Content uploaded by Ebenezer Esenogho
Author content
All content in this area was uploaded by Ebenezer Esenogho on Nov 04, 2023
Content may be subject to copyright.
2023 IEEE AFRICON
979-8-3503-3621-4/23/$31.00 ©2023 IEEE
AI-based Palm Print Recognition System for High-
security Applications
Abraham S. Martey
Department of Electrical and
Electronics Engineering Technology
University of Johannesburg
Johannesburg, South Africa
martey235@gmail.com
Ahmed Ali
Department of Electrical and
Electronics Engineering Technology
University of Johannesburg
Johannesburg, South Africa
aali@uj.ac.za
Esenogho Ebenezer
Department of Electrical and
Electronic Engineering
University of Botswana
Gaborone, Botswana
drebenic4real@gmail.com
Abstract— In recent years, many studies have failed to
implement an effective palm print recognition system for high-
security applications. This study focuses on developing a novel
palm print recognition system using novel data processing
techniques. The study proposes an embedded zero-tree wavelet
(EZW) and principal component analysis (PCA) feature
extraction technique concerning palm print recognition. The
database contains palm print image samples from right and
left palm images. 200 images of 5 people were captured with each
person, and 40 shots were used. 150 images were used in the SVM
training, and 50 images were used in the SVM testing. The
spectral feature extraction of the palm print image is processed by
the EZW. The spatial feature extraction of the palm print image is
processed by PCA. The minimum distance classifier is used for
the comparison of results. Finally, the palm print images are
trained and classified with Support Vector Machine (SVM). The
researcher concluded that, when compared to the other evaluated
approaches and classifiers, the palm print recognition system that
combines EZW and PCA as a method of feature extraction is the
most accurate. The overall testing results show that the proposed
approach yields a maximum of 90.4% recognition accuracy.
Keywords— Palm print recognition, EZW, PCA
I. INTRODUCTION (HEADING 1)
Identification and verification have become crucial
matters in today’s world. Passwords are frequently used;
however, this authentication method is insufficient in many
cases. Since passwords are easily lost or stolen, they have
multiple restrictions. Hence, many studies have been
undertaken to discover optimal plans to get around them. In
state-of-the-art biometrical engineering, which has
significantly enhanced human authentication and security [1],
human verification is carried out in accordance with the proper
physiological or behavioural characteristics. Palm prints,
irises, and fingerprints are examples of physiological
characteristics, while voice, signature, and gait represent
behavioural characteristics. In contemporary applications
such as law enforcement, e-commerce applications, laptops,
and mobile phones, biometrics stand as the most efficient and
secure solution for human authentication. Its excellent
capability for security authentication has made the palm print
recognition an increasingly important topic of study over the
past few decades [2]. The intricate morphology of palm prints
has given rise to this high level of security. These
distinguishing characteristics include ridges, wrinkles, and
principal lines specific to every person [3].
A pattern recognition tool termed a biometrics system is
used to verify the user's authentication with regard to
particular behavioural or physiological characteristics [4].
This is concerned with categorising individuals based on their
physiological factors, including their face, fingerprint, palm
print, iris, or different traits such as their signature or voice
[5]. Due to this, it has been regarded as a promising topic of
study. Palm print features can be divided into two groups
according to the industries in which palm print systems are
employed. Wrinkles and principal lines make up the first
group of features; these are observable and simple to obtain
because they can be extracted from images with low resolution
(less than 100 dpi) and used in commercial applications.
Fig 1: Palm print features [6]
The minutiae point, ridges, and singular point, which make
up the other group of features, are particular details that must
be extracted from the images with high resolution (greater
than 100 dpi). This group can be employed in forensic
applications such as those related to law enforcement [7]. The
low- and high-resolution features in a palm print are depicted
in Figure 1.
This study examines the reliability of various feature
extraction and palm print identification methods to determine
which is the most reliable and accurate when used on a
specific dataset. This study aims to determine the most
appropriate feature extraction and palm print identification
methods with regard to cost, resources required, simplicity,
and accuracy. The study compares the outcomes of the highly
popular feature extraction and palm print identification
methods on the same dataset. To give the images high
visibility and ensure that they required less work during the
stages involved in feature extraction, the author added a pre-
processing step to the images. The study also includes an
intensive literature survey on palm print recognition studies
and the importance of embedded zero-tree wavelet (EZW) and
principal component analysis (PCA) feature extraction
techniques.
