INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue XI November 2025
Page 459
www.rsisinternational.org
Development of Iris Image Classification Framework using Multi-
Layer CNN Architecture
R. D. Bhoyar
1
, D. R. Solanke
2
, S. D. Pachpande
3
1,3
Department of Computer Science, SGB Amravati University, Amravati, India (MS)
2
Department of Applied Electronics, SGB Amravati University, Amravati, India (MS)
*Corresponding Author
DOI: https://dx.doi.org/10.51584/IJRIAS.2025.101100043
Received: 22 November 2025; Accepted: 28 November 2025; Published: 09 December 2025
ABSTRACT:
This paper proposes a multi-layer Convolutional Neural Network (CNN) framework for iris image
classification, targeting left and right eye recognition across 46 subjects. A custom five-layer CNN was
trained for 200 epochs with a learning rate of 0.0001, effectively learning discriminative features from iris
textures. The model achieved a training accuracy of 97.90% with a loss of 0.4116, and a testing accuracy of
93.09% with a loss of 0.6837, demonstrating robust generalization to unseen data. The results highlight the
potential of multi-layer CNN architectures for reliable iris-based biometric systems, enabling accurate and
automated eye classification. The key contribution of this work is the demonstration that a compact five-
layer CNN can achieve high accuracy in binary left-right iris classification, offering an efficient and scalable
solution for biometric authentication.
Keywords: Authentication, Biometric, Classification, Deep Learning, Feature Extraction, Iris Recognition.
INTRODUCTION
Iris recognition has emerged as one of the most reliable and secure biometric modalities due to the
uniqueness and stability of iris patterns across individuals. Unlike other biometric traits such as fingerprints
or facial features, the iris provides rich textural information that remains largely unchanged over time,
making it an ideal candidate for accurate identification and authentication systems. The increasing demand
for secure access control in sectors such as banking, defence, and personal devices has motivated researchers
to develop automated iris classification frameworks capable of high precision and robustness. Conventional
iris recognition methods rely on hand-crafted feature extraction techniques such as Gabor filters, wavelet
transforms, or Local Binary Patterns (LBP). While these approaches can achieve reasonable accuracy, they
often require extensive pre-processing, careful parameter tuning, and may fail to generalize effectively
across large or diverse datasets. With the advent of deep learning, Convolutional Neural Networks (CNNs)
have demonstrated remarkable capability in automatically learning hierarchical features directly from raw
image data, eliminating the need for manual feature engineering.
In this study, we propose a custom five-layer CNN architecture for iris image classification, specifically
designed to distinguish between left and right eyes across 46 subjects. The model was trained over 200
epochs with a learning rate of 0.0001, allowing it to progressively capture discriminative patterns in iris
textures. The training process yielded an accuracy of 97.90% with a corresponding loss of 0.4116, while the
model achieved a testing accuracy of 93.09% and a loss of 0.6837, indicating strong generalization
performance on unseen data. These results validate the effectiveness of multi-layer CNN architectures for
iris classification and demonstrate that even a relatively compact network can provide high accuracy in
practical biometric applications. The proposed framework not only automates the classification process but
also offers scalability and efficiency, making it suitable for deployment in real-time biometric authentication
systems. The high accuracy obtained in distinguishing left and right eyes underscores the potential of CNN-
based approaches to enhance the reliability and security of iris-based identification systems. Future
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue XI November 2025
Page 460
www.rsisinternational.org
extensions of this work may include increasing the dataset size, integrating advanced data augmentation
strategies, and exploring deeper or hybrid CNN architectures to further improve classification performance.
