 0.9841 74.87 1.6
CSAN 99.89 99.12 98.87 98.35 0.9943 75 1.5

8
S.S. Salve and S.P. Narote Journal of Visual Communication and Image Representation 103 (2024) 104262

Across the three datasets, the accuracy levels are high, ranging from
96.11 % to 97.55 %. Precision values are consistently high, ranging from
95.27 % to 95.59 %, indicating the ability to classify positive instances
accurately. The recall percentages, measuring the proportion of actual
positive cases correctly identified, range from 96.12 % to 96.83 %. F1-
scores, which provide a balance between precision and recall, also
exhibit strong performance, ranging from 97.12 % to 97.67 %. Table 8
presents the results obtained for the proposed dataset with the seg­
mentation process applied.
Across the three datasets, the accuracy levels are especially high,
ranging from 99.35 % to 99.98 %. Precision values indicate the ability to
accurately classify positive instances and range from 98.54 % to 99.12
%. Recall percentages, range from 98.27 % to 98.87 %. F1-scores,
ranging from 98.15 % to 98.25 %. The time taken for evaluation is
also provided, indicating the efficiency of the method in processing the
data.
Fig. 6. False match rate of proposed classification methods. The results demonstrate a significant improvement in accuracy and
precision when segmentation is applied to the dataset. The time taken
for evaluation is also provided, indicating the efficiency of the method
lower loss values signifying improved model performance.
with segmentation in processing the data. These outcomes underscore
A confusion matrix summarizes a classification model’s performance
the effectiveness of the segmentation process in enhancing the perfor­
by comparing predicted and actual labels, aiding model evaluation and
mance of the proposed dataset.
metric selection in machine learning.
Fig. 12 represents the confusion matrices for the proposed datasets,
specifically for (a) the MMU iris dataset (b) the MRL eye dataset (c) the 4.3. Statistical analysis
CASIA v3 iris dataset and (d) the CASIA V4 iris dataset. Analyzing these
confusion matrices helps in understanding the strengths and weaknesses Statistical analysis is fundamental for assessing and interpreting the
of the classification models used for iris recognition in each dataset, performance of iris segmentation and recognition methods, providing
allows for further refinement and optimization of the system. Table 7 insights into the significance of observed differences across datasets.
provides the obtained results for the proposed dataset without the seg­ Measures such as p-values are essential, indicating the likelihood of
mentation process. observing results under the null hypothesis of no difference between
datasets. A low p-value (typically < 0.05) suggests a statistically

Fig. 7. ROC for (a) MMU iris dataset, (b) MRL eye dataset, (c) CASIA V3 iris dataset and (d) CASIA V4 iris dataset.

9
S.S. Salve and S.P. Narote Journal of Visual Communication and Image Representation 103 (2024) 104262

Fig. 8. (a) Accuracy and (b) loss obtained by the trained CSAN method on the MMU iris dataset.

Fig. 9. (a) Accuracy and (b) loss obtained by the trained CSAN method on the MRL eye dataset.

Fig. 10. (a) Accuracy and (b) loss obtained by the trained CSAN method on the CASIA V3 iris dataset.

significant difference, elucidating which method or dataset excels in 4.4. Comparitive analysis with exixting approaches
segmentation and classification metrics. Table 9 showcases the results of
the DIMNet segmentation method evaluated across multiple datasets Table 10 represents a comprehensive comparison of existing
using performance metrics including F1-score, accuracy, and recall. methods and their obtained accuracy in iris recognition tasks. The
The ANOVA statistical test assesses whether there are significant proposed CSAN method, when compare with other classification
differences in DIMNet’s segmentation performance across these data­ methods, achieves high accuracy across multiple datasets and obtains up
sets. This analysis aids in understanding which dataset influences the to 99.98 %. This showcases the efficiency of the proposed technique in
segmentation outcomes and highlights the efficacy of the DIMNet iris recognition and positions it as a competitive approach in the field.
method in varying contexts.

10
S.S. Salve and S.P. Narote Journal of Visual Communication and Image Representation 103 (2024) 104262

Fig. 11. (a) Accuracy and (b) loss obtained by the trained CSAN method on the CASIA V4 iris dataset.