2023 IEEE AFRICON | 979-8-3503-3621-4/23/$31.00 ©2023 IEEE | DOI: 10.1109/AFRICON55910.2023.10293345
Authorized licensed use limited to: University of Botswana. Downloaded on November 04,2023 at 16:10:49 UTC from IEEE Xplore. Restrictions apply.
2023 IEEE AFRICON
II. RELATED WORK
A. Studies on palm print recognition system's classification
techniques
A study by Ito et al. focused on the pre-processing stage
for palm print images to detect the palm region for the
classification and feature extraction stages. The authors noted
that this stage would impact the recognition system's accuracy
[8]. To obtain high precision, the authors had to pay attention
to this stage. This study's local dataset, which included 1,932
palm print images, was used. Additionally, there was a
presumption that the right and left hands were obtained from
unique sources. At the end of the study, their system was
assessed by applying the CASIA Palmprint Image Dataset to
demonstrate its efficacy [8]. However, most studies utilised
any one palm print image while presenting the techniques for
recognising palm prints. A study by Xu et al. used two palm
print images for the right hand. They held hands to compare
the accuracy they obtained individually from the palm prints.
Scholars use three different types of scores to match these
palm print images. The identification technique has been used
in the first two sources, while the last uses a unique algorithm.
The two palm print images are the main aim of this algorithm.
The results of this study showed how similar they were [9].
B. Studies on palm print recognition system's feature
extraction techniques
Rajput proposed a palm print recognition system that
utilises the fewest resources possible. Fourier descriptors,
Kekre's fast codebook generation, and the discrete cosine
transform are the algorithms used to implement this system.
Additionally, Rajput’s study compared the outcomes of these
algorithms. This study used a database that contained 180
images with a resolution of 640*480 pixels; three of those
images were utilised for testing and an additional three were
used for the training process for each of the 18 people. The
Fourier descriptors, Kekre's Fast Codebook Generation, and
the discrete cosine transform had recognition rates of 66.25%,
88.89%, and 92.85%, respectively, for this system [10]. A
study by Cappelli et al. used similar steps in fingerprint
recognition to present a new palm print recognition system.
However, these steps were optimised to serve as feasible
systems for quickly recognising palm prints and obtaining
more accuracy. With fewer forged features than in any
scenario, these steps helped to identify the majority of features
of the palm print. This system achieved high accuracy
compared to earlier fieldwork [11].
Sun and Abdulla used the hamming distance of the
classification step and the curvelet transform in the feature
extraction step to present a palm print recognition system. The
Hong Kong PolyU multispectral palm print database was used
to test and train systems. This database had 500 unique palm
images and six palms per image. The recognition rate used
various curvelet coefficients, resulting in a matching error rate
of 0.73 percent and scales 2 and 3, with storage equivalent to
900 bytes. As a suggestion for further research, the authors
provided suggestions to combine the outcomes of the three
scales' coefficients and create a system for categorising palm
print recognition [12].
A study by Shrivas et al. proposed a new identification
technique that used palm print images and depended on
feature extraction using the MATLAB Image Processing
Toolbox and classification using back propagation neural
networks. The recognition features were mean orientation,
primary axis length, convex area, solidity, extent, Equiv
diameter, Euler number, total area, and no_objects. Their
system had a 95.81 percent recognition rate with six epochs
and 50 neurons, a 98.62 percent rate with 14 epochs and ten
neurons, and a 99.99 percent rate with nine epochs and 20
neurons [13]. Elaydi et al. developed a palm print recognition
system using various techniques for feature extraction, such as
contourlet, curvelet, ridgelet, and 2D discrete wavelet. The
PolyU hyperspectral palm print database was used in this
study. Methodology
III. PROPOSED METHOD
This study proposes an EZW and PCA feature extraction
technique concerning palm print recognition. The database
contains palm print image samples from right and left palm
images. The image pre-processing steps have been applied.