REVIEW OF LITERATURE
The development of an iris image classification framework using a multi-layer Convolutional Neural
Network (CNN) architecture is a promising approach in the field of biometric authentication. This method
leverages the unique and intricate patterns found in the human iris to achieve high accuracy in identification
tasks. The use of CNNs allows for effective feature extraction and classification, making them well-suited
for handling the challenges associated with iris recognition, such as variations in lighting, occlusions, and
noise
Iris recognition has established itself as a cornerstone of biometric authentication systems due to the
uniqueness, stability, and richness of iris patterns. With increasing demands for accuracy and robustness
under real-world conditions, researchers have turned to deep learning particularly Convolutional Neural
Networks (CNNs) to enhance feature extraction and classification in iris images. This review examines the
recent developments in iris image classification, with a focus on multi-layer CNN architectures, and situates
them within the broader landscape of hybrid, lightweight, and transfer learning-based models. A
foundational contribution in this area is the work by Pambudi et al. [1], who proposed a custom CNN model
with three convolutional layers using filter sizes of 32, 64, and 128. The model integrated ReLU activations,
batch normalization, max pooling, and dropout regularization, ultimately achieving a classification accuracy
of 97.33%. This architecture was specifically designed for iris images and demonstrated robustness under
varying illumination and image noise characteristics essential for real-world biometric systems.
Comparative approaches include the use of hybrid deep learning models. For instance, Almsaadi et al. [2]
introduced HDN-Net, a hybrid CNN framework that fused features from both left and right irises to
overcome occlusions from eyeglasses and image blur. This model achieved up to 98.79% accuracy on
benchmark datasets such as UBIRIS.V2 and CASIA-Iris.V4. In a similar vein, Rahman et al. [3] explored
transfer learning techniques using pre-trained networks like VGG16, VGG19, and ResNet50 to extract
features from iris images, followed by classification via multiple machine learning algorithms. Their best-
performing CNN classifier attained an accuracy of 93.40%, demonstrating the utility of deep feature transfer
in iris recognition. Other studies have focused on simplifying CNN pipelines while maintaining high
accuracy. For example, S and Mathew [4] developed a CNN model coupled with a Softmax classifier,
achieving 96% accuracy on IITD and CASIA datasets without requiring domain-specific pre-processing.
Bhatnagar et al. [5] combined CNN feature extraction with Support Vector Machines (SVM) for
classification, offering a modular architecture adaptable to other biometric tasks. A different angle is
explored by Nguyen et al. [6], who proposed the WAHET-CNN framework emphasizing precise iris
segmentation and pattern classification. Although it achieved a lower accuracy (90%) on the CASIA dataset,
the study highlighted the critical role of segmentation in improving downstream classification. In the same
year, Minaee and Abdolrashidi [7] introduced DeepIris, a residual CNN that learned both feature
representation and classification in a joint framework, especially effective with limited training samples.
Focusing on spoof detection, Yan et al. [8] developed a hierarchical CNN architecture capable of
distinguishing fake irises with nearly 100% accuracy, showcasing the potential of CNNs in security-critical
applications. Meanwhile, Pasha et al. [9] achieved an impressive 99.89% accuracy by leveraging CNNs for
iris localization and boundary detection, outperforming classical segmentation methods such as U-Net.
Research addressing specific challenges like contact lens interference includes the ContactLensIris system
by Kaur and Saini [10], which used ORB and BRISK descriptors combined with CNN layers, achieving
98% correct classification rate. Shirke et al. [11] integrated optimization techniques into a CNN-based
framework (BW-CNN), although it did not explicitly detail a multi-layer structure. The work by
Anegsiripong et al. [12] fused periocular features with iris recognition via CNNs, boosting accuracy from
93.64% to 99.80%.Further innovation comes from Liu et al. [13], who developed a compact 2-channel CNN
with radial attention and pruning strategies to optimize performance with fewer training samples. Prasad
[14] employed a CNN-Softmax pipeline with AdaGrad optimization and achieved a Rank-1 identification
rate of 99.8%, confirming the model’s utility in rapid recognition tasks. Lightweight CNNs with attention
mechanisms have also gained traction. Zou et al. [15] proposed a streamlined CNN incorporating channel-
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue XI November 2025
Page 461
www.rsisinternational.org
wise attention, improving both speed and accuracy. Kawakami et al. [16] segmented the iris into four regions
for selective feature learning and matching, enhancing recognition precision. Other studies like Menon and
Mukherjee [17] applied deep residual networks and reached 99.8% recognition rates, confirming the strength
of deep CNN architectures. In contrast, Feng et al. [18] introduced Iris R-CNN for segmentation rather than
classification, employing dual-region proposal networks for non-cooperative environments. Zambrano et al.
