Intelligent Human Iris Recognition System Based on Deep Learning

Models

Andreea Negoit

escu

Faculty of Mathematics and Computer Science, “Babes

-Bolyai” University, Cluj-Napoca, Romania

Keywords:

Deep Learning, Iris, Biometrics, Segmentation, Recognition.

Abstract:

This research paper presents the development of an intelligent biometric system which performs human iris

recognition. The software application that incorporates it is called KEYE. Deep learning models are imple-

mented to segment and recognize the users’ irises at authentication. Iris segmentation uses a modiﬁed version

of the U-Net convolutional neural network, trained and validated on images from the I-SOCIAL-DB dataset.

The experimental results prove a maximum validation accuracy of 98.98% and a Dice score of 0.93. The

extraction of features from the segmented images is done using part of the layers of the pre-trained DenseNet-

201 neural network. For classiﬁcation, the KEYE-DB dataset with visible light spectrum images was created.

The accuracy obtained after testing the recognition model is 99.98%. The precision, speciﬁcity, recall and

F1 score exceed 0.9955, while the error and the false positive rate are almost zero, following the conducted

experiments. The performance of the biometric system has proven to be gratifying.

1 INTRODUCTION

Due to the increasing interest in the development of

science and technology worldwide, there is also an

intense focus on security and, implicitly, on the de-

velopment of intelligent systems that use biometric

recognition for human identiﬁcation and veriﬁcation.

Such an authentication system represents the basis

of the KEYE mobile application developed from this

study, which aims to keep users’ credentials and pho-

tos safe from impostors.

Biometrics is deﬁned, in (Tahir and Anghelus

2019), as the technology that analyzes the physiolog-

ical and behavioral features of people, with the aim of

identifying and authorizing them. According to (Ab-

dulkader et al., 2015), it is the most secure human

authentication method among the existing ones: bio-

metric, knowledge-based and possession-based. Bio-

metrics is a vast ﬁeld and is intensively studied by

researchers, because it provides information used in

the design and implementation of security technolo-

gies. It involves a wide range of human recognition

techniques and portrays the unique and detailed char-

acteristics of individuals.

The complexity and uniqueness of the human iris

is fascinating compared to other biometric traits. The

arrangement of pigments, the pattern of the collarette,

the distribution of ﬁbers and blood vessels, give this

natural structure a huge potential for use in the ﬁeld

of security. Thus, the aim of this study is to demon-

strate the reliability and accuracy of human iris fea-

ture recognition using artiﬁcial intelligence. The ob-

jectives of this study are: researching and implement-

ing innovative methods in the ﬁeld of iris biomet-

rics, obtaining performant results after applying deep

learning algorithms, demonstrating the uniqueness of

the iris as a biometric characteristic and ensuring a

high degree of personal data security.

From the ﬁrst studies on iris recognition, there has

been remarkable progress in the diversity and perfor-

mance of the algorithms used for this purpose. Start-

ing from images captured in infrared light, more and

more emphasis has been placed on the use of datasets

containing images from the visible light spectrum.

These images are captured in uncontrolled environ-

ments, where iris region visibility conditions are not

necessarily favorable, as the ones in the datasets used

in this study. It is resorted to the development of

methods with an increased degree of complexity for

the purpose of iris segmentation and classiﬁcation,

based mainly on machine learning techniques, or to

the improvement of already existing ones.

This paper focuses on the implementation of both

segmentation and recognition deep learning methods.

The unique contributions of this study consist, ﬁrstly,

in using a deep learning model that consists of a vari-

Negoi¸tescu, A.

Intelligent Human Iris Recognition System Based on Deep Learning Models.

DOI: 10.5220/0013037000003890

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2025) - Volume 3, pages 15-23

ISBN: 978-989-758-737-5; ISSN: 2184-433X

ation of the U-Net architecture and trained on the I-

SOCIAL-DB dataset, for iris segmentation. Secondly,

the KEYE-DB dataset is created and used for the ﬁrst

time. Its scope is to help training and validating an iris

recognition model to extract relevant features from

irises and classify them, based on the DenseNet-201

neural network. Due to these implemented innova-

tions, the biometric authentication system of KEYE

mobile application proves outstanding performances.

Compared to the present study, others focus either

on segmentation or recognition, or are simply not suit-

able for use on a mobile phone, either because of their

computational complexity or due to the fact that most

of them use images from the infrared light spectrum.

