Hybrid Deep Learning Model Based on Autoencoder and CNN for Palmprint Authentication

Abstract

Palmprint authentication has received a lot of attention as one of the most prevalent biometric methods. A palmprint is a portion of the palm's surface that has special characteristics that could be used for authentication. Getting the most valuable features out of a palmprint, on the other hand, is a major challenge. Another challenge is coming up with an efficient strategy to authentication that uses fewer images, especially with approaches that require a high number of images in the training phase, which is a major issue. The majority of recently developed approaches rely on primary lines, wrinkles, and creases, which in some situations are insufficient to separate two people due to resemblance. Deep learning methods have recently been viewed as a critical component for extracting deep features such as texture features in these types of tasks. We concentrated on the palmprint authentication challenge in this work by creating a hybrid model called AE+CNN, which is based on the autoencoder (AE) model and a convolutional neural network (CNN) model. The proposed model comprises three phases: pre-processing, region of interest extraction (ROI), and feature extraction and matching using hybrid AE+CNN. The experiments used the COEP palmprint database, which has a limited number of palmprint images, posing a significant challenge for deep learning models that require a large number of images for training. The F1-score and the accuracy metric were both employed in the evaluation. with a score of 97.85 % for accuracy and 96.81 % for F1-score. Gradient-weighted class activation mapping (Grad-CAM) was also applied to figure out which parts of the palmprint are the most discriminative for class classification.

Full text

Received: February 24, 2022. Revised: April 1, 2022. 488
International Journal of Intelligent Engineering and Systems, Vol.15, No.3, 2022 DOI: 10.22266/ijies2022.0630.41
Hybrid Deep Learning Model Based on Autoencoder and CNN for Palmprint
Authentication
Firas Muneam Bachay1* Mohammed Hasan Abdulameer2
1 Department of Computer Science, Faculty of Computer Science and Mathematics, University of Kufa, Iraq
2 Department of Computer Science, Faculty of Education for Women, University of Kufa, Iraq
* Corresponding author’s Email: firasm.bachay@uokufa.edu.iq
Abstract: Palmprint authentication has received a lot of attention as one of the most prevalent biometric methods. A
palmprint is a portion of the palm’s surface that has special characteristics that could be used for authentication.
Getting the most valuable features out of a palmprint, on the other hand, is a major challenge. Another challenge is
coming up with an efficient strategy to authentication that uses fewer images, especially with approaches that require
a high number of images in the training phase, which is a major issue. The majority of recently developed
approaches rely on primary lines, wrinkles, and creases, which in some situations are insufficient to separate two
people due to resemblance. Deep learning methods have recently been viewed as a critical component for extracting
deep features such as texture features in these types of tasks. We concentrated on the palmprint authentication
challenge in this work by creating a hybrid model called AE+CNN, which is based on the autoencoder (AE) model
and a convolutional neural network (CNN) model. The proposed model comprises three phases: pre-processing,
region of interest extraction (ROI), and feature extraction and matching using hybrid AE+CNN. The experiments
used the COEP palmprint database, which has a limited number of palmprint images, posing a significant challenge
for deep learning models that require a large number of images for training. The F1-score and the accuracy metric
were both employed in the evaluation. with a score of 97.85 % for accuracy and 96.81 % for F1-score. Gradient-
weighted class activation mapping (Grad-CAM) was also applied to figure out which parts of the palmprint are the
most discriminative for class classification.
Keywords: Palmprint, Biometric, Deep learning, Autoencoder, Authentication.
1. Introduction
Biometrics is a method for identifying or
verifying an individual’s identity based on their
physical or behavioural traits. Human identity and
recognition have relied on biometric traits such as
the face, iris, fingerprint, hand shape, palmprint, and
signature [1]. Among all of these capabilities,
palmprint recognition has recently acquired a lot of
interest as a viable personal identification method.
In terms of principal lines, creases, and wrinkles,
Palmprint contains a lot of detail. Principal lines,
geometry, and texture have all been studied for
discerning features [2].
Several recent studies have worked on palmprint
authentication using various methods, where
researchers suggest in [3] a palmprint feature
extraction and recognition method based on double
half-orientation with three different datasets.
Though, they didn’t represent the orientation
property of cross points with a single dominating
orientation. They did not rely on extracting other
types of distinct features accessible in the palmprint,
instead depending on just one. The authors in [4]
proposes palmprint recognition system (PRS), which
combines direction, local binary pattern (LBP)
features, C5.0, and K-nearest neighbour (KNN)
techniques for two datasets and shows promising
results. The proposed method, on the other hand,
separates the image into several little sections,
resulting in the loss of many critical elements that
are dependent on the entire image structure. The
study [5] uses the dual-tree complex wavelet