[19] further enhanced CNNs by mitigating aliasing and adapting circular padding, thereby improving
performance under head rotation scenarios. Sharma et al. [20] diverged from CNNs, using kernel
discriminant analysis with probabilistic neural networks to classify iris images with high robustness.
Zhou et al. [21] adopted an improved MobileNetV2 architecture to extract features efficiently in constrained
settings, while Zambrano et al. [22] proposed an unsupervised method leveraging low-level CNN layers
without training, ideal for fast matching. Finally, Minaee and Abdolrashidi [23] presented an updated version
of DeepIris with visualization modules, and Kawakami et al. [24] expanded their earlier region-based CNN
model with weighted matching and attention mechanisms.
Dataset
The Multimedia University (MMU) Iris Dataset is a publicly available collection of eye images developed
to support research and development in iris-based biometric systems, particularly for applications such as
biometric attendance. Iris patterns are unique for each individual, making them highly reliable for personal
identification. This dataset consists of a total of 460 grayscale images corresponding to 46 individuals, with
each individual having five close-up images of the left eye and five of the right eye. The dataset is organized
into 46 directories, each representing one subject, and within each directory, there are two subdirectories
labelled ‘left’ and ‘right’, containing the respective images. The structured nature of the dataset makes it
convenient for tasks such as iris segmentation, feature extraction, and classification. The MMU dataset has
been widely used for iris recognition research. Iris classification can be performed to classify the eye patterns
in two categories, which allows precise feature extraction. These features can then be used to classify an iris
image according to a stored database, enabling accurate individual identification.
Fig. 1 Iris image dataset distribution for implementation of CNN
Convolutional Neural Networks (CNNs) have been commonly employed for iris classification on this
dataset, with transfer learning from pre-trained models improving the recognition performance. Researchers
have also applied pre-processing techniques such as histogram equalization and contrast enhancement to
improve image quality and training efficiency. Several studies have reported promising results using the
MMU dataset. For present work, CNN-based models have achieved high classification accuracy by learning
discriminative features from the iris patterns, and hybrid approaches combining CNNs with other machine
learning classifiers such as support vector machines have also been explored.
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue XI November 2025
Page 462
www.rsisinternational.org
Fig. 2 Random Iris images for Training and Testing
Convolution Neural Network
In this study, a supervised deep learning approach using a custom five-layer Convolutional Neural Network
(CNN) was employed for iris image classification. The dataset consisted of 460 annotated samples from 46
subjects, comprising both left and right eye images. The network architecture included multiple 2D
convolutional and max-pooling layers with ReLU activation functions, followed by a flatten layer and a
softmax output layer with two classes: Left and Right. By training on these labelled samples, the model
automatically learned discriminative features from the iris textures, enabling accurate left-right eye
classification and demonstrating the effectiveness of a tailored CNN architecture for biometric recognition
tasks.
Fig. 3 Proposed Convolution Neural Network
Implementation of Architecture:
The proposed Convolutional Neural Network (CNN) model follows a sequential deep learning architecture
specifically designed for the binary classification of iris images into left and right categories. The network
takes grayscale or RGB iris (200, 200, 3) images as input and processes them through a series of
convolutional and pooling operations to extract hierarchical spatial features. The input image is first passed
through a 2D convolutional layer comprising 16 filters of size 3×3, generating an output feature map of
dimensions 198×198×16. This is followed by a 2×2 max-pooling layer that reduces the spatial resolution to
99×99 while maintaining the 16 feature channels, thus preserving essential features while reducing
computation. The second convolutional layer employs 32 filters, also of size 3×3, producing feature maps
of shape 97×97×32. This is again followed by a 2×2 max-pooling operation, reducing the spatial size to
48×48×32. The third convolutional layer increases the depth to 64 feature maps and outputs a 46×46×64
tensor, which is down sampled to 23×23×64 using another max-pooling layer. Continuing the depth
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue XI November 2025
Page 463
www.rsisinternational.org
expansion, the fourth convolutional layer applies 128 filters, resulting in a 21×21×128 output, followed by
max pooling to reduce it to 10×10×128.
