https://doi.org/10.31449/inf.v47i9.4675 Informatica 47 (2023) 63–80 63 
An Effective Hyperspectral Palmprint Identification System Based on 
Deep Learning and Band Selection Approaches 
Maarouf Korichi
1
, Djamel Samai
1
, Azeddine Benlamoudi
1
, Abdallah Meraoumia
2
, Khaled Bensid
1
 
1
Fac. des nouvelles technologies de l’information et de la communication. Lab. de Génie électrique (LAGE), Univ 
Ouargla, 30000 Ouargla, Algeria. 
2
Laboratory of Mathematics Informatics and Systems (LAMIS) University of Larbi Tebessi Tebessa, Algeria. 
E-mail:  korichi.maarouf@univ-ouargla.dz, samai.djamel@univ-ouargla.dz, benlamoudi.azeddine@univ-ouargla.dz, 
ameraoumia@univ-tebessa.dz, bensid.khaled@univ-ouargla.dz 
Keywords: biometrics, identification, hyperspectral palmprint, band selection, OCF, deep learning 
Received: February 11, 2023  
Over the past two decades, biometric technologies have exploded because anything that identifies a person 
provides a source of information. The palmprint modality is a biometric characteristic of great interest to 
researchers, and its traits can be found in various representations, including grayscale, colour, and 
multi/hyperspectral representations. The most difficult challenge in developing a hyperspectral palmprint-
based recognition system is determining how to use all the information available in these spectral bands. 
In this paper, we propose a hyperspectral palmprint identification system. In the first stage, an Optimal 
Clustering Framework (OCF) is proposed to extract the most representative bands. Then, two types of 
feature extraction methods (handcrafted and deep learning approaches) were used to determine the best 
method to describe palmprint features. After setting the number of selected bands to 4, we performed our 
set of experiments using the Hong Kong Polytechnic University (Poly U), which consists of 69 spectral 
bands. The results indicated that the proposed system offers the best performance, which qualifies it to be 
intended for usage in high-security situations. 
Povzetek:glavni cilj tega dela je oblikovati biometrični identifikacijski sistem z uporabo hiperspektralnega 
slikanja odtisa dlani (HSP) in nove sheme izbire pasu. 
 
1 Introduction 
Due to the progression of information security technology, 
the number of risks to the recognition system and its 
resources has increased, necessitating the need to 
strengthen security measures. The necessity for faster and 
more reliable user authentication approaches has 
increased security concerns in these days of rapid 
improvements in communication, networking, and 
mobility [1]. 
Researchers suggest the use of human biometric traits as a 
solution to guarantee the identity of system clients [2]. 
Biometrics is a term that refers to technologies that use a 
person's unique physiological or behavioural features for 
identification and/or authentication. Fingerprints, 
palmprints, facial patterns, iris patterns, voice, signatures, 
and gait are unique biometric attributes, often known as 
"biometric data". Nowadays, many biometric systems are 
being designed and used, and the utility of biometric 
systems is becoming more commonly understood and 
accepted because of the uniqueness of human biometrics, 
which is essential in preventing imposter threats. These 
authentication systems have defeated traditional security 
measures based on knowledge (such as a password) or  
 
 
 
 
possession (such as a key), which can be faked or broken 
[3], [4]. 
Palmprint recognition has recently been identified as a 
novel biometric recognition technology, and it has 
received great interest. Palmprint recognition is the term 
for the technology that uses images of the palm skin on a 
human’s hands to verify that the person is who they say 
they are [5]. The palmprint is of great interest to 
researchers because it is the most advanced biometric 
technique and has been used for over a century. However, 
this interest is due to its many advantages over other 
biometric modalities, such as usability, long-term 
stability, affordability, availability of materials, higher 
accuracy, and so on [6]. 
The research in palmprint imaging-based systems can be 
found in various representations. One can find a multi-
spectral (MSP) based palmprint representation that 
provides additional discriminating information. Because 
four spectral bands have peaks at various light 
wavelengths, such as vein networks where the near-
infrared band (NIR) penetrates the skin, visualisation of 
the vein pattern is allowed [7]. Therefore, hyperspectral-
based palmprint imaging is a new research area in which 
researchers in numerous domains have been interested 
because of its high ability to distinguish between 
customers (users). In comparison to multi-spectral 