Received: February 24, 2022. Revised: April 1, 2022. 489
International Journal of Intelligent Engineering and Systems, Vol.15, No.3, 2022 DOI: 10.22266/ijies2022.0630.41
transform to offer a palmprint detection image
feature descriptor (DT-CWT). The approach makes
use of spatial structural information, which divides
each image into many subblocks. As a result, many
crucial characteristics that depend on the entire
image structure are lost. Also, database is also
acquired by a scanner, which reduces user
satisfaction and inconvenient situations. By
integrating the discrete cosine transform (DCT) and
an autoregressive (AR) signal modelling a palmprint
identification method is introduced by [6].
Nevertheless, the technique is interested in the
texture features of palmprints and ignores other
features. The study [7] suggests a palmprint personal
identification system based on the merging of local
and global data. The discrete orthonormal stockwell
transform is used to extract the local features of the
improved palmprint. By lowering the scale of the
discrete orthonormal stockwell transform to infinity,
the global feature is achieved. Though, it is difficult
to precisely get the local features due to the
sensitivity of the palmprint picture, since the
performance of the local features declines if the
palmprint quality and capture area are inadequate.
Furthermore, the combination with the global
features took a long time to compute. Deep learning
approaches have recently been recognized as
effective methods for a variety of domains, one of
which is biometrics. For instance, convolutional
neural networks have been investigated with
palmprints by [8]. The experimental results showed
that this approach achieves very good accuracy on
the PolyU dataset. Though, the suggested network is
made up of eleven convolutional layers, which
increases the model’s computational complexity
while also increasing network training time.
Likewise, the researchers in [9] suggest an enhanced
deep convolutional generative adversarial net
(DCGAN) to create high-resolution palmprint
images by swapping the convolutional transpose
layer by linear upsampling and introducing the
structure similarity (SSIM) index into the loss
function. The authors employ traditional data
augmentation techniques and proposed deep
convolutional GAN to increase data (DCGAN).
Conversely, the majority of the researches showed
effective results without using any data
augmentation techniques. In [10] also employs the
Alexnet convolutional neural network (CNN)
structure. At initial, only the ROI region of the
palmprint was cut away. The ROI region is then
used as the input to the convolutional neural
network after it has been processed. They used a
standard convolutional neural network architecture
only and the database in their work is massive as
well as multispectral in nature. Which is make that
the training duration is excessively long and also the
training images’ environment is not realistic. The
researchers in [11] developed a deep neural network
termed a palm convolutional neural network to solve
the palmprint verification challenge (PCNN). The
model has a simple structure with only one
convolutional layer, which may not be able to
extract deep and important characteristics in the
palm. In recent times, autoencoder a deep learning
model has demonstrated impressive image
reconstruction skills based on image discriminative
properties with feature reduction skills [12, 13, 14].
In this paper, we introduce a new hybrid model
based on an autoencoder and a convolutional neural
network (AE+CNN) that combines the advantages
of both models into a single hybrid model. The
following is how the rest of the paper is ordered:
Part 2 gives the problem statement; Part 3 gives the
theoretical background; Part 4 the proposed
palmprint authentication technique; Part 5 displays
and discusses the experimental results; and Part 6
conclusion the paper.
2. Problem statement
Palmprint authentication necessitates the
extraction of palmprint features prior to
classification, which has an essential effect on the
classification rate. The most significant features, on
the other hand, have a direct impact on the ultimate
authentication process. Another issue is
classification of the extracted palm features, which
can be a problem for any authentication method,
therefore choosing the optimal classification
approach is crucial too. The majority of traditional
solutions rely on fundamental lines, wrinkles, and
creases features that are missing to distinguish two
people precisely due to likeness. Also, there are
many advanced features on the palm that can be
utilized to create authenticity. Furthermore, past
palm systems relied on touch directly between the
palm pattern and the capture system device, which
could reduce user adoption. As a result, current
research has concentrated on contact-free solutions,
which make it more pleasant and hygienic by
eliminating the need for physical contact.
Furthermore, the acquisition device may be costly
for capturing high-resolution photos, particularly for
the palmprint biometric feature. For several works,
this resulted in the use of low-resolution hand
images in the acquisition module. Deep learning
approaches like CNN have recently been successful
in a variety of disciplines, including biometrics.
However, the vast majority of deep learning models,