Fig. 12. Confusion matrix for(a) the MMU iris dataset (b) the MRL eye dataset (c) the CASIA v3 iris dataset and (d) the CASIA V4 iris dataset.

4.5. Ablation study resulting in a modified network termed NSCNet. This alteration allowed
for a direct comparison of the performance between DIMNet and
To assess the effectiveness of DIMNet and CSAN within the proposed NSCNet in the context of iris segmentation tasks.
unified framework, a series of ablation experiments are conducted to Table 11 presents a comparative analysis of iris segmentation and
analyze the impact of specific architectural components. Specifically, recognition outcomes achieved by two distinct approaches: DIMNet and
the skip connection introduced in DIMNet was thoroughly evaluated, NSCNet, across multiple datasets. For segmentation performance, the
and the fully connected layer in CSAN is replaced with global average metrics include the error rate (E) and F1-score, assessing the accuracy
pooling. and precision of isolating iris regions. In terms of classification, the table
In these experiments, DIMNet was further examined by replacing its includes metrics such as the GAR at a False Match Rate (FMR) of 0.1 %
skip connections with Normal Skip Connections Network (NSCNet), and the Equal Error Rate (EER), which reflect the effectiveness of each

11
S.S. Salve and S.P. Narote Journal of Visual Communication and Image Representation 103 (2024) 104262

Table 7 Table 10
Obtained results for proposed dataset without segmentation. Comparison of existing methods with obtained accuracy.
Datasets Performance Evaluation Time References Feature Classification Datasets Obtained
(sec) extraction methods accuracy
Accuracy Precision Recall F1-score
methods
(%) (%) (%) (%)
[11] CNN CNN and IITD 94.88 %
MMU [11] 97.34 95.27 96.12 97.35 1.871
Hamming CASIA-Iris- 96.56 %
MRL [11] 97.55 95.44 96.65 97.67 2.345
Distance Interval V4,
CASIA V3 96.11 95.59 96.83 97.12 1.976
MMU 98.01 %
[16]
[13] SDD CNN CASIA 95.5 %
CASIA V4 97.06 96.57 96.72 97.07 1.876
MMU 95 %
[11]
[14] EH-WOA CNN IITD 92 %
MMU 95 %
[15] LDA DNN MMU 98.2 %
[16] DL CNN, SVM LFW 99.15 %
Table 8
FERET 98.35 %
Obtained results for proposed dataset with segmentation.
IITD 97.89 %
Datasets Performance Evaluation Time CASIA- 95.48 %
(sec) IrisV3
Accuracy Precision Recall F1-score
[17] PFFT KNN, SVM, CASIA 1.1 99.99 %
(%) (%) (%) (%)
PNN CASIA 4.0 98.26 %
MMU [11] 99.98 99.12 98.35 97.56 0.875 SCCSIE&T 99.25 %
MRL [11] 99.35 98.78 98.87 98.25 1.890 VISA Iris 96.67 %
CASIA V3 99.45 98.54 98.27 98.15 1.654 [18] MCFFN CNN CASIA-Iris- 99.55 %
[16] Syn
CASIA V4 99.65 98.25 98.27 98.15 1.625 CASIA-Iris- 98.32 %
[11] Lamp
[19] OCFON Deep learning CASIA-Iris- 99.94 %
methods Twins
CASIA-Iris- 99.57 %
Table 9 Lamp
Segmentation method using statistical methods. [20] − MTL CASIA-Iris- 99.95 %
Lamp
Method Datasets Performance metrices Proposedmethods DMELA CSAN, SVM, MMU 99.98 %
F1-score Accuracy Recall KNN, RF MRL 99.35 %
CASIA V3 99.45 %
DIMNet MMU 0.048721 0.04726 0.03672 CASIA V4 99.65 %
MRL 0.041243 0.03672 0.04261
CASIA V3 0.039821 0.04261 0.04726
CASIA V4 0.038232 0.04321 0.04321