64 Informatica 47 (2023) 63–80 M. Korichi et al. 
palmprint identification, the hyperspectral palmprint 
sensor collects spectral information varying from 420 nm 
to 1100 nm from 69 different bands, which could provide 
more discriminative information. 
Nevertheless, two major difficulties must be solved before 
hyperspectral biometrics can be used. Firstly, how many 
spectral bands are required to distinguish between distinct 
palms? On the one hand, typically, more feature bands 
provide more information, implying more accuracy. In 
contrast, increasing the number of feature bands increases 
feature extraction and matching costs. Moreover, 
providing more information may not always improve 
accuracy due to redundancies between distinct spectra. 
Secondly, how do we choose representative spectra for a 
given number of feature bands? In hyperspectral palmprint 
identification, optimal band selection and band 
combination are major research topics [8–9]. 
However, exploiting all the possible information from 
hyperspectral imaging is a difficult challenge. To do that, 
two main strategies can be found in the literature to utilise 
all the possible information. Clustering, band 
combination, and band selection are the most commonly 
used approaches in hyperspectral-based palmprint 
recognition systems. In this paper, to maintain all of the 
important information about the palmprint and reduce the 
number of bands as much as possible, a selection of bands 
should be made first before palmprint identification. Thus, 
a band selection scheme is developed and adapted to 
extract the most representative hyperspectral palmprint 
bands. A multimodal palmprint-based recognition system 
is proposed, with two feature extraction approaches: 
handcrafted-based and deep learning-based. After 
extracting the relevant information from the selected 
hyperspectral palmprint band images, a Deep Rule-Based 
classifier (DRB) is deployed to determine whether the 
person who tries to access the system is genuine or an 
impostor. 
To summarise our work, the following contributions can 
be cited: 
▪ An Optimal Clustering Framework (OCF) is 
proposed for automatically selecting the most 
discriminative hyperspectral palmprint bands. 
▪ Efficient unimodal and multimodal HSP 
palmprint identification systems have been made 
based on the chosen bands. 
▪ An extensive comparative analysis between the 
handcrafted features and the deep learning is 
performed to choose the best techniques that 
better describe the palmprint features. 
▪ The impact of incorporating a deep rule-based 
classifier (DRB) for classification tasks was 
investigated. 
To construct our paper, we followed the following steps: 
Firstly, in Section 2, we introduce the state of the art of 
hyperspectral palmprint-based recognition systems. Then, 
Section 3 discusses the proposed scheme for an HSP 
palmprint-based identification system. After that, the 
optimal clustering framework for hyperspectral band 
selection is outlined in Section 4. Section 5 briefly 
describes the feature extraction descriptors used 
(handcrafted-based and deep learning-based). The 
architecture of the Deep Rule-Based classifier (DRB) is 
described in Section 6. Section 7 then presents the 
experimental results on the PolyU hyperspectral palmprint 
database. Section 8 concludes with conclusions and future 
work. 
2 Hyperspectral palmprint 
recognition: A literature review 
Relevant research has been published for the identification 
of a person’s palmprint traits. So, many approaches for 
extracting features from palmprint scans have recently 
attracted a variety of research interests. The biggest 
challenge for constructing an HSP- or even MSP-based 
palmprint recognition system is how to exploit all the 
possible information from those images. Therefore, one of 
the crucial tasks in the HSP palmprint-based recognition 
system is the extraction of features. This step is all about 
getting the most important discriminative information 
from the region of interest (ROI) that has been identified 
and making the palmprint of different subjects more 
separable. In this context, we are going to talk in this 
section about the most pertinent studies that treat the 
subject of HSP-based palmprint recognition systems. 
Although the literature on hyperspectral palmprint 
identification is relatively limited, some works extract 2D 
features from each band image and fuse them for 
recognition. In contrast, clustering and band selection 
techniques are used in other works to select the most 
informative bands. 
Hereafter, we classify the feature extraction approaches 
into two primary categories: handcrafted-based techniques 
and deep learning-based techniques. 
2.1 Handcrafted-based approaches 
Handcrafted features-based palmprint recognition 
depends on texture characteristics, which are low-level 
features that characterise a specified area of an image and 
give accurate features for the highest palmprint 
recognition rate [10]. 
Shen et al. [11] developed a 3D Gabor wavelet-based 
approach to concurrently extract information in the spatial 
and spectral domains. However, a series of three-
dimensional Gabor wavelets with different frequencies 
and orientations were created and convolved with the cube 
to extract discriminative information. The authors of this 
study calculated the similarity between two hyperspectral 
cubes using the humming distance. 
As an extended version of their work, Shen in [12] and 
[13] suggested using clustering techniques and 3D Gabor 
wavelets to reduce the band’s number. To choose the most 
representative bands, they propose the use of a clustering-
based method known as Affinity Propagation (AP) in their 
work. 
Another research work is based on using a 2D Gabor filter 
response for hyperspectral palmprint recognition [14]. The 
features descriptor is applied to each hyperspectral band 