Received: February 24, 2022. Revised: April 1, 2022. 490
International Journal of Intelligent Engineering and Systems, Vol.15, No.3, 2022 DOI: 10.22266/ijies2022.0630.41
Figure. 1 The standard structure of the palmprint authentication procedure
on the other hand, necessitate a huge dataset in order
to process and train the neural network. According
to the over mentioned issues, our objectives are to
provide a deep network-based approach that
increases accuracy, lowers the cost of the biometric
system, effective with little data, and enhances user
acceptance by focusing on only the most important
aspects of the image. Therefore, we developed a
hybrid model that can benefit from both an
autoencoder network and a CNN simultaneously.
3. Theoretical background
This section provides an overview of palmprint
authentication before moving on to a quick
explanation of deep learning networks (autoencoder
and CNN).
3.1 Palmprint authentication procedure in
general
Palm print collecting, pre-processing, region of
interest extraction (ROI), feature extraction,
matching, and decision are the five components of a
typical palmprint authentication system, as depicted
in Fig. 1.
3.2 Autoencoder network
An auto-encoder is a deep neural network that
encodes and decodes data effectively via
unsupervised feature learning. It can learn robust
features from unlabelled data automatically. This
approach is divided into two stages: encoding and
decoding. contains one or more hidden layers with
an input and output layer. The input is compressed
into a lower-dimensional feature with a meaningful
representation during the encoding stage. Fig. 2
depicts the simple architecture of an autoencoder
[15].
3.3 Convolutional neural network
Deep learning is an artificial neural network
with numerous layers. One of the most widely used
deep neural networks is CNN. Fig. 3 shows CNN’s
Figure. 2 The standard structure of autoencoder
numerous layers, which include a convolution layer,
a pooling layer, and a fully-connected layer (FC).
Convolution is the initial layer of CNN, and it
consists of a set of filters. These filters are initialized
by the convolutional neural network to make them
more suitable for the task at hand. There is potential
to add additional layers after the input layer to
improve the efficiency of this method. Each layer
can have its own set of filters. Stride, padding, and
an activation function are among the various
adjustments available. Eq. (1) explains the
computations for this layer.
 

   














(1)
Where: Mr,x,yf is an output of the convolutional
layer, (r,x) is a pixel coordinate, 
 is a channel bias,


 is the kernel weighs, 
 and 
 are
respectively width and height of convolutional layer
kernel, y is channel number, f is the present layer,
and f−1 is the before layer.
Pooling is the next layer following convolution.
Pooling's main purpose is to down-sample
succeeding layers to minimize their complexity. It's
equivalent to decreasing the resolution in the context
of image processing. Pooling has no effect on the
number of filters. Max-pooling is one of the most

Received: February 24, 2022. Revised: April 1, 2022. 491
International Journal of Intelligent Engineering and Systems, Vol.15, No.3, 2022 DOI: 10.22266/ijies2022.0630.41
diffuse kinds of pooling strategies. It separates the
image into sub-region rectangles and only returns
the largest value of each. The maximum pooling
computations can be demonstrated using Eq. (2).


 (2)
Where:  is an output of pooling layer,  
 
 
 is height of pooled channel,   

 
 is width pooled channel, 0 ≤ y < yl = yl-1, ph
is the height of the pooled window and pw is width
of the pooled window.
The last layer is the fully-connected layer where
each node in it is directly related to every node in
the layers above and below it. A fully connected
layer also contains a large number of parameters that
necessitate complex training computing. As a result,
the dropout technique can be used to minimize the
number of nodes and connections in a network [16].
Eq. (3) shows how this layer is calculated [11].


󰇛󰇜     









 (3)
Where:  is an output of fully connected layer,

 is width of previous channel, 
 is height of
previous channel, 
 is number of previous
channels, 󰇛󰇜 is vector of pooling layer outputs,

 is the weights between the pooling and fully
connected layers, and ml is the required number of
categories.
Figure. 3 The standard CNN structure
3.4 Hybrid network
A hybrid model is made up of two or more
generic machine learning models that are combined
to improve the model’s overall performance.
Typically, a single machine learning algorithm is
created for a specific purpose. When two or more
algorithms are integrated, the hybrid model’s
performance improves dramatically. CNN+GAN,
CNN+AE, GAN+RL...etc. examples to hybrid
models [15].
4. The proposed palmprint authentication
technique
The suggested palmprint authentication
technique is described in this part, and it comprises
of two primary stages: region of interest (ROI)
extraction and feature extraction and matching using
the hybrid model. The proposed palmprint
authentication technique is shown in Fig. 4.
4.1 Region of interest extraction
The ROI of palmprint images must be split
before characteristics can be extracted. For the
region of interest extraction approach in this paper,
we used the method described by [17]. To smooth
down the noise in the input palmprint, the low-pass
Gaussian filter is applied initially. The palmprint
border is calculated using a boundary tracking
algorithm after the smoothed image is transformed
to a binary image by direct thresholding. Third,
reference points are located at the bottom of the
gaps among the index and middle fingers, as well as
the ring and little fingers. Fourth, the perpendicular
bisector of the line segment among two reference
landmarks is set up to locate the point of ROI.
Finally, at a specific place, the subimage is cropped
and scaled [17,18]. The ROI of the palmprint image
in this paper is 192×192 pixels. The ROI steps can
be briefed as illustrated in Fig. 5.
Figure. 4 The suggested palmprint authentication technique