Table 11
Iris segmentation and recognition results obtained through various approaches
method in accurately identifying and classifying iris patterns. This su­ are compared.
perior performance is largely attributed to the unique skip connection
Methods Datasets Segmentation Classification
strategy employed within DIMNet, which effectively compensates for
both semantic and spatial information deficits associated with low-level E (%) F1-score (%) GAR(FMR=0.1) EER (%)
(%)
and high-level features. This compensation enhances feature fusion
within the network, thereby enabling more accurate differentiation DIMNet MMU 0.43 98.38 99.01 0.27
MRL 0.34 97.65 99.26 0.06
between iris and non-iris regions. The skip connection mechanism in
CASIA V3 1.02 99.42 99.53 0.43
DIMNet plays a crucial role in information propagation across different CASIA V4 1.06 98.76 99.17 0.32
network layers, facilitating improved feature learning and representa­ NSCNet MMU 0.43 97.32 98.21 0.37
tion. By addressing missing semantic and spatial cues inherent in MRL 0.36 96.34 98.25 0.18
traditional skip connections, DIMNet achieves enhanced performance in CASIA V3 1.03 98.43 99.05 0.56
CASIA V4 1.05 97.26 98.87 0.47
iris segmentation tasks, underscoring its effectiveness within the pro­
posed framework.
5. Conclusion
4.6. Discussion
This research presents a comprehensive framework for improving
The comprehensive evaluation of iris recognition methods presented iris recognition accuracy and efficiency, addressing challenges in seg­
in this study underscores the effectiveness of various classification ap­ mentation and classification. Proposed CSAN classification method has
proaches across multiple datasets. High accuracy percentages achieved demonstrated better accuracy, precision and recall, making it a prom­
by RF, KNN, SVM, and CSAN demonstrate their robustness in classifying ising choice for real-world applications that demand robust security and
iris patterns accurately. Segmentation significantly enhances recogni­ reliable user identification. The experimental study was conducted for
tion methods by isolating the iris, improving feature extraction, and both segmented and non-segmented classification processes, and it was
classification accuracy. CSAN emerges as the top performer with 99.89 observed that segmentation significantly outperforms the non-
% accuracy, demonstrating superior precision and recall. ROC values segmented approach in terms of recognition accuracy. Limitations of
indicate strong discrimination ability, although efforts are needed to the current work include the reliance on traditional segmentation
minimize FAR and FRR for optimal authentication performance. Sta­ techniques, which may be less robust for complex iris patterns. Inves­
tistical analyses, including ANOVA, provide insights into segmentation tigating more advanced and robust segmentation techniques could
performance variations across datasets, guiding methodological further improve the accuracy and speed of iris recognition. Exploring
improvements. deep learning-based segmentation methods may yield even better
results.