An Effective Hyperspectral Palmprint Identification System Based… Informatica 47 (2023) 63–80 65 
 
separately after a dimensionality reduction using (2D)
2
 
LDA. The similarity measure is performed through the use 
of SVM and KNN classifiers. 
Exploiting all the information offered by the hyperspectral 
images is a difficult challenge. Therefore, Korichi et al. 
[15] introduced a new feature descriptor called 3D local 
binary pattern (3DLBP). The proposed features descriptor 
is an extended version of the well-known descriptor LBP 
(Local Binary Pattern), and it can be applied to any image 
type (grayscale, colour, multi-spectral, and hyperspectral 
images). 
In their study [16], Guo et al. found that the spectral bands 
700 nm and 960 nm provided the strongest discriminatory 
information. The extracted features are based on using a 
feature descriptor called (2D)
2
 PCA, where the bands are 
selected according to their recognition accuracy. 
Also, using Principal Component Analysis (PCA) and the 
Hidden Markov Model, Meraoumia et al. [17] presented a 
biometric recognition system based on multispectral and 
hyperspectral palmprint images. Firstly, the feature 
vectors are extracted using PCA [18], and after that, the 
feature vectors are modelled using HMM. 
Sun [19] and Chlaoua [20] introduced the use of k-mean-
based clustering techniques, which are aimed at reducing 
the number of spectral bands. Their method eliminates 
low-quality bands by computing image entropy and EER 
with a 2D Gabor filter response. In addition, the optimal 
band combination is chosen using a clustering technique 
that is validated by the band fusion technique after this 
procedure. 
Guo et al. [21] use hyperspectral palmprint data to 
investigate feature band selections. Although there were 
some promising results in accuracy and anti-spoofing 
capacity, the feature band selection from many 
hyperspectral palmprints needed to be faster, making it 
unsuitable for real-time deployment. 
2.2 Deep learning-based approaches 
Deep learning is regarded as a breakthrough in computer 
vision and has been effectively applied in different 
domains. Neural networks and deep learning approaches 
are the most recent categories of methods. Because of their 
adaptability and representational capacity, these methods 
are gaining popularity in various fields, particularly 
palmprint-based biometric systems [18]. 
Zhao et al. [22] presented a joint deep convolution feature 
representation to recognise hyperspectral palmprints. All 
spectral bands can be used with a CNN stack to create a 
single convolutional feature. They tested their model's 
performance on a hyperspectral palmprint dataset with 53 
bands. They were able to attain an EER of 0.01 %. 
The same author, Zhao [23], proposes combining features 
from different hyperspectral palmprints. This strategy 
combines global Deep Convolutional Neural Network 
(DCNN) features, texture LBP features, and direction 
LDP features. The hyperspectral palmprint images are 
subjected to the LBP, LDP, and DCNN algorithms to 
extract their features, resulting in three feature matrices. 
The 2D PCA is then applied to each feature matrix to 
remove redundant information and reduce the dimension. 
An EER of 0.11% was reported for a database including 
53 distinct bands. 
Based on PCANet's deep learning features, Meraoumia et 
al. [24] propose a palmprint biometric identification 
system. After extracting the PCANet features, four 
classifiers (SVM, RBF, RFT, and KNN) are used to 
differentiate between the distinct palmprint feature 
vectors. In their research work, two multi-spectral 
palmprint databases are used. 
However, a recent research study introduced by Trabelsi 
and her team [10] aimed to construct a multispectral 
palmprint identification system. The suggested system 
relies on PalmNet and Log-Gabor capabilities, utilising 
feature selection and dimensionality reduction algorithms. 
The proposed method is evaluated using three different 
multispectral palmprint databases. 
Bensid et al. [25] suggested using a deep learning-based 
feature extraction technique called discrete cosine 
transform network (DCTNet) to construct a multi-spectral 
palmprint recognition system. The CASIA and PolyU 
palmprint databases are the most commonly used 
benchmarks for evaluating their proposed systems. 
Hence, after analyzing a variety of research articles that 
treat the use of deep learning-based approaches, one can 
find that the majority of those research articles use a 
database of 4 or 6 bands maximally (multi-spectral 
databases), where they try to select which band or which 
combination is the most appropriate to represent a 
palmprint identification system without the use of a band 
selection algorithm. On one hand, as we previously 
mentioned, hyperspectral imaging offers more 
information that can be useful to differentiate between 
users. On the other hand, the more the number of bands 
there are, the more redundant information we have, which 
influences the system accuracy. For this reason, we 
suggest using a new bands selection algorithm before 
constructing a deep learning-based hyperspectral 
palmprint system. This will help choose the most useful 
bands and eliminate redundant information. 
3 Proposed system design 
The objective of this study is to develop a hyperspectral 
palmprint identification system and an Optimal Clustering 
Framework (OCF) for the selection of hyperspectral 
bands. To do this, an OCF is first applied to the utilised 
hyperspectral palmprint database (the Hong Kong 
University hyperspectral palmprint database), where the 
objective is to extract the most representative bands. After 
choosing the bands, the proposed hyperspectral palmprint 
recognition system is made up of two separate subsystems 
that interface with each other at the matching score level. 
As shown in Figure 1, there are four steps in each 
subsystem: preprocessing, feature extraction, matching, 
and making a decision. To get into the system's database 
(enrollment), a person has to give a set of training 
palmprint modalities. Typically, two approaches for 
feature vector extraction are used, namely, a handcrafted 
approach where Local Phase Quantization (LPQ) and 
Binarized Statistical Image Feature (BSIF) are used. The 
second approach is based on the use of a deep learning 