Received: February 24, 2022. Revised: April 1, 2022. 492
International Journal of Intelligent Engineering and Systems, Vol.15, No.3, 2022 DOI: 10.22266/ijies2022.0630.41
Figure. 5 The region of interest extraction steps
Figure. 6 The proposed autoencoder model
4.2 Autoencoder model training
The many layers that appear in the proposed
autoencoder-based initial model are described in
depth in this subsection. We chose the deep
convolutional autoencoder because we are dealing
with images. There are multiple layers to the model.
The input layer of the encoder is set to receive a
grayscale ROI palmprint image with a resolution of
192×192 pixels. After that, the convolution layer is
used. There are three convolution layers proposed.
The activation function of each convolution layer is
a rectifier linear unit (ReLU). After each
convolution layer, the pooling layer is applied.
Windowing and maximum operations are employed.
During the decoding process the encoder employs
three convolutional transpose layers, which is the
same number as the encoder’s convolutional layers.
A rectifier linear unit (ReLU) activation function is
present in each convolutional transpose layer. Each
convolution transpose layer is followed by an
upsampling layer. The output layer is set to create a
grayscale ROI palmprint image with the same
dimensions as the input image. The autoencoder
model is used to train the network to extract
powerful features, and then the trained encoder part
is combined with the CNN network to create the
Figure. 7 The reconstructed images
suggested authentication method. Table 1
summarizes the proposed autoencoder architecture.
In addition, the following Fig. 6 depicts the structure
of the suggested autoencoder model.
Furthermore, the suggested autoencoder model’s
reconstructed images are shown in Fig. 7,
demonstrating that the features derived from the
reconstructed image using the autoencoder are
expected to be robust and important.
4.3 Feature extraction and matching via hybrid
model
The proposed hybrid model, which we
developed and built exclusively for palmprint
images, is described in depth in this subsection. The

Received: February 24, 2022. Revised: April 1, 2022. 493
International Journal of Intelligent Engineering and Systems, Vol.15, No.3, 2022 DOI: 10.22266/ijies2022.0630.41
Table1. The suggested autoencoder architecture
Layer Kind
Filter size
No. of filter
Input Shape
Output Shape
No. of parameter
Convolution 1
3x3
32
192,192,1
192,192,32
320
Max pooling 1
2x2
-
190,190,32
96,96,32
0
Convolution 2
3x3
16
96,96,32
96,96,16
4624
Max pooling 2
2x2
-
96,96,16
48,48,16
0
Convolution 3
3x3
16
48,48,16
48,48,16
2320
Max pooling 3
2x2
-
48,48,16
24,24,16
0
Conv transpose 1
3x3
16
24,24,16
24,24,16
2320
Upsampling 1
2x2
-
24,24,16
48,48,16
-
Conv transpose 2
3x3
16
48,48,16
48,48,16
2320
Upsampling 2
2x2
-
48,48,16
96,96,16
-
Conv transpose 3
3x3
32
96,96,16
96,96,32
4640
Upsampling 3
2x2
-
96,96,32
192,192,32
-
Output
3x3
1
192,192,32
192,192,1
289
Total parameters: 16,833
Trainable parameters: 16,833
Non-trainable parameters: 0
Table2. The suggest hybrid model architecture
Layer Kind
Filter
size
Activation
Function
No. of
filter
FC units
Layer
Input
Shape
Output
Shape
No. of
parameter
Encoder
Convolution 1
3x3
ReLU
32
-
192,192,1
190,190,12
320
Max pooling 1
2x2
-
-
-
190,190,32
96,96,32
0
Convolution 2
3x3
ReLU
16
-
96,96,32
96,96,16
4624
Max pooling 2
2x2
-
-
-
96,96,16
48,48,16
0
Convolution 3
3x3
ReLU
16
-
48,48,16
48,48,16
2320
Max pooling 3
2x2
-
-
-
48,48,16
24,24,16
0
CNN
Convolution 4
3x3
LeakyReLU
8
-
24,24,16
22,22,8
1160
Max pooling 4
4x4
-
-
-
22,22,8
5,5,8
-
Flatten
-
-
-
-
-
200
-
Fully connected
-
Leaky ReLU
-
512
-
-
102912
Dropout (rate = 0.5)
-
-
-
512
-
-
0
Fully connected
-
Softmax
-
163
-
-
83619
Total params: 194,955
Trainable params: 187,691
Non-trainable params: 7,264
Figure. 8 The proposed hybrid model
encoder element of the proposed autoencoder model
and a proposed convolutional neural network make
up the model. We first used the encoder portion,
which had been trained and prepared to accept a
grayscale ROI palmprint image with a size of
192×192 pixels and extract the majority of the