2 RELATED WORK

2.1 The Basis of Human Iris

Recognition

The ﬁrst patent on iris recognition was developed

in 1987 (Flom and Saﬁr, 1987). Most of the cur-

rently existing iris recognition methods have their ba-

sis on the algorithm of the British researcher John

Daugman, patented in 1994. In his work, (Daugman,

1994), iris localization implies a integro-differential

operator in order to demarcate the inner and the outer

contours of the iris. Then, geometric normalization is

performed and Gabor ﬁlters are used to represent the

obtained rectangular image of the iris in binary code.

In the iris code matching process, authenticity veriﬁ-

cation is performed by calculating the Hamming dis-

tance between pairs of codes. In order for two codes

to deﬁne the iris of the same person, the value of the

Hamming distance, scored between 0 and 1 inclusive,

must be as close as possible to 0. The work (Wildes,

1997) investigates the application of the Hough trans-

form for the purpose of detecting the iris and Gaussian

ﬁlters for the representation of its code.

2.2 Recent Studies Regarding Human

Iris Segmentation and Recognition

A study that presents a complex approach is (Gang-

war et al., 2019). From the visible light spectrum,

it uses the UBIRIS.v1, UBIRIS.v2, UTIRIS V.1 and

MICHE-1 datasets in various combinations. Iris

segmentation is performed using a pair of convolu-

tional neural networks. The ﬁrst network, inspired

by YOLO (Redmon et al., 2016), locates the iris and

pupil. It receives as input an image of 448×448 pixels

and the obtained accuracy is 96.78%. The second net-

work, similar to SegNet (Badrinarayanan et al., 2017),

receives an input of size 100×100 pixels. It performs

pixel-level segmentation of the localized region, re-

sulting in an F1 score of 96.98%. For iris binary

code generation, the paper proposes the DeepIrisNet2

architecture, with approximately 100 layers, which

achieves remarkable results without the need for pre-

cise segmentation of images or their normalization.

For the UBIRIS.v2 database, an error EER = 8.51% is

obtained, while for MICHE-1 it varies between 1.05%

and 3.98%. Such a model is extremely computational

expensive to be used on a mobile phone.

Another approach that accepts segmented but non-

normalized images is the ThirdEye system, described

in (Ahmad and Fuller, 2019). It consists of triple

convolutional neural networks, obtained by modify-

ing the architecture of ResNet-50 (He et al., 2016).

The model is trained using three input irises at once,

each of size 200×200 pixels: two are from the same

class and one is from a different class. The recogni-

tion error for the UBIRIS.v2 dataset is EER = 9.20%

and the false rejection rate is FRR = 60%.

The (Ahmadi et al., 2019) study combines, for

image feature extraction, two dimensional Gabor ﬁl-

ters, step ﬁltering and polynomial ﬁltering. Then, for

matching purposes, it uses a neural network with ba-

sic radial functions along with a genetic algorithm.

Using the UBIRIS.v1 database, it achieves an ac-

curacy of 99.9869% after only 10 iterations, with

10 neurons per layer and the following parameters:

population of 150, maximum number of generations

equal to 10, selection factor equal to 3, mutation of

0.35, crossover of 0.5 and recombination of 0.15.

However, the process of locating, segmenting and

normalizing the iris region is not speciﬁed.

The work (Yang et al., 2021) provides the

encoder-decoder architecture of DualSANet. Be-

ing included in the pre-trained ResNet-18 (He et al.,

2016) network, the encoder represents spatially cor-

responding features at multiple levels. For these fea-

tures to fusion, a module based on spatial attention,

integrated in the decoder, is introduced. It gener-

ates dual feature representations that contain comple-

mentary discriminative information. The described

recognition model proves a great performance, hav-

ing a minimum error EER = 0.27% and a rate FRR =

0.31%. It does not specify the behavior of the network

on images in the visible light spectrum, as the exper-

iments are performed on infrared images. They are

ﬁrst segmented, then normalized using the Daugman

Rubber Sheet Model method and resized to 64×512

pixels, before feature extraction.

In the study (Lee et al., 2021), iris recognition

is experimented with NICE.II and MICHE databases.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

The iris region is detected then normalized along with

two periocular regions. The reconstruction of the nor-

malized blurred regions of the iris is done using the

DeblurGAN model (Kupyn et al., 2018). Each triplet

of normalized images becomes the input of a convo-

lutional neural network that extracts a feature vector.

4096 dimensional features are distinguished across

the layers. To check whether two irises correspond

to the same person, the Euclidean distances between

three pairs of feature vectors are calculated. This re-

sults in three scores are merged into one using a sup-

port vector machine. An error of 14.18% is obtained

for images captured with a Samsung Galaxy S4 phone

and 17.02% for those taken with an iPhone 5.