The processes involved in the proposed methodology are RGB
to grayscale conversion, image resizing, and image
enhancement. The spectral feature extraction of the palm print
image is processed by the EZW. The spatial feature extraction
of the palm print image is processed by principal component
analysis PCA. The minimum distance classifier is used for the
comparison of results. Finally, the palm print images are
trained and classified using SVM. The resulting palm print is
enhanced, and the noise is removed during pre-processing.
Following pre-processing, the palm print undergoes two
stages of feature extraction, the first of which involves
producing a frequency response feature representation by
wavelet decomposition and the second of which involves a
spatial level using PCA. Finally, we trained and classified
using SVM. The vectors are combined in the following stage,
and the databases examine the result to calculate the minimal
distance that categorises the fingerprint according to the
classification it belongs to. A total of 200 photographs of 5
people were taken, each containing 20 left- and 20 right-hand
images. Of those images, 150 (15 left-hand and 15 right-hand
images of 5 people) were utilised for training, while the
remaining 50 (5 left- and five right-hand images of 5 people)
were used for testing.
A. Pre-processing of image data
A wide range of techniques are used in enhancement
procedures to successfully enhance each image's visual
quality. Additionally, these techniques support the conversion
of the image into a format more suitable for examination by
appropriate technologies. Depending on the needs of the
image, enhancement techniques such as histogram
equalisation, contrast enhancement, thinning, inverting, and
binarization are utilised. The palm print is then reduced in size
to 512 x 512. To improve the quality of the contours, image
enhancement is applied to the acquired palm print picture. The
grayscale input image is then transformed into a binary image.
The results from pre-processing are displayed in the results
and discussion sections.
Segmenting palm print pictures during the preprocessing
step prepares them for feature extraction during the following
stage. The fingerprints are processed further after
preprocessing. Improvement techniques vary depending on
the database and palm prints. The algorithms can be improved
with the help of the improved palm prints, which were of poor
quality. A grayscale digital picture that primarily contains
image information has a single context for each pixel's result
when analysing palm print photos. Images of this type,
Authorized licensed use limited to: University of Botswana. Downloaded on November 04,2023 at 16:10:49 UTC from IEEE Xplore. Restrictions apply.
2023 IEEE AFRICON
commonly referred to as black and white, mostly hold grey
content, ranging in intensity from white at the lightest to black
at the deepest. With the bitonal white and black pictures,
grayscale images are always displayed and are unique from
one bit to the next.
The black-and-white photographs were selected with
computer imaging in mind. Grayscale images feature a variety
of grayscale tones throughout. The light intensity for every
pixel inside a specific spectrum of electromagnetic waves is
frequently measured to create grayscale pictures. Grayscale
pictures can occasionally be monochromatic when a single
frequency is recorded. However, they may also be converted
from the colour image into a black-and-white image. The
procedure of expanding a digital image is called image
scaling. Effectiveness, accuracy, and smoothness must be
traded off in the non-trivial process of scaling. The pixels that
make up an image in bitmap graphics become more evident as
the dimension of the photograph is lowered or raised. When
considering vector graphics, the cost of re-rendering the image
might be computational power. When considering computer
animation's frame rate, frame skipping, and still graphics, the
representation of the image was visible. Digital photographs
are modified during the image enhancement process to
provide outcomes better suited for presentation and additional
image analysis. The working methodology of the proposed
system is shown in Figure 2.
Fig 2. Working methodology of the proposed palm print
recognition system
The proposed methodology for the current system is
shown in Figure 2. The SVM training and testing classifier
was used in this study. Initially, in this process, the left and
right palm prints undergo feature extraction processes. The
feature extraction process includes the EZW and PCA for the
SVM classification. After that, the comparison scores are
evaluated for palm print recognition.
B. Feature Extraction of Embedded Zero-tree Wavelet
(EZW) and Principal Component Analysis (PCA)
The pre-processed image is transformed into a set of
features using the feature extraction step. This procedure is
the fundamental phase in the proposed system's development
of features to differentiate between people's palm prints. Even
though there are several methods of feature extraction, the
researchers for this study chose the following: principal
component analysis (PCA) and embedded zero-tree wavelet
(EZW).