Received: February 24, 2022. Revised: April 1, 2022. 494
International Journal of Intelligent Engineering and Systems, Vol.15, No.3, 2022 DOI: 10.22266/ijies2022.0630.41
features. Then the proposed convolutional network
is added. There are numerous layers to CNN: In the
beginning, the convolution layer which has a
rectifier linear unit (leaky ReLU) activation function,
which is based on the ReLU notion as well. There is
a difference only when the input value is negative.
Leaky ReLU multiplies the value by a small integer
α instead of zeroing it like ReLU does (generally
0.01). As a result, the negative part gains a value,
although a very small one. It’s an attempt to solve
the dying ReLU problem. After the convolution
layer, the pooling layer is applied. Windowing and
maximum operations are employed. After that, a
fully connected layer is applied. This layer adjusts
for the number of nodes in the before layer and the
number of categories needed. After the fully
connected layer, we add a dropout layer to produce a
model that has better generalization and is less prone
to overfitting the training data. Finally, we add a
final layer that consists of softmax functions, which
are widely used to solve multiple categorization
problems. This layer has been modified to accept the
number of categories. The appropriate size of the
suggested hybrid model design is determined
empirically by gradually modifying the number of
convolutions and max-pooling, then the number of
filters, and finally selecting the network with the
highest performance. Table 2 summarizes the
proposed hybrid model design. The proposed hybrid
model’s structure can be described by the Fig. 8.
5. Experimental results
The suggested palmprint authentication
technique is built using Google Colab, which
provides Python3 in a free environment, using a
GPU Nvidia Tesla k80 12GB processor with 13GB
of RAM. Furthermore, the region of interest is
extracted using MATLAB v13.a. The performance
of the recommended technique is evaluated using
the COEP palmprint database, taken from the
college of engineering, Pune-411005, and the
proposed authentication technique as a whole is
evaluated in terms of accuracy
5.1 COEP dataset
According to the file’s attribute detail, the palm
images in this database were captured with a Canon
PowerShot SX120 IS camera with a resolution of
1600×1200×3 pixels and an image density of 180
dots per inch (DPI). The database took a year to
compile, and the download collection has 1304
photos of 163 palms, each containing eight images.
Fig. 9 shows palmprint samples from the COEP
dataset.
Figure. 9 Samples of palmprint images from the COEP
database
Figure. 10 Samples of region of interest
5.2 Region of interest extraction results
We utilized [17] to extract ROIs from 1304
images from the COEP dataset, with each extracted
ROI image having a size of 192×192 pixels and
resulting in a grayscale image. Fig. 10 shows a
variety of ROI palm print images.
5.3 Evaluation metrics
The proposed technique is assessed using
accuracy and F1-score metrics in this study. Eq. (4)
defines the first metric, which is accuracy [19, 20].
ACC= TP+TN
TP+ TN + FP + FN (4)
Where TP (true positives) is the numeral of
palmprints that are classified correctly and
recognized, TN (true negatives) represents the
numeral of palmprints that are negatively classified,
FP (false positives) denotes the numeral of
palmprints that are classified incorrectly, and FN
(false negatives) indicates the numeral of palmprints
that are identified as misclassified. The harmonic
mean of precision (P) and recall (R) is the F1-score,
which is calculated as follows Eq. (5).
F1_score= 2×P×R
P+R (5)
Precision and recall are two evaluation metrics
applied to assess the effectiveness of a procedure,
and they are represented in Eqs. (6) and (7) [20].
P= TP
TP + FP (6)
R= TP
TP+FN (7)