A variant of the algorithm that does not require the

use of artiﬁcial neural networks is described in (Singh

et al., 2020). The illumination and contrast of the im-

ages are improved, then the median ﬁlter is applied to

reduce their noise. For iris localization and segmenta-

tion, the circular Hough transform and the total rela-

tive variation model measure and regularize the local

variation of pixels. The obtained region is normal-

ized with the Daugman Rubber Sheet Model method

and decomposed using the four-level integer wavelet

transform (IWT), which generates 256 frequency sub-

bands. Only the lower 192 sub-bands, which pro-

duce a 192-bit binary code, are considered, by com-

paring their energies with a previously calculated cor-

responding threshold value. This is done by ﬁnding

the Hamming distance between them. The algorithm

achieves, on the UBIRIS.v2 dataset, an accuracy of

98.9% in segmentation and 98.02% in recognition.

Regarding the semantic segmentation of the iris,

one of the most recent approaches is mentioned in

(Pourafkham and Khotanlou, 2023). This presents

the ES-Net architecture, which uses an ESP (Efﬁ-

cient Spatial Pyramid) block (Mehta et al., 2018) to

minimize the time complexity of a network model in-

spired by the U-Net architecture (Ronneberger et al.,

2015), but also an attention mechanism (Vaswani

et al., 2017) to enhance performance. Through the ex-

periments, a MIOU (Mean Intersection Over Union)

score of 93.61% and an F1 score of 97.03% are ob-

tained for the UBIRIS.v2 dataset.

The study (Nourmohammadi Khiarak et al., 2023)

proposes a new dataset, called KartalOl. It contains

images from the visible light spectrum, captured us-

ing a mobile phone camera. As a segmentation ar-

chitecture, Mobile-Unet is built, consisting of the pre-

trained MobileNetV2 model (Sandler et al., 2018), in-

tegrated into the encoder part of the U-Net network.

It achieves 98% accuracy on validation data.

3 PROPOSED IRIS

SEGMENTATION APPROACH

3.1 I-SOCIAL-DB Dataset

For segmentation, the I-SOCIAL-DB, namely Iris So-

cial Database dataset (Donida Labati et al., 2021) was

used. It contains 3286 color images from the visible

light spectrum, collected from a sample of 400 sub-

jects, in uncontrolled environments. These were ob-

tained by extracting two eye regions of 300×350 pix-

els each, corresponding to the left and right eye, from

1643 high-resolution portrait images. Because they

were collected from various online public sources,

both the devices that captured the images and the dis-

tances from which they were taken are unknown.

Figure 1: I-SOCIAL-DB ocular region sample.

Each image in this dataset corresponds to a manu-

ally constructed segmentation mask at the pixel level,

as in the example in Figure 1. The mask highlights,

through white pixels, the iris as the region of interest,

excluding reﬂections and other possible occlusions.

The portion of the iris after segmentation represents,

on average, 71.4% of the total area of the ring formed

by the circles that approximate the inner and outer

border of the iris.

3.2 U-Net Architecture

The architecture of the U-Net convolutional neural

network was ﬁrstly introduced in the paper (Ron-

neberger et al., 2015). Even though it was originally

intended for the processing of microscopic biomedi-

cal images, it also proves extraordinary results in the

case of semantic segmentation of human irises pho-

tographed under various conditions. The advantage of

this network lies, in addition to the speed of segmen-

tation, in the useful ability to learn from a relatively

small set of data. This is proven by the great per-

formances shown by the network through the experi-

ments conducted in this research, as the used dataset

for learning contains only 3286 images. Also, being

a fully convolutional network, the sizes of the outputs

adapt to those of the input image, so their resolutions

and number of channels can vary. The U-Net archi-

tecture is of encoder-decoder type, being formed of

Intelligent Human Iris Recognition System Based on Deep Learning Models

a contraction path, followed by an expansion path.

They are connected to each other symmetrically, to

preserve information lost by contracting. The net-

work contains 23 convolutional layers.

The contraction path consists of repetitive steps

which extract relevant features from images. Each

step involves the application of two convolutions with

a ﬁlter of 3×3. Each of them is followed by a ReLU

function for activation, at the end of which a 2×2 max-

pooling operation is performed with a step of 2. To

compensate for this reduction in spatial dimensional-

ity caused by subsampling, the number of channels of

the feature maps is doubled at each iteration.

The expansion path is relatively symmetrical and

achieves a precise localization, at pixel level, of the

region of interest. Each step represents an upsampling

of the feature map. Then, a 2×2 ﬁlter convolution that

halves the number of channels is applied, a concate-

nation with a copy of the clipped feature map from

the corresponding step of the shrinking path and two

3×3 convolutions, followed by one ReLU activation

function each. The ﬁnal layer maps each 64-element

feature vector to the desired number of classes.