After the final convolutional and pooling layers, the 3D output tensor is flattened into a 1D vector of 12,800
units. This flattened vector is fed into a fully connected (dense) layer with 256 neurons, introducing non-
linearity and learning complex, high-level features. The final dense layer contains 2 output units
corresponding to the two target classes: left and right iris. A softmax activation function is used in the output
layer to convert the logits into probability scores, enabling the model to classify each input image into one
of the two categories. Throughout the network, ReLU (Rectified Linear Unit) is used as the activation
function for all convolutional and dense layers (except the final output), which ensures faster convergence
and mitigates vanishing gradient issues. The model consists of a total of 3,375,010 trainable parameters,
distributed across convolutional filters and dense layers, and has no non-trainable parameters, making it a
fully learnable system. The total memory footprint is approximately 12.87 MB, making it computationally
efficient for mid-scale biometric applications. This layered architecture enables the network to progressively
extract both low-level and high-level discriminative features from the iris patterns. The inclusion of multiple
convolutional layers facilitates robust feature learning, while the max-pooling layers help in reducing spatial
redundancy and overfitting. This design is particularly effective for biometric classification tasks like iris
recognition, where subtle differences in texture, shape, and radial patterns need to be captured and
distinguished across subjects.
Model Evaluation:
Let the dataset be
(1)
In k-fold cross validation, D is split into disjoint subset folds (k)
For i
th
fold
(3)
Let the model trained in the i-th fold produce a performance metric is the average (accuracy and loss etc.) M
(i)
. Then, after training and evaluating across all k folds, the overall i-fold performance is computed as the
average:
(4)
This equation formalizes that the final performance metric is the average of the metric from all k folds,
providing a robust estimate of the models generalization
Let
denotes the predicted label for sample x
j
in fold i, and y
i
the true label. Then accuracy for fold i is:
(5)
Where 1 is an indicator function that equals 1 of the condition is true, and 0 otherwise.
The overall k-fold accuracy is the average of the accuracy across folds:
(6)
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue XI November 2025
Page 464
www.rsisinternational.org
If
is the predicted probability distribution over C classes for sample in
fold,
and
is the true class label (one-hot encoded as
), the cross-entropy loss for fold is:
(7)
The overall k-fold cross-entropy (Log loss) is the average over folds:
RESULTS AND DISCUSSION
The proposed multi-layer CNN-based iris image classification framework was evaluated on a labelled
dataset using standard performance metrics including accuracy, loss, precision, recall, F1-score, and
confusion matrix analysis. The network architecture, comprising three convolutional layers with increasing
filter sizes (32, 64, and 128), ReLU activation, max pooling, and dropout, was trained over 100 epochs.
Fig. 4 Training and validation curves for loss (left) and accuracy (right) over 100 epochs.
Fig. 1 illustrates the training and validation loss and accuracy curves. The training loss decreased rapidly
and stabilized after approximately 10 epochs, with a final average training loss of 0.4116 and validation loss
of 0.6837. The model’s ability to generalize is evident from the minimal gap between training and validation
loss beyond epoch 20, suggesting effective mitigation of overfitting through dropout and batch
normalization techniques. Concurrently, the training accuracy reached an average of 97.89%, while the
validation accuracy stabilized at 93.09%, as shown in Fig. 1 (right). This high performance indicates the
model’s capability to learn robust feature representations of the iris even in the presence of noise, occlusion,
and intra-class variability.
Table 1 Final training and testing metrics showing average accuracy and loss values.
The classification report in Table. 2 further confirms the model's reliability. For binary class labels (0 and
1), both classes achieved an identical F1-score of 0.93, although they exhibited inverse trade-offs between
precision and recall. Class 0 (possibly representing normal/ideal iris images) attained a precision of 0.88 and
a recall of 0.98, while class 1 (noisy or occluded images) achieved a precision of 0.98 and a recall of 0.87.
This behavior suggests the model is highly sensitive to class “Left”, possibly due to better feature
consistency in those images.