Received: February 24, 2022. Revised: April 1, 2022. 495
International Journal of Intelligent Engineering and Systems, Vol.15, No.3, 2022 DOI: 10.22266/ijies2022.0630.41
Table 3. The performance results of assessing the proposed technique
No. of filters
in Conv
Maxpooling
Activation
function
No. of hidden
layers
Accuracy
Precision
Recall
F1_score
16
4x4
Leaky ReLU
1x512
96.93%
98.43%
96.32%
96.61%
12
4x4
Leaky ReLU
1x512
95.71%
96.58%
95.40%
95.44%
8
4x4
Leaky ReLU
1x512
97.85%
97.53%
96.93%
96.81%
8
6x6
Leaky ReLU
1x512
93.25%
95.81%
91.10%
91.94%
8
8x8
Leaky ReLU
1x512
75.77%
93.68%
54.60%
57.91%
8
2x2
Leaky ReLU
1x512
95.09%
96.88%
95.09%
95.15%
8
4x4
ReLU
1x512
92.02%
95.81%
91.10%
91.88%
8
4x4
ELU
1x512
96.01%
97.20%
95.71%
95.71%
8
4x4
Swish
1x512
94.17%
95.02%
93.56%
93.60%
8
4x4
Leaky ReLU
1x256
95.40%
96.26%
94.79%
94.74%
8
4x4
Leaky ReLU
2x256
95.40%
97.16%
94.48%
94.81%
5.4 Evaluation of the hybrid model via accuracy
and F1-score
The ROI was retrieved and divided into 978
training palmprint images and 326 testing palmprint
images. After that, the network is trained on Keras, a
prominent deep learning platform. Many tests were
built and analyzed to find the proper autoencoder
model parameters. After obtaining the best results,
the coding part was used in the hybrid model with
CNN. In hybrid model, the RMSprop algorithm was
used with a learning rate of 0.001, and 100 epochs
of training. Many experiments were set up and
analyzed in order to determine the right CNN
parameters in the hybrid model. After multiple
rounds of training, the best accuracy rate was
97.85 % when the number of filters in the
convolutional layer was 8, with maxpooling 4×4.
The results of evaluating various CNN part
parameters are shown in Table 3. The parameters of
the convolutional and pooling layers are evaluated
one by one in this table, by tweaking one parameter
and settling the values of the other parameters. The
proposed technique’s assessment metrics are
accuracy, precision, recall, and F1-score. Table 4
shows the changing learning rate values in the
RMSprop algorithm and their effect on assessment
results. Fig. 11 depicts the relationship between the
proposed network’s accuracy and the number of
iterations used during the training phase. Also
illustrated Fig. 12 the relationship between the
proposed network’s loss function
(categorical_crossentropy) and the number of
iterations used during the training phase.
From the above figure, we can notice the start of
the accuracy with a dramatic rise. After that, the
indicator begins to gradually increase at epoch 40,
and the accuracy of access is relatively stable after
epoch 80.
Table 4. The performance results of changing the learning
rate values in RMSprop technique
Learning
rate
Accuracy
Precision
Recall
F1_score
0.0001
39.57%
100%
0.61%
0.61%
0.0005
93.25%
95.51%
91.41%
92.45%
0.001
97.85%
97.53%
96.93%
96.81%
0.005
95.40%
95.96%
94.79%
94.76%
0.01
95.09%
95.38%
95.09%
94.67%
0.05
84.05%
84.05%
84.05%
82.73%
0.1
86.20%
86.20%
86.20%
84.14%
0.5
91.10%
91.10%
91.10%
90.43%
Figure. 11 The accuracy of the proposed model
Figure. 12 The loss function of the proposed model

Received: February 24, 2022. Revised: April 1, 2022. 496
International Journal of Intelligent Engineering and Systems, Vol.15, No.3, 2022 DOI: 10.22266/ijies2022.0630.41
Table 5. comparison of the proposed model with the previous models
Method
No. Of con.
layers
No. of FC
layers
Dataset
No. of images
No. of cases
Accuracy
CNN [8]
11
2
PolyU
6000
500
99.95%
CNN [10]
5
3
PolyU
Multi-Spectral
24000
500
99.99%
CNN [11]
1
3
PolyUC
1000
100
97.67%
The proposed AE+CNN
4
2
COEP
1304
163
97.85%
From the above figure, we can notice the start of the
loss function with a large drop. After that, the
indicator begins to steadily fall at epoch 40, and the
loss function is relatively stable after epoch 80.
To evaluate the proposed hybrid model, we
compared it with recent developed works, and the
comparison contained two parts. The first part was a
comparison with some deep learning methods in
terms of the number of convolutional layers used,
the number of fully connected layers, the database
used, its size, and the attained accuracy of the
method. The results of the proposed technique were
promising despite the small number of images used
for the database, which is considered an important
characteristic of our proposed technique. The
comparison showed that the proposed technique was
well-organized in terms of building the model, as it
does not use a large number of layers, which leads
to the computational complexity of the model and
increases the training time. At the same time, our
model does not resemble a simplistic model with a
single layer that may not be capable of extracting
deep features. Table 5 shows the details of the
comparison results. The other part of the comparison
was with the traditional methods that used the COEP
database used in our research, noting that the
database was not used with deep learning methods,
and we believe that this is due to the small number
of images in it, which is one of the biggest
challenges in deep learning techniques, as we
showed the method used and the results obtained
compared to the proposed technique in Table 6. The
proposed technique also has an additional advantage,
which is to ensure that the features extracted from
the palm print are important features and able to be
distinguished with high accuracy. This is based on
the autoencoder model used in the hybrid model that
uses the extracted features for the purpose of
reconstructing the image with high accuracy, which
can highlight the strength of the proposed model on
this side as we proved in Fig. 13 which provided the
gradient weighted class activation map (Grad-CAM)
for the entire model.
Table 6. comparison of the proposed model with the
classical methods
Method
Accuracy
Stockwell transform [7]
100%
PRS [4]
99.7%
The proposed AE+CNN
97.85
(a) (b) (c)
Figure. 13 Illustration of the GRAD-CAM from the
model: (a) original image, (b) points of interest in the
final convolutional layer, and (c) Softmax activation
output
5.5 Evaluate by Grad-CAM result
To show regions of interest in the final
convolutional layer and the softmax activations
output from the model, we developed a gradient
weighted class activation map (Grad-CAM). As
shown in Fig. 13, the GRAD-CAM has identified
the specific areas of the palmprint that CNN finds
relevant for class differentiation. The network
concentrates on the parts highlighted in yellow,
which represent the most relevant spots in the palm
image, and red, which signifies the least important
points, in order to make the classification decision.
5.6 Using the feature maps technique to visualize
convolution layers
The feature maps technique is used to display
the applied filters in this section. The first six filters