1). Embedded Zero-tree Wavelet (EZW)
An encoder created specifically for use with wavelet
transformations in image processing is known as an EZW
encoder. The word “wavelet” in its name refers to the wavelet
transform. Although the EZW encoding was initially intended
to work with photos and other two-dimensional data, it is also
used for communications in different dimensions. The EZW
encoder accurately uses progressive encoding to convert an
image into a digital signal. This EZW encoder is based on two
significant findings:
• In general, natural photos have a very low bandpass
filter spectrum. The intensity of the sub-bands drops as the
scale gets smaller whenever the image is converted into a
wavelet. Additionally, the wavelet coefficients in the higher
sub-bands would be lower and medium compared to the lower
sub-bands. As a result, wavelet-converted images were
naturally compressed using progressive encoding, and the
higher sub-bands added information to the image conversion.
• Wavelet coefficients with large values are more
significant than those with modest values.
The EZW encoding strategy uses these two findings by
repeatedly encoding individual variables in decreasing order.
A threshold is selected for each pass, and all variables are
evaluated. A wavelet coefficient was encrypted and deleted
from the photo if it exceeded the threshold value.
Figure 3. Inter-relationship among the sub-bands and
wavelet coefficients [14]
The threshold is reduced, and the photo is examined once
again to add extra information to the previously encrypted
photo once all wavelet coefficients are examined. Before all
wavelet coefficients were fully encoded, another requirement
of the image processing met the original requirement and
repeated this process. Substantial portions of the photo that
fall below the existing threshold can be effectively encoded by
taking advantage of the dependence among the wavelet
coefficients
throughout various scales. The location in time is best
described as the space-time position when the input is an
Authorized licensed use limited to: University of Botswana. Downloaded on November 04,2023 at 16:10:49 UTC from IEEE Xplore. Restrictions apply.
2023 IEEE AFRICON
image. Due to the sub-sampling during the wavelet transform,
researchers represent an image after it has been transformed
using trees. A quadtree called a zero-tree has nodes that are all
smaller than or equal to those of the base. The tree is encoded
with a single signal, and the decoder reconstructs it as a
quadtree made entirely of zeros. To further complicate this
condition, researchers must state that the root must be less than
the existing cut-off point used to evaluate wavelet coefficients.
Based on the knowledge that wavelet transforms drop with
scale, the EZW encoder uses the zero tree. It assumes that
when the base of a quadtree is below a given threshold, there
is a very strong likelihood that all variables will also be
smaller than all the cut-off points.
Furthermore, suppose the picture is examined under the
specified conditions and in a predetermined order, from the
higher level to the lower level. In that case, many points are
automatically coded using zero-tree signals. The zero-tree
constraint will undoubtedly be broken frequently, but overall,
the likelihood is still relatively high, as is shown in practice.
The cost is that the zero-tree sign is now part of the coding
language. The EZW encoder uses progressive encoding to
accurately compress a picture into a bit stream. This indicates
that when additional bits are introduced to the streaming, the
decoded image will have a comparable quality to JPEG-
encoded pictures because it will have more information.
2). Principal Component Analysis (PCA)
Principal component analysis (PCA) enables the
construction of part of the analysis by reducing the number of
correlated variables to a large number of variables. By a linear
transformation procedure, these are known as significant
components of image processing. The key features that follow
are essentially “linear mixtures of the original dataset”, which
effectively captures much of the volatility in the information.
Creating the following primary components from high-
dimensional data, the original characteristics can be
normalised by subtracting the sampling distribution, and the
result is divided by the standard deviation. PCA is a practical
and linear approach mostly adopted in statistical data analysis,
data processing, and image processing. It is a statistical
technique for reducing the dimensions of a heterogeneous
dataset for analysis, display, or data reduction. The framework
in PCA's representation of the data represents the patterns of
the data's highest variance. New fundamental vectors are
constructed and used for the specific data collection of the
palm prints.
The computing costs of PCA's versatility are significantly
greater than those of the fast Fourier transform (FFT). PCA is
among the statistical analyses often used in signal analysis for
data correlation or data feature reduction. Two different PCA
image-processing applications are discussed in this study. The
first use involves reducing the three colour components of a
picture into one that contains most of the details. The
eigenvalue characteristics are utilised in the second
application of PCA in order to determine the position of a
chosen palm print image. Several techniques may be applied
for earlier object detection. The effectiveness of the image
segmentation procedure also affects the outcome of the PCA-
based object orientation assessment described in the following
section.