12
S.S. Salve and S.P. Narote Journal of Visual Communication and Image Representation 103 (2024) 104262

Funding [12] S.N. Khan, S.U. Khan, O.J. Nwobodo, K. Adam, Cyran, Iris recognition through
edge detection methods: application in flight simulator user identification, Int. J.
This research did not receive any specific grant from funding
Adv. Comput. Sci. Appl. 14 (2023), https://doi.org/10.14569/
agencies in the public, commercial, or not-for-profit sectors. ijacsa.2023.0140425.
[13] E. Hussein Ali, H. Abbas Jaber, N. Naji Kadhim, New algorithm for localization of
CRediT authorship contribution statement iris recognition using Deep Learning Neural Networks, Indonesian J. Electr. Eng.
Comput. Sci. 29 (2022) 110, https://doi.org/10.11591/ijeecs.v29.i1.pp110-119.
[14] G. Babu, P.A. Khayum, Elephant herding with whale optimization enabled orb
Sushilkumar S. Salve: Conceptualization. Sandipann P Narote: features and CNN for Iris recognition, Multimed. Tools Appl. 81 (2021)
Supervision. 5761–5794, https://doi.org/10.1007/s11042-021-11746-7.
[15] R.A. Abdulhasan, S.T. Abd Al-latief, S.M. Kadhim, Instant learning based on deep
neural network with linear discriminant analysis features extraction for accurate
Declaration of competing interest iris recognition system, Multimed. Tools Appl. (2023), https://doi.org/10.1007/
s11042-023-16751-6.
[16] E. Abdellatef, R.F. Soliman, E.M. Omran, N.A. Ismail, S.E. Elrahman, K.N. Ismail, et
The authors declare that they have no known competing financial al., Cancelable face and Iris recognition system based on Deep Learning, Optical
interests or personal relationships that could have appeared to influence Quantum Electron. 54 (2022), https://doi.org/10.1007/s11082-022-03770-0.
the work reported in this paper. [17] V.C. Kagawade, S.A. Angadi, A new scheme of Polar Fast Fourier transform code
for iris recognition through symbolic modelling approach, Exp. Syst. Appl. 197
(2022) 116745, https://doi.org/10.1016/j.eswa.2022.116745.
Data availability [18] J. Sun, S. Zhao, Y. Yu, X. Wang, L. Zhou, Iris recognition based on local circular
Gabor filters and multi-scale Convolution Feature Fusion Network, Multimed.
Tools Appl. 81 (2022) 33051–33065, https://doi.org/10.1007/s11042-022-13098-
Data sharing not applicable to this article as no datasets were
2.
generated or analyzed during the current study. [19] J. Sun, S. Zhao, S. Miao, X. Wang, Y. Yu, Open-set iris recognition based on Deep
Learning, IET Image Proc. 16 (2022) 2361–2372, https://doi.org/10.1049/
Acknowledgement ipr2.12493.
[20] S. Lei, B. Dong, A. Shan, Y. Li, W. Zhang, F. Xiao, Attention meta-transfer learning
approach for few-shot Iris recognition, Comput. Electr. Eng. 99 (2022) 107848,
None. https://doi.org/10.1016/j.compeleceng.2022.107848.
[21] J. Zhou, J. Yao, W. Zhang, D. Zhang, Multi-scale retinex-based adaptive gray-scale
transformation method for underwater image enhancement, Multimed. Tools Appl.
References 81 (2021) 1811–1831, https://doi.org/10.1007/s11042-021-11327-8.
[22] F. Bulut, Low dynamic range histogram equalization (LDR-HE) via quantized Haar
[1] K. Nguyen, H. Proença, F. Alonso-Fernandez, Deep Learning for Iris Recognition: A wavelet transform, Visual Comput. 38 (2021) 2239–2255, https://doi.org/
Survey. arXiv Preprint arXiv:2210.05866. (2022). 10.1007/s00371-021-02281-5.
[2] J. Wei, H. Huang, Y. Wang, R. He, Z. Sun, Towards more discriminative and robust [23] X. Soria, A. Sappa, P. Humanante, A. Akbarinia, Dense extreme inception network
iris recognition by learning uncertain factors, IEEE Trans. Inf. Forensics Secur. 17 for edge detection, Pattern Recogn. 139 (2023) 109461, https://doi.org/10.1016/
(2022) 865–879, https://doi.org/10.1109/tifs.2022.3154240. j.patcog.2023.109461.