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue XI November 2025
Page 465
www.rsisinternational.org
Table 2 Classification report displaying precision, recall, and F1-scores.
The confusion matrix in Fig. 4 substantiates this observation, showing 63 true positives and 64 true
negatives, with 8 false negatives and 0 false positives. These results indicate that the classifier exhibits high
specificity, avoiding false alarms, but has slightly reduced sensitivity in detecting class “Right” images.
Fig. 5 Confusion matrix illustrating actual vs. predicted class distributions.
The framework achieved a macro-average accuracy of 93%, with macro and weighted F1-scores also
standing at 0.93, demonstrating balanced classification performance across both classes. Compared to
traditional or transfer learning-based methods, the proposed dedicated multi-layer CNN shows superior
learning efficiency and better adaptation to domain-specific features, eliminating dependency on pre-trained
external networks.
Numerical data and graphical representation shown above indicating for quantitative analysis purpose but it
will be nice if results are shown ( fig. 3) in terms of real time images for better understanding that how our
model perform on custom image data. The following are the images actually predicted by our proposed deep
learning model.
Fig. 6 The images actually predicted
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue XI November 2025
Page 466
www.rsisinternational.org
CONCLUSION
In this study, a custom five-layer Convolutional Neural Network (CNN) was developed for iris image
classification to distinguish left and right eyes across 46 subjects. The model demonstrated strong learning
capability, achieving a training accuracy of 97.90% and a testing accuracy of 93.09%, with corresponding
losses of 0.4116 and 0.6837. These results confirm that the proposed multi-layer CNN framework effectively
captures the discriminative features of iris textures and generalizes well to unseen data. The study highlights
the potential of compact CNN architectures for reliable and efficient iris-based biometric systems, providing
a practical solution for automated eye classification and authentication applications. Future work could focus
on expanding the dataset and incorporating additional pre-processing or augmentation techniques to further
enhance accuracy and robustness.
REFERENCES
1. D. Pambudi, F. Fadly, M. H. Kurniawan, and H. Haryanto, “The Eye’s Signature: Innovative
Approaches to Iris Detection,” International Journal of Advanced Artificial Intelligence and
Machine Learning, vol. 2, no. 1, pp. 38–43, 2025.
2. J. R. M. Almsaadi, N. Manshor, L. S. Affendey, A. A. Halin, and M. R. Mirza, “HDN-Net: A
Hybrid Deep Neural Network to Improve Iris Recognition in Unconstrained Environments with
Eyeglasses,” Journal of Advanced Research in Applied Sciences and Engineering Technology, vol.
54, no. 2, pp. 238–251, 2024.
3. M. A. Rahman, L. E. Ali, S. Mistry, A. Mukherjee, and Md. Z. Islam, “Distant Iris Recognition
Through Machine Learning Models with Deep Features Transfer for Human Identification,”
Journal of Innovative Image Processing, vol. 7, no. 3, pp. 850–860, 2025.
4. S. S and S. Mathew, “CNN Based Iris Recognition System: A Novel Approach,” Social Science
Research Network, 2020.
5. S. Bhatnagar, S. Raja, and A. Preeya, “Improving Iris Recognition Systems With CNN Based
Image Processing and Machine Learning Approach,” in Proc. International Conference on
Computer Applications & Mathematical Sciences (ICCAMS), pp. 1–7, 2023.
6. K. Q. Nguyen, M. T. Ha, N. P. Dang, B. X. Viet, and C. V. Phan, “Advancements in Iris
Recognition: WAHET-CNN Framework for Accurate Segmentation and Pattern Classification,”
EAI Endorsed Transactions on Context-Aware Systems and Applications, vol. 9, no. 1, 2023.
7. S. Minaee and A. Abdolrashidi, “DeepIris: Iris Recognition Using Deep Learning Approach,”
arXiv preprint arXiv: 1907.09380, 2019.
8. Z. Yan, L. He, M. Zhang, Z. Sun, and T. Tan, “Hierarchical Multi-class Iris Classification for
Liveness Detection,” in Proc. International Conference on Biometrics (ICB), pp. 47–53, 2018.