3). Support Vector Machine (SVM)
The support vector machine (SVM) is among the most
recent learning tools that may be used for both pattern
classification and prediction. The unsupervised learning
models are considered to function using the associated
learning algorithms that carry out data processing and pattern
categorisation. Algorithms are employed in both classification
and regression analyses. Additionally, with positive
outcomes, SVMs were applied to curvatures, coefficients,
radial base function networks, and multi-layered perceptron.
Because they are based on the fundamental risk reduction
concept, the accuracy and level of difficulty of SVM solutions
are independent of the size of the input data.
IV. RESULTS AND DISCUSSION
This section discusses the overall level of palm print
recognition achieved through EZW and PCA in this study.
The analysis progressed with various research methodologies,
including the EZW and PCA feature extraction methods and
machine learning classifiers. The SVM classifier is processed
for the chosen palm prints concerning their recognition and
accuracy.
Fig 4. Input palm print image
Fig 5. Pre-processed palm print image
Fig 6. Reversed palm print image
Authorized licensed use limited to: University of Botswana. Downloaded on November 04,2023 at 16:10:49 UTC from IEEE Xplore. Restrictions apply.
2023 IEEE AFRICON
Figure 4 shows the input palm print image processed in the
system. The collected samples are clearly shown in the
MATLAB GUI platform. After that, the input image is pre-
processed; the pre-processed image is shown in Figure 5. The
pre-processed image is thoroughly analysed by processing its
orientation with the reversal image. The reversed palm print
image is shown in Figure 6.
Fig 7. EZW Palm print recognition
Fig 8. PCA results of palm print
Figures 7 and 8 show the EZW and PCA results of the
chosen palm print images. The image results show the impact
of feature extraction techniques on the palm print image. The
feature extraction techniques integrate the palm print image
into a two-dimensional array of co-efficient. The PCA
resolves the dimensionality issue.
Fig 9. Palm print recognition result of Person 1
Fig 10. Palm print recognition result of Person 2
Fig 11. Palm print recognition result of Person 3
Figures 9, 10, and 11 show the palm print recognition
results of Persons 1, 2, and 3. The process of palm print
recognition involves selecting the input image, pre-
processing, reversing the image, and conducting feature
extraction using EZW and PCA. At the end of the process, the
SVM classifier identifies the given person's palm print. Based
on the chosen operations, the palm print is recognised
effectively, with a maximum accuracy of 90.4%.
Fig 12. Confusion matrix results of palm print recognition
TABLE I
Comparison results of palm print recognition
Algorithm
Accuracy (%)
Hamming distance
64.46
Euclidean distance
70.00
Point-wise matching
75.83
MDC
81.67
Proposed system
90.4
Authorized licensed use limited to: University of Botswana. Downloaded on November 04,2023 at 16:10:49 UTC from IEEE Xplore. Restrictions apply.
2023 IEEE AFRICON
The confusion matrix results are shown in Figure 12.
These results show the impact of the proposed system on the
five people included in the study. Two hundred images were
captured for each of those five people; out of those 150 images
(15 left-hand and 15 right- hand images of five people), 40 (20
left-hand and 20 right-hand images) were used for training and
50 (five left and five right hands of 5 people) were used for
testing. The overall testing results show that the proposed
system yields a maximum percentage of 90.4% recognition
accuracy. Comparison research was also conducted in this
study. Other algorithms such as the Hamming distance
algorithm, the Euclidean distance algorithm, the point-wise
matching algorithm, and the MDC algorithm were also
processed in this study. The comparison results show that the
Hamming distance algorithm achieved 64.46% recognition
accuracy, the Euclidean distance algorithm achieved 70.00%
recognition accuracy, and the point-wise matching algorithm
achieved 75.83% recognition accuracy. The MDC algorithm
achieved 81.67% recognition accuracy, and our proposed
system yields a maximum recognition accuracy of 90.4%.