In this study, the U-Net neural network is adapted

to act as a binary classiﬁer, assigning each pixel in

each input image a corresponding class, iris or non-

iris. The difference from the original model described

in (Ronneberger et al., 2015) consists, ﬁrstly, in the

use of padding in the case of convolutions and the

BatchNorm2d layer (Ioffe and Szegedy, 2015), which

has the role of normalizing the activations between

network layers. Since the bias will be canceled by this

normalization layer, its existence is no longer neces-

sary. Also, because color input images are provided

to the network, the input layer contains 3 channels.

For semantic segmentation, a binary classiﬁcation of

the pixels is performed, so the existence of a single

output channel is sufﬁcient.

Another important step is to ensure that the net-

work works properly for any input image dimensions

by performing resizing when concatenating the fea-

ture maps. Otherwise, if the input dimensions were

not divisible by 2 at each of the four steps at which

the max-pooling operation is performed within the

contracting path, some pixels would be lost. For ex-

ample, if max-pooling is performed on an image of

size 175×175, an image of 87×87 pixels will result.

In the expansion path, when oversampled, it will end

up being only 174×174 pixels in size. In order to be

concatenated with the original image, they must be

brought to the same size.

3.3 Training and Validation

The described model was trained and validated on 3-

channel color images from the I-SOCIAL-DB dataset,

both original size of 300×350 pixels and resized to

160×240 pixels. The ﬁrst 3000 images, respectively

masks, were kept for training and the next 286 im-

ages, respectively masks, for validation. In this way,

the train : validation ratio is approximately 90 : 10.

By feeding the network batches of 16 images, 188

steps are performed in each training epoch. The cho-

sen loss function is Binary Cross Entropy. For opti-

mization, the Adam algorithm is used, with a constant

learning rate of 0.001. The model is saved in a check-

point every time the validation accuracy increases fol-

lowing the completion of an epoch.

3.4 Experimental Results

For images of 300×350 pixels, the model proves out-

standing performance even from the ﬁrst epoch. On

the training dataset, an accuracy of 98.01%, an aver-

age loss of 0.67731, a ﬁnal loss of 0.276 and a Dice

score of 0.83326 are obtained. Upon validation, an

accuracy of 97.85%, and an average loss equal to

0.67600 is obtained. The Dice score increases by

approximately 0.00222, reaching a value of 0.83548.

The model is trained over 18 epochs, each taking be-

tween 2 and 4 hours to run. This number was chosen

because after epoch 18 the model performance does

not improve anymore. The results can be observed in

the graphics from Figure 2 and Figure 3, where the

color blue is used for training evolution and the or-

ange color corresponds to validation evolution.

Figure 2: Analysis of accuracy and Dice score after training

and validating the model on 300×350 pixels images.

Figure 3: Average and ﬁnal loss progress of the model dur-

ing training and validation on 300×350 pixels images.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

Table 1: Comparison of model performance for images of different sizes.

Criterion

Training Validation

300×350 160×240 300×350 160×240

Accuracy

ﬁrst epoch 98.01% 98.33% 97.85% 98.07%

last epoch 99.30% 99.26% 98.98% 98.97%

maximum 99.33% 99.27% 98.98% 98.97%

Dice score

ﬁrst epoch 0.83326 0.86242 0.83548 0.85592

last epoch 0.94474 0.94195 0.92860 0.92833

maximum 0.94765 0.94195 0.92860 0.92833

Medium loss

ﬁrst epoch 0.67731 0.67570 0.67600 0.67488

last epoch 0.67174 0.67185 0.67141 0.67138

minimum 0.67163 0.67179 0.67141 0.67138

Final loss

ﬁrst epoch 0.276 0.184 - -

last epoch 0.0182 0.0188 - -

minimum 0.0182 0.0173 - -

At the end of the 18th epoch, an accuracy of

99.30% is noted on the training data, 0.03% lower

than the maximum, which was achieved in the 17th

epoch. The training Dice score is maximum in the

penultimate epoch, reaching 0.94765. In epoch 18,

it drops to 0.94474. The average loss in epoch 18

is equal to 0.67174, while in the previous epoch it is

lower, reaching the minimum value of 0.67163. Con-

sidering the value of the ﬁnal loss at the end of each

of the 18 epochs, without calculating the average of

its values within them, its decrease during training is

achieved progressively, from 0.276 in the ﬁrst epoch

to 0.0182 in the last. Upon validation, the accuracy

and the Dice score reach the maximum of 98.98%,

respectively 0.92860, in the last epoch, in which the

average loss is also minimal, being equal to 0.67141.

From the ﬁrst to the last epoch, a 1.13% increase in

accuracy is reported. The Dice score also increases

by about 0.09312 and the average loss decreases by

about 0.0046.