[3] S. Jamaludin, A.F. Ayob, M.F. Akhbar, A.A. Ali, M.M. Imran, S.M. Norzeli, et al., [24] Y. Zhou, B. Sun, W. Sun, Z. Lei, Tool wear condition monitoring based on a two-
Efficient, accurate and fast pupil segmentation for pupillary boundary in Iris layer angle kernel extreme learning machine using sound sensor for milling
recognition, Adv. Eng. Softw. 175 (2023) 103352, https://doi.org/10.1016/j. process, J. Intell. Manuf. 33 (2020) 247–258, https://doi.org/10.1007/s10845-
advengsoft.2022.103352. 020-01663-1.
[4] B. Karthik, G. Ramkumar, Comparison of feature extraction technique for [25] M.A. Mohammed, A. Lakhan, K.H. Abdulkareem, B. Garcia-Zapirain, A hybrid
segmentation in human iris recognition under uncontrolled environment using cancer prediction based on multi-omics data and Reinforcement Learning State
CNN algorithm with SVM classifier, ECS Trans. 107 (2022) 16785–16795, https:// Action Reward State Action (Sarsa), Comput. Biol. Med. 154 (2023) 106617,
doi.org/10.1149/10701.16785ecst. https://doi.org/10.1016/j.compbiomed.2023.106617.
[5] A. Boyd, S. Yadav, T. Swearingen, A. Kuehlkamp, M. Trokielewicz, E. Benjamin, et [26] C. Yan, B. Gong, Y. Wei, Y. Gao, Deep multi-view enhancement hashing for image
al., Post-mortem iris recognition—a survey and assessment of the state of the art, retrieval, IEEE Trans. Pattern Anal. Mach. Intell. (2020), https://doi.org/10.1109/
IEEE Access 8 (2020) 136570–136593, https://doi.org/10.1109/ TPAMI.2020.2975798.
access.2020.3011364. [27] C. Yan, Z. Li, Y. Zhang, Y. Liu, X. Ji, Y. Zhang, Depth image denoising using nuclear
[6] N. Ahmadi, M. Nilashi, S. Samad, T.A. Rashid, H. Ahmadi, An intelligent method norm and learning graph model, ACM Trans. Multimed. Comput. Commun. Appl.
for Iris recognition using supervised machine learning techniques, Optics Laser (2020), https://doi.org/10.1145/3404374.
Technol. 120 (2019) 105701, https://doi.org/10.1016/j.optlastec.2019.105701. [28] C. Yan, Y. Hao, L. Li, J. Yin, A. Liu, Z. Mao, Z. Chen, X. Gao, Task-adaptive
[7] H. Hafeez, M.N. Zafar, C.A. Abbas, H. Elahi, M.O. Ali, Real-time human attention for image captioning, IEEE Trans. Circuits Syst. Video Technol. 32 (2021)
authentication system based on Iris recognition, Eng. 3 (2022) 693–708, https:// 42–51, https://doi.org/10.1109/TCSVT.2021.3067449.
doi.org/10.3390/eng3040047. [29] C. Yan, T. Teng, Y. Liu, Y. Zhang, H. Wang, X. Ji, Precise no-reference image
[8] B. Abd El-Rahiem, F.E. Abd El Samie, M. Amin, Efficient cancellable multi- quality evaluation based on distortion identification, ACM Trans. Multimed.
biometric recognition system based on Deep Learning and bio-hashing, Appl. Intell. Comput. Commun. Appl. 17 (2021) 1–21, https://doi.org/10.1145/3468872.
53 (2022) 1792–1806, https://doi.org/10.1007/s10489-021-03153-0. [30] C. Yan, L. Meng, L. Li, J. Zhang, Z. Wang, J. Yin, J. Zhang, Y. Sun, B. Zheng, Age-
[9] N.D. AL-Shakarchy, H.K. Obayes, Z.N. Abdullah, Person identification based on invariant face recognition by multi-feature fusion and decomposition with self-
voice biometric using deep neural network. Int. J. Inf. Technol. 15 (2022) 789–795. attention, ACM Trans. Multimed. Comput. Commun. Appl. 18 (2021) 1–8, https://
DOI: 10.1007/s41870-022-01142-1. doi.org/10.1145/3472810.
[10] N. Alay, H.H. Al-Baity, Deep Learning Approach for multimodal biometric [31] C. Yan, Y. Sun, H. Zhong, C. Zhu, Z. Zhu, B. Zheng, X. Zhou, Review of omnimedia
recognition system based on fusion of Iris, face, and finger vein traits, Sens. 20 content quality evaluation, J. Signal Process. 38 (2022) 1111–1143, https://doi.
(2020) 5523, https://doi.org/10.3390/s20195523. org/10.16798/j.issn.1003-0530.2022.06.001.
[11] R.H. Farouk, H. Mohsen, Y.M. El-Latif, A proposed biometric technique for
improving iris recognition, Int. J. Comput. Intell. Syst. 15 (2022), https://doi.org/
10.1007/s44196-022-00135-z.

13