Received: February 24, 2022. Revised: April 1, 2022. 497
International Journal of Intelligent Engineering and Systems, Vol.15, No.3, 2022 DOI: 10.22266/ijies2022.0630.41
are displayed in Fig. 14. Each filter is represented by
a figure with six rows of three images. Dark region
symbolizes little or inhibitory weights, while light
region represents large or excitatory weights.
We also use the feature maps method to evaluate
and understand the model’s predictions. This way
Figure. 14 Feature map technique for first six filters
Figure. 15 An illustration of the first convolution layer,
which uses the feature maps technique to interpret 32
filters.
Figure. 16 An illustration of the fourth convolution layer,
which incorporates interpret eight feature maps-based
filters.
helps to understand how the model learns different
filters and how data is transferred through the layers.
The premise is that feature maps near the input
identify small or fine-grained details, while feature
maps near the model’s output capture more generic
information. The visualized feature maps for the
first convolution layer, which contains 32 filters, are
shown in Fig. 15. We can see from the figure that
the consequence of applying the filters in the first
convolutional layer gives different sorts of features.
Fig. 16 also depicts feature maps for the fourth
convolution layer, which includes eight filters.
Where the feature maps display fewer information,
we can tell that the model works deeply. This
context is expected at this convolution level, but it
has the potential to provide useful features for
categorization. We lose our capacity to read these
deeper feature maps in most cases, yet they are very
apparent for the model.
6. Conclusion
In this work, we developed a palmprint
authentication technique (AE+CNN) based on deep
learning approaches, specifically a hybrid of
autoencoder and CNN. With the suggested approach,
we obtained deep features from the hybrid model
that led to valuable authentication results with a
smaller number of images, which is a major issue in
any deep learning model. We used the COEP dataset,
which has never been used before with palmprint
authentication based on deep learning. Several
experiments have been conducted to see how the
parameters of various network hybrid layers evolve
over time. The network has shown that it can adapt
to a wide variety of palmprints, resulting in strong
authentication. We compared our findings with
conventional approaches and deep learning models.
In compared to other methodologies, the best results
obtained were 97.85 % accuracy and 96.81 % F1-
score, which are excellent findings. In future work,
we suggest adopting deep learning networks to be
used in different stages of the palmprint
authentication process, from ROI extraction to
classification. We came to the conclusion that not all
deep learning models require a large number of
images to provide effective results.
Conflicts of interest
The authors declare no conflict of interest.