The segmentation results of 4 validation images,

along with their original masks above, are illustrated

in Figure 4. After training the model for 18 epochs, it

recognizes iris reﬂections with high accuracy.

Figure 4: Original vs. predicted iris masks.

To reduce the execution time of an epoch to a

maximum of one hour, it is experimented with im-

ages resized to 160×240 pixels. For these, the model

proves, in the ﬁrst of the 17 total epochs, better results

compared to the previously described approach. This

number was chosen because after epoch 17 the model

performance does not improve anymore. A compari-

son of the results from the ﬁrst and ﬁnal epochs and

the best values of the metrics obtained for both dimen-

sions of the images, is made in Table 1.

4 PROPOSED IRIS

RECOGNITION APPROACH

4.1 KEYE-DB Dataset Creation

The KEYE-DB dataset contains 1370 3-channel color

images from the visible light spectrum captured by

various mobile phones with high-resolution cameras.

To capture them, both the front and back cameras of

the devices were used, with and without ﬂash. They

were positioned at distances between 7 and 10 cen-

timeters from the eyeball of the subjects, who were in

various environments with natural or artiﬁcial light.

This study involved 36 subjects, 22 women and 14

men, with irises of various colors and shades. The

subjects belong to several age categories: 5-20 years

(14%), 20-35 years (33%), 35-50 years (25%), 50-65

years (20%) and 65-80 years (8%). Most of them are

between 20 and 35 years old. Both the left and the

right iris were photographed for each. Between 25

and 50 photographs were collected for each individ-

ual, with an average of approximately 38 photographs

per person. As the privacy of the subjects is priori-

tized, the dataset is not made publicly accessible.

Most images have been cropped to approximate

3:4 or 4:3 aspect ratios. They were then resized to

300×350 pixels height×width for further segmenta-

tion. In the case of older subjects, a cropping of the

images that more closely frames the eye region than

Intelligent Human Iris Recognition System Based on Deep Learning Models

in the case of the others was considered. Thus, a pre-

cise segmentation was ensured, which is not disturbed

by the uneven distribution of light on the skin folds.

Each binary mask obtained after segmentation

was transformed back to the original image dimen-

sions. To avoid false positive regions in the predicted

mask as much as possible, the largest area of white

pixels is found, as it is most likely to describe the iris

region. Then, the radius and the center of the smallest

enclosing circle are calculated to simulate the outer

boundaries of the iris. With their help, the coordinates

of the square that inscribes this circle are determined.

Next, a multiplication of the pixel values of the origi-

nal image with those of the mask pixels is performed,

to obtain an image in which the white pixels in the

mask are replaced by the corresponding ones in the

original image. The resulting image is cropped based

on the coordinates of the previously obtained square,

resized to 300×300 pixels and saved in the folder cor-

responding to the subject to which the iris belongs.

These stages are summarized in Figure 5.

Figure 5: General steps of obtaining KEYE-DB images.

Figure 6: Iris images augmentations.

The entire process is applied to all captured im-

ages. After completion, the dataset can be augmented

by rotating each image by -60, -40, -20, 20, 40, and

60 degrees, respectively. For each image among the

approximately 30 of a subject, 6 more images are ob-

tained, as in Figure 6. In this way, each subject will

have 7 times more images of their own irises than

originally. This type of augmentation is necessary be-

cause various factors can obscure the iris region and

cause signiﬁcant areas of black pixels, whose orienta-

tion is not relevant.

4.2 DenseNet-201 Architecture

The DenseNet-201 convolutional neural network ar-

chitecture was ﬁrst introduced in the study (Huang

et al., 2017). Its major advantage is the presence of

dense blocks, where each layer has direct connections

to all the others. With their help, the risk of losing

information through the network layers is consider-

ably reduced, while the direction of its transmission

remains constant. Dense connections reduce the num-

ber of parameters and avoid possible overﬁtting ten-

dencies, which is why DenseNet-201 was chosen to

be used in this research.

Considering that each Conv layer corresponds to

the triplet of BN (Batch Normalization) (Ioffe and

Szegedy, 2015), ReLU and Conv layers, an example

of the DenseNet-201 network architecture is shown in

Table 2.

Table 2: DenseNet-201 architecture with a growth rate of

32 (Huang et al., 2017).