V. CONCLUSION
This study proposes effective machine learning techniques
by implementing effective feature extraction techniques. The
researcher concludes that, compared to the other evaluated
approaches and classifiers, the palm print recognition system,
which combines EZW and PCA for feature extraction, is more
accurate than other methods. The algorithm was only used to
generate prints that were optically scanned. According to our
dataset, the proposed method's maximum success rate for
palm print identification is about 90.4%. The proposed
algorithms have the main advantages of high recognition
accuracy and greater concentration on both spatial and
spectral features.
REFERENCES
[1] M. Trikoš, I. Tot, J. Bajčetić, K. Lalović, B. Jovanović, and D.
Bogićević, "Biometric security standardization," in 2019 Zooming
Innovation in Consumer Technologies Conference (ZINC), Novi Sad,
Servia, May 2019, pp. 17–20.
[2] P. Drozdowski, C. Rathgeb, A. Dantcheva, N. Damer, and C . Busch,
"Demographic bias in biometrics: A survey on an emerging challenge,"
IEEE Trans. Technol. Soc., vol. 1, no. 2, pp. 89–103, 2020.
[3] G. Brostoff, "How AI and biometrics are driving next- generation
authentication," Biom. Technol. Today, vol. 2019, no. 6, pp. 7–9, 2019.
[4] Ahmed Ali, Sanjeevikumar Padmanaban, Bhekisipho Twala,
Tshilidzi Marwala, "Electric Power Grids Distribution Generation
System For Optimal Location and Sizing-A Case Study Investigation
by Various Optimization Algorithms", Energies Journal, vol. 10, no. 7,
2017
[5] D. Zhang, W.-K. Kong, J. You, and M. Wong, "Online palmprint
identification," IEEE Trans. Pattern Anal. Mach. Intell., vol. 25, no. 9,
pp. 1041–1050, 2003.
[6] H. Elaydi, M. A. Abukmeil, and M. Alhanjouri, "Palmprint recognition
using multiscale transform, linear discriminate analysis, and neural
network," Sci. J. Circuits Syst. Signal Process., vol. 2, no. 5, pp. 112–
118, 2013.
[7] M. Sharkas, I. El-Rube, and M. A. Mostafa, "The contourlet transform
with the principal component analysis for palmprint recognition," in
2010 2nd International Conference on Computational Intelligence,
Communication Systems and Networks, Liverpool, United Kingdom,
July 2010, pp. 262–267.
[8] K. Ito, T. Sato, S. Aoyama, S. Sakai, S. Yusa, and T. Aoki, "Palm
region extraction for contactless palmprint recognition," in 2015
International Conference on Biometrics (ICB), Phuket, Thailand, May
2015, pp. 334–340.
[9] Y. Xu, L. Fei, and D. Zhang, "Combining left and right palmprint
images for more accurate personal identification," IEEE Trans. Image
Process., vol. 24, no. 2, pp. 549–559, 2014.
[10] K. Y. Rajput, M. Amanna, M. Jagawat, and M. Sharma, "Palmprint
Recognition Using Image Processing," Int. J. Comput. Sci. Commun.
Technol., vol. 3, no. 2, pp. 618– 621, 2011.
[11] R. Cappelli, M. Ferrara, and D. Maio, "A fast and accurate palmprint
recognition system based on minutiae," IEEE Trans. Syst. Man Cybern.
Part B Cybern., vol. 42, no. 3, pp. 956–962, 2012.
[12] A. M. Farayola, Y. Sun, and A. Ali, "ANN-PSO Optimization of PV
systems under different weather conditions," in 7th International
Conference on Renewable Energy Research and Applications
(ICRERA), Paris, France, 2018
[13] R. Shrivas, N. Verma, and V. Singh, "Palmprint Biometrics using Feed
Forward Back Propagation Neural Network," Int. J. Comput. Appl.,
vol. 74, no. 3, 2013.
[14] C. N. Obiora, A. Ali and A. N. Hasan, "Estimation of Hourly Global
Solar Radiation Using Deep Learning Algorithms," 2020 11th
International Renewable Energy Congress (IREC), 2020, pp. 1-6, DOI:
10.1109/IREC48820.2020.9310381
Authorized licensed use limited to: University of Botswana. Downloaded on November 04,2023 at 16:10:49 UTC from IEEE Xplore. Restrictions apply.