Received: February 24, 2022. Revised: April 1, 2022. 498
International Journal of Intelligent Engineering and Systems, Vol.15, No.3, 2022 DOI: 10.22266/ijies2022.0630.41
Author contributions
The paper conceptualization, Mohammed and
Firas; methodology, Firas; software, Firas;
validation, Mohammed and Firas; formal analysis,
Firas; investigation, Mohammed; resources, Firas;
data curation, Mohammed and Firas; writing—
original draft preparation, Mohammed and Firas;
writing—review and editing, Mohammed;
visualization, Firas; supervision, Mohammed;
project administration, Mohammed.
References
[1] Z. Guo, D. Zhang, L. Zhang, and W. Zuo,
“Palmprint verification using binary orientation
co-occurrence vector”, Pattern Recognition
Letters, Vol. 30, No. 13, pp. 1219-1227, 2009.
[2] S. Verma and P. Mishra. “A survey paper on
palm prints based biometric authentication
system”, International Journal of Electrical
and Electronics Engineering (IJEEE), Vol. 1,
No. 3, pp. 20-27, 2012.
[3] L. Fei, Y. Xu, and D. Zhang. “Half-orientation
extraction of palmprint features”, Pattern
Recognition Letters, Vol. 69, pp. 35-41, 2016.
[4] M. Kadhm, H. Ayad, and M. Mohammed,
“Palmprint Recognition System Based on
Proposed Features Extraction and (C5. 0)
Decision Tree, K-Nearest Neighbour (KNN)
Classification Approaches”, J. Eng. Sci.
Technol, Vol. 16, No. 1, pp. 816-831, 2021.
[5] M. Mu, Q. Ruan, and Y. Ming, “Shape
parameters of Gaussian as descriptor for
palmprint recognition based on dual-tree
complex wavelet transform”, In: Proc. of IEEE
10th International Conf. on SIGNAL
PROCESSING PROCEEDINGS, Beijing,
China, pp. 1406-1409, 2010.
[6] B. Ergen, “Scale invariant and fixed-length
feature extraction by integrating discrete cosine
transform and autoregressive signal modeling
for palmprint identification”, Turkish Journal
of Electrical Engineering & Computer
Sciences, Vol. 24, No. 3, pp. 1768-1781, 2016.
[7] N. Kumar and K. Premalatha, “Palmprint
authentication system based on local and global
feature fusion using DOST”, Journal of
Applied Mathematics, Vol. 2014, pp. 1-11,
2014.
[8] X. Dong, L. Mei, and J. Zhang, “Palmprint
recognition based on deep convolutional neural
networks”, In: Prof. of 2018 2nd International
Conf. on Computer Science and Intelligent
Communication (CSIC 2018), Leipzig,
Germany, pp. 82-88, 2018.
[9] G. Wang, W. Kang, Q. Wu, Z. Wang, and J.
Gao, “Generative adversarial network (GAN)
based data augmentation for palmprint
recognition”, In: Proc. of IEEE International
Conf. 2018 Digital Image Computing:
Techniques and Applications (DICTA),
Canberra, ACT, Australia, pp. 1-7, 2018.
[10] W. Gong, X. Zhang, B. Deng, and X. Xu,
“Palmprint recognition based on convolutional
neural network-AlexNet”, In: Proc. of 2019
Federated Conf. On Computer Science and
Information Systems (FedCSIS), Leipzig,
Germany, pp. 313-316, 2019.
[11] L. Albak, R. A. Nima, and A. Salih, “Palm
print verification based deep learning”,
Telkomnika, Vol. 19, No. 3, pp. 851-857, 2021.
[12] J. Mehta and A. Majumdar, “Rodeo: robust de-
aliasing autoencoder for real-time medical
image reconstruction”, Pattern Recognition,
Vol. 63, pp. 499-510, 2017.
[13] C. Chaitanya, A. Kaplanyan, C. Schied, M.
Salvi, A. Lefohn, D. Nowrouzezahrai, and T.
Aila, “Interactive reconstruction of Monte
Carlo image sequences using a recurrent
denoising autoencoder”, ACM Transactions on
Graphics (TOG), Vol. 36, No. 4, pp. 1-12,
2017.
[14] J. Zheng and L. Peng, “An autoencoder-based
image reconstruction for electrical capacitance
tomography”, IEEE Sensors Journal, Vol. 18,
No. 13, pp. 5464-5474, 2018.
[15] M. Fuad, A. Fime, D. Sikder, M. Iftee, J. Rabbi,
M. A. Rakhami, A. Gumae, O. Sen, M. Fuad,
and M. Islam, “Recent Advances in Deep
Learning Techniques for Face Recognition”,
IEEE Access, Vol. 9, pp. 99112-99142, 2021.
[16] S. Albawi, T. Mohammed, and S. A. Zawi,
“Understanding of a convolutional neural
network”, In: Proc. of 2017 International Conf.
On Engineering and Technology (ICET),
Antalya, Turkey, pp. 1-6, 2017.
[17] W. Li, B. Zhang, L. Zhang, and J. Yan,
“Principal line-based alignment refinement for
palmprint recognition”, IEEE Transactions on
Systems, Man, and Cybernetics, Part C
(Applications and Reviews), Vol. 42, No. 6, pp.
1491-1499, 2012.
[18] L. Fei, G. Lu, W. Jia, S. Teng, and D. Zhang,
“Feature extraction methods for palmprint
recognition: A survey and evaluation”, IEEE
Transactions on Systems, Man, and
Cybernetics: Systems, Vol. 49, No. 2, pp. 346-
363, 2018.

Received: February 24, 2022. Revised: April 1, 2022. 499
International Journal of Intelligent Engineering and Systems, Vol.15, No.3, 2022 DOI: 10.22266/ijies2022.0630.41
[19] P. Dhandapani and A. Varadarajan, “Multi-
Channel Convolutional Neural Network for
Prediction of Leaf Disease and Soil Properties”,
International Journal of Intelligent
Engineering and Systems, Vol. 15, No. 1, pp.
318-328, 2022.
[20] S. Behdenna, B. Fatiha, and G. Belalem,
“Ontology-Based Approach to Enhance
Explicit Aspect Extraction in Standard Arabic
Reviews”, International Journal of Computing
and Digital Systems, Vol. 11, No. 1, pp. 277-
287, 2022.