Layers

Output

dimension

DenseNet-201

Convolution 112×112 7×7 Conv, step=2

Pooling 56×56

3×3 max-pooling,

step=2

Dense

Block (1)

56×56

[1x1 Conv and

3x3 Conv] x 6

Transition

layer (1)

56×56 1×1 Conv

28×28

2×2 average

pooling, step=2

Dense

block (2)

28×28

[1x1 Conv and

3x3 Conv] x 12

Transition

layer (2)

28×28 1×1 Conv

14×14

2×2 average

pooling, step=2

Dense

block (3)

14×14

[1x1 Conv and

3x3 Conv] x 48

Transition

layer (3)

14×14 1×1 Conv

7×7

2×2 average

pooling, step=2

Dense

Block (4)

7×7

[1x1 Conv and

3x3 Conv] x 32

Classiﬁcation

layer

1×1

7×7 global average

pooling

completely connected

and Softmax

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

The inputs of each layer are represented by the

concatenation of feature maps received from previous

layers and the total number of connections between L

layers has a value equal to c

L · (L + 1)

(1)

The layer l that performs the nonlinear transfor-

mations denoted by H

receives, at input, the concate-

nated outputs x

, x

, up to x

l−1

, of the previous layers.

Thus, its output is given by x

= H

l−1

) + x

l−1

= H

([x

, . . . , x

l−1

]) (2)

If each function H

produces k feature maps,

which represent the growth rate of the network, layer

l will have, at input, m

feature maps.

= k

+ k · (l − 1) (3)

4.3 Feature Extraction and

Classiﬁcation

In order to extract the relevant features from the pre-

viously segmented iris images, the DenseNet-201 net-

work is used, from which the last 58 layers are re-

moved. It is pre-trained on the ImageNet dataset,

which contains 1281167 images, divided into 1000

classes. The features are extracted using the weights

resulting from learning based on the data in that set

and become the inputs of a multi-class classiﬁer rep-

resented by a simple artiﬁcial neural network using

a Flatten and a Dense layer with Softmax activation.

For each feature, the prediction of the classiﬁer is

a probability distribution, indicating the percentage

match of the image to each of the existing classes.

4.4 Training and Validation

The classiﬁcation model was trained and validated

over 50 epochs using batches of 8 images from the

KEYE-DB dataset, with and without augmentation.

This batch size was chosen as it proved the best vali-

dation results. It is small enough to provide frequent

gradient update and great generalization potential, but

also large enough to be used on a relative small dataset

like KEYE-DB. The images were randomly split so

that 30% of them were dedicated to validation and

70% to training. The used loss is Categorical Cross

Entropy and, as an optimizer, the Adam algorithm

was chosen, with a learning rate equal to 0.001. The

model is saved in a checkpoint each time the valida-

tion accuracy increases following the completion of

an epoch. This technique is very useful, as it provides

the possibility of continuing the training starting from

the iteration that proved the best result previously.

4.5 Experimental Results

After training the model on the KEYE-DB dataset

without augmentation, a maximum accuracy of 100%

is achieved on the training dataset after only 4 epochs,

which is maintained up to epoch 50. The maximum

validation accuracy is 96.594% and results at the end

of epoch 20, when the validation loss is 0.1590 and

the training loss is 0.0016. It then oscillates for 30

epochs, without exceeding the mentioned maximum

percentage and having, in the ﬁnal epoch, a value of

96.11%. The minimum training, respectively valida-

tion loss, is recorded in the last epoch, with the values

of 1.9594e-04, respectively 0.1428. The model per-

formance over epochs is shown in Figure 7.

Figure 7: Analysis of accuracy and loss after training and

validating on the KEYE-DB dataset without augmentation.

To increase the performance of the recognition

model, the images from the KEYE-DB dataset are

augmented as previously speciﬁed. Thus, after train-

ing the model, a maximum accuracy of 100% is ob-

tained over 50 epochs on the training dataset, which

is maintained from epoch 23 to the end, and 99.583%

on the validation dataset, at epoch 47. In epoch 50

the same accuracy values are recorded as in epoch 47,

but the loss decreases, from 6.1620e-09 to 4.4750e-

09, respectively from 0.0299 to 0.0295. The model

performance over epochs is shown in Figure 8.

Table 3 compares the performances of the model

considering the average value of each metric for all

classes. It proves the advantage of data augmentation.

Table 3: Comparative analysis of the validation perfor-

mance of the model before and after data augmentation.

Performance

metric

Before data

augmentation

After data

augmentation

Accuracy 0.998107597 0.999768277

Error 0.001892403 0.000231723

Precision 0.967860534 0.995949288

Recall 0.967048978 0.995510867

Speciﬁcity 0.999024939 0.999880640

False Positive

Rate

0.000975061 0.000119360

F1 score 0.966338857 0.995705300

Intelligent Human Iris Recognition System Based on Deep Learning Models

Table 4: Comparative analysis of proposed segmentation and recognition approaches with the existing ones in literature.