9. S. N. Pasha, A. D, R. Anbunathan, G. C, and M. G. K, “An Efficient Novel Approach for Iris
Recognition and Segmentation Based on the Utilization of Deep Learning,” in Proc. IEEE
International Conference on Emerging Applications of Science and Technology (EASCT), pp. 1–
5, 2023.
10. B. Kaur and S. S. Saini, “ContactLensIris: A Robust Iris Recognition System based on Integrated
Deep Learning Approach,” in Proc. International Conference on Electrical, Electronics and
Communication Technologies (ICEECT), pp. 1–5, 2024.
11. S. Shirke, D. Midhun chakkaravarthy, and V. Deshpande, “A Novel Approach to Iris Recognition
at-a-Distance: Leveraging BW-CNN Framework,” Engineering Research Express, 2024.
12. K. Anegsiripong, J. Karnjana, S. Duangpummet, H. Kaneko, and W. Kongprawechnon, “Personal
Identification Based on Fusion of Iris and Periocular Information Using Convolutional Neural
Networks,” in Proc. SICE Annual Conference, pp. 531–536, 2023.
13. G. Liu, W. Zhou, L. Tian, W. Liu, Y. Liu, and H. Xu, “An Efficient and Accurate Iris Recognition
Algorithm Based on a Novel Condensed 2-ch Deep Convolutional Neural Network,” Sensors, vol.
21, no. 11, 3721, 2021.
14. P. S. Prasad, “Deep Learning Based Iris Recognition System,” International Journal of Emerging
Trends in Engineering and Technology, vol. 8, no. 4, pp. 3567–3571, 2018.
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue XI November 2025
Page 467
www.rsisinternational.org
15. Q. Zou, Y. Sun, J. Qin, Z. Huang, and B. Chen, “Toward More Efficient Iris Recognition Using a
Lightweight CNN Framework with Attention Mechanism,” in Proc. SPIE 12290, Image
Processing: Algorithms and Systems XX, 122900F, 2022.
16. S. Kawakami, H. Kawai, K. Ito, T. Aoki, Y. Yasumura, M. Fujio, Y. Kaga, and K. Takahashi, “A
Simple and Accurate CNN for Iris Recognition,” in Proc. Asia-Pacific Signal and Information
Processing Association Annual Summit and Conference (APSIPA ASC), pp. 1566–1571, 2022.
17. H. Menon and A. Mukherjee, “Iris Biometrics Using Deep Convolutional Networks,” in Proc. IEEE
Instrumentation and Measurement Technology Conference (I2MTC), pp. 1–5, 2018.
18. C. Feng, Y. Sun, and X. Li, “Iris R-CNN: Accurate Iris Segmentation in Non-cooperative
Environment,” arXiv preprint arXiv: 1903.10140, 2019.
19. J. E. Zambrano, J. I. Pilataxi, C. A. Perez, and K. W. Bowyer, “Iris Recognition Using an Enhanced
Pre-trained Backbone Based on Anti-Aliased CNNs,” IEEE Access, vol. 12, 2024.
20. R. Sharma, A. Singh, A. Joshi, and A. Gangwar, “Robust Iris Classification Through a Combination
of Kernel Discriminant Analysis and Parzen Based Probabilistic Neural Networks,” in Proc.
International Conference on Emerging Applications of Information Technology (EAIT), pp. 268–
272, 2014.
21. L. Zhou, S. Jiang, S. Zhang, and X. Zhang, “Iris Recognition Based on Convolutional Neural
Network and Active Iterative Optimization,” in Proc. AEECA, pp. 523–528, 2023.
22. J. E. Zambrano, D. P. Benalcazar, C. A. Perez and K. W. Bowyer, "Iris Recognition Using Low-
Level CNN Layers Without Training and Single Matching," in IEEE Access, vol. 10, pp. 41276-
41286, 2022
23. S. Minaee and A. Abdolrashidi, “DeepIris: Iris Recognition Using A Deep Learning Approach,”
arXiv preprint arXiv: 1907.09380.
24. S. Kawakami et al., “A Simple and Accurate CNN for Iris Recognition,” in Proc. APSIPA ASC,
2022.