Study Segmentation performance Recognition performance

proposed

accuracy = 98.98%

Dice = 0.93

accuracy = 99.98%

F1 ≈ 99.57%

EER, FRR ≈ 0

(Gangwar et al., 2019)

accuracy = 96.78% YOLO

F1 = 96.98% SegNet

EER = 8.51% UBIRIS.v1

1.05% ≤ EER ≤ 3.98% MICHE-1

(Ahmad and Fuller, 2019) -

EER = 9.20%

FRR = 60%

(Ahmadi et al., 2019) - accuracy = 99.9869%

(Yang et al., 2021) -

EER = 0.27%

FRR = 0.31%

(Lee et al., 2021) -

EER = 14.18% Samsung Galaxy S4

EER = 17.02% iPhone 5

(Singh et al., 2020) accuracy = 98.9% accuracy = 98.02%

(Pourafkham and Khotanlou, 2023)

MIOU = 93.61%

F1 = 97.03%

(Nourmohammadi Khiarak et al., 2023) accuracy = 98% -

Figure 8: Analysis of accuracy and loss after training and

validating on the KEYE-DB dataset after augmentation.

5 DISCUSSION

After analyzing the obtained experimental results, in

the case of segmentation, it was concluded that the

model trained on images of size 300×350 is the most

suitable. Despite the longer training time, the decision

was made considering the increased performance of

the model. It proves a maximum validation accuracy

equal to 98.98% and a Dice score of about 0.93.

For iris classiﬁcation based on the features ex-

tracted from the segmented images, data augmenta-

tion was chosen, with the model obtaining an accu-

racy of 99.98% and an almost insigniﬁcant error, as

well as the false positive rate. The precision, speci-

ﬁcity, recall and F1 score all exceed the value of

0.9955. Thus, the chance of unauthorized persons

logging into the application is almost zero. This state-

ment is made considering that no photos of digital or

printed iris pictures are used, as these cases have not

yet been extensively tested to reach a ﬁrm conclusion.

As seen in Table 4, this study shows promis-

ing results and even competitive with those obtained

in other studies in existing literature. The study

(Gangwar et al., 2019) performs recognition using

iris matching, as well as classiﬁcation, while (Ahmad

and Fuller, 2019), (Ahmadi et al., 2019), (Yang et al.,

2021) and (Lee et al., 2021) describe, in essence, iris

matching approaches.

The choice for using a classiﬁcation approach in

this study has been made due to the fact that by trying

several variants of iris matching algorithms, no satis-

factory results were obtained, considering the limited

public datasets resources that contain images from the

visible light spectrum. The comparison in Table 4

is made with the mention that the described studies

do not use the same datasets nor approaches as in the

present work. An exact comparison cannot be made

because there are no relevant studies in the literature

that address the problem in this paper using the I-

SOCIAL-DB dataset for segmentation. Also, there

are no studies that use the KEYE-DB dataset since it

is created within this work. However, an attempt was

made to select related studies that use similar datasets.

The limitations of this study consist in the number

of subjects who agreed to participate in the research

by providing photos of their irises to the KEYE-DB

dataset. As a future improvement, it is desired to ex-

pand the sample of users recognized by the KEYE

application. Also, additional veriﬁcation at authenti-

cation should be implemented to conﬁrm the physical

presence of the user who tries to access the applica-

tion, such as the live recording of subtle but continu-

ous movements of the pupil. This should stop fraudu-

lent intents of authentication using photos of iris pic-

tures or artiﬁcial irises.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

6 CONCLUSIONS

In conclusion, following the extensive research car-

ried out in this paper, it is conﬁrmed with certainty

that the iris is a biometric feature that meets all the

necessary conditions to be used in the implementation

of a reliable biometric recognition system. The exper-

imental results are gratifying for the development of

the KEYE mobile application, so the objectives of this

research were achieved.

REFERENCES

Abdulkader, S., Atia, A., and Mostafa, M.-S. (2015). Au-

thentication systems: principles and threats. Com-

puter and Information Science, 8.

Ahmad, S. and Fuller, B. (2019). Thirdeye: Triplet based

iris recognition without normalization. 2019 IEEE

10th International Conference on Biometrics Theory,

Applications and Systems (BTAS), pages 1–9.

Ahmadi, N., Nilashi, M., Samad, S., Rashid, T., and Ah-

madi, H. (2019). An intelligent method for iris recog-

nition using supervised machine learning techniques.

Optics & Laser Technology, 120.

Badrinarayanan, V., Kendall, A., and Cipolla, R. (2017).

Segnet: A deep convolutional encoder-decoder ar-

chitecture for image segmentation. IEEE Transac-

tions on Pattern Analysis and Machine Intelligence,

39(12):2481–2495.

Daugman, J. (1994). Biometric personal identiﬁcation sys-

tem based on iris analysis.

Donida Labati, R., Genovese, A., Piuri, V., Scotti, F.,

and Vishwakarma, S. (2021). I-social-db: A labeled

database of images collected from websites and social

media for iris recognition. Image and Vision Comput-

ing, 105(104058):1–9. 0262-8856.

Flom, L. and Saﬁr, A. (1987). Iris recognition system.

Gangwar, A., Joshi, A., Joshi, P., and Ramachandra, R.

(2019). Deepirisnet2: Learning deep-iriscodes from

scratch for segmentation-robust visible wavelength

and near infrared iris recognition.

He, K., Zhang, X., Ren, S., and Sun, J. (2016). Deep resid-

ual learning for image recognition. In 2016 IEEE Con-

ference on Computer Vision and Pattern Recognition

(CVPR), pages 770–778.

Huang, G., Liu, Z., Van Der Maaten, L., and Weinberger,

K. Q. (2017). Densely connected convolutional net-

works. In 2017 IEEE Conference on Computer Vision

and Pattern Recognition (CVPR), pages 2261–2269.

Ioffe, S. and Szegedy, C. (2015). Batch normalization: Ac-

celerating deep network training by reducing internal

covariate shift. In Proceedings of the 32nd Interna-

tional Conference on Machine Learning, volume 37

of Proceedings of Machine Learning Research, pages

448–456, Lille, France. PMLR.

Kupyn, O., Budzan, V., Mykhailych, M., Mishkin, D., and

Matas, J. (2018). Deblurgan: Blind motion deblurring

using conditional adversarial networks. IEEE/CVF

Conference on Computer Vision and Pattern Recog-

nition, pages 8183–8192.

Lee, M. B., Kang, J. K., Yoon, H. S., and Park, K. R. (2021).

Enhanced iris recognition method by generative ad-

versarial network-based image reconstruction. IEEE

Access, 9:10120–10135.

Mehta, S., Rastegari, M., Caspi, A., Shapiro, L. G., and Ha-

jishirzi, H. (2018). Espnet: Efﬁcient spatial pyramid

of dilated convolutions for semantic segmentation. In

European Conference on Computer Vision.

Nourmohammadi Khiarak, J., Golzari Oskouei, A.,

Salehi Nasab, S., Jaryani, F., Moaﬁnejad, N., Pourmo-

hamad, R., Amini, Y., and Noshad, M. (2023). Kar-

talol: Transfer learning using deep neural network for

iris segmentation and localization: New dataset for iris

segmentation. Iran Journal of Computer Science, 6:1–

13.

Pourafkham, B. and Khotanlou, H. (2023). Es-net: Unet-

based model for the semantic segmentation of iris.

Redmon, J., Divvala, S., Girshick, R., and Farhadi, A.

(2016). You only look once: Uniﬁed, real-time ob-

ject detection. IEEE Conference on Computer Vision

and Pattern Recognition (CVPR), pages 779–788.

Ronneberger, O., Fischer, P., and Brox, T. (2015). U-

net: Convolutional networks for biomedical im-

age segmentation. In Medical Image Computing

and Computer-Assisted Intervention – MICCAI 2015,

pages 234–241, Cham. Springer International Pub-

lishing.

Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., and

Chen, L.-C. (2018). Mobilenetv2: Inverted residu-

als and linear bottlenecks. In 2018 IEEE/CVF Con-

ference on Computer Vision and Pattern Recognition,

pages 4510–4520.

Singh, G., Singh, R., Saha, R., and Agarwal, N. (2020). Iwt

based iris recognition for image authentication. Pro-

cedia Computer Science, 171:1868–1876.

Tahir, A. A. K. and Anghelus

, S. (2019). Human biometrics

and biometric recognition systems; an overview. In

The 19th International Multidisciplinary Conference

”Professor Dorin Pavel - the founder of Romanian hy-

dropower”, volume 35, pages 431–446.

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones,

L., Gomez, A. N., Kaiser, L., and Polosukhin, I.

(2017). Attention is all you need. volume 30, pages

5998–6008. Curran Associates Inc.

Wildes, R. (1997). Iris recognition: an emerging biomet-

ric technology. Proceedings of the IEEE, 85(9):1348–

1363.

Yang, K., Xu, Z., and Fei, J. (2021). Dualsanet: Dual spa-

tial attention network for iris recognition. In IEEE

Winter Conference on Applications of Computer Vi-

sion (WACV), pages 888–896.

Intelligent Human Iris Recognition System Based on Deep Learning Models