66 Informatica 47 (2023) 63–80 M. Korichi et al. 
approach, specifically the use of well-known pre-trained 
networks like AlexNet, VGG16, and VGG19. For 
identification, the same feature descriptors are used to 
extract the feature vectors from the test palmprint. Then, 
the Deep Rule-Based classifier (DRB) is used to compare 
the test palmprint templates with the database of reference 
feature vectors. Finally, for the multimodal system, each 
subsystem generates its own matching score, which is then 
combined to obtain a final score (fusion at the score level). 
This last score is used to determine whether the user 
should be accepted or rejected. 
4 Hyperspectral palmprint band 
selection 
From such an implementation viewpoint, a well-chosen 
band set should maintain just the most informative bands 
while removing those that offer little discriminating 
information. Because the first step of our work is selecting 
the representative hyperspectral bands, we suggest the use 
of an Optimal Clustering Framework (OCF). In this 
section, we will briefly present the proposed method for 
selecting the appropriate hyperspectral palmprint bands. 
The next three steps in installing the proposed framework 
are as follows: 
▪ Firstly, an optimal clustering framework is 
developed to find the best HSI clustering 
structure. However, the hyperspectral palmprint 
database can be analysed using the proposed 
OCF, which can determine the optimal solution 
under an acceptable constraint. Furthermore, 
because the proposed OCF is a global 
framework, many types of objective functions 
can be optimised using the same approach as long 
as they accord with the framework's specific 
form. Also, the suggested OCF is a global 
framework. This implies that the same method 
can be used to optimise different types of 
objective functions, as long as they fit the 
specific form of the framework. 
▪ After the clustering structure is achieved, an RCS 
(ranking on clusters strategy) is proposed as a 
useful criterion to select the most representative 
bands. With the addition of an arbitrary ranking 
algorithm, the RCS may better use the 
advantages of both clustering and ranking-based 
techniques. This will produce band subsets with 
lower correlation and more discriminative 
information. Some ranking criteria, such as 
MVPCA (Maximum-Variance Principal 
Component Analysis), Enhanced Fast Density-
Peak-based Clustering (E-FDPC), and 
Information Entropy (IE), can also be used to 
finish the RCS task. 
▪ Finally, an automated approach for determining 
the required number of bands is proposed 
because the goal is to reduce the correlation 
between bands to determine how much unique 
information may be generated by a given number 
of bands. 
The overall procedure for designing a band selection 
algorithm is shown in Figure 2. 
5 Feature extraction 
One of the difficult tasks in constructing such a biometric-
based identification system is extracting the relevant 
information that can be used to differentiate between 
users. Typically, several algorithms are utilised in the 
feature extraction process; the approach differs based on 
the type of biometric identification being used. Our study 
uses two feature extraction approaches: a handcrafted-
based and a deep learning-based approach. In this section, 
we will briefly explain the feature descriptors used. 
5.1 Local phase quantization (LPQ) 
Ojansivu et al. [26] proposed local phase quantization as a 
texture descriptor. Also, it was difficult to blur and 
outperform for texture classification compared to other 
descriptors. For the first time, the LPQ descriptor has been 
designated for use in the classification of texture blur. LPQ 
is designed to preserve an image's local invariant 
 
Figure 1: Architecture of the proposed hyperspectral palmprint recognition system. 
 
 
 

An Effective Hyperspectral Palmprint Identification System Based… Informatica 47 (2023) 63–80 67 
 
information against artefacts caused by various types of 
blurring. 
5.2 Binarized statistical image features 
(BSIF) 
Kannala et al. [27] developed a texture descriptor called 
Binarized Statistical Image Features. The BSIF is a binary 
code string representing a given image's pixels. A pixel's 
code value is regarded as a local descriptor of the image in 
its surroundings. BSIF has been developing local image 
descriptors that efficiently encapsulate texture information 
and can be used to represent image regions using 
histograms. The method generates a binary code for each 
pixel by linearly projecting local image patches onto a 
subspace whose basis vectors are binarized by 
thresholding after being learned from natural images using 
Independent Component Analysis (ICA). The number of 
basis vectors determines the length of the binary code 
string. Histograms of pixel binary codes can be used to 
represent image regions. 
5.3 AlexNet 
AlexNet has eight weighted layers, five convolutional 
layers, and three fully linked layers. Except for the last 
layer, which outputs a softmax with a distribution over the 
1000 class labels, ReLu activation is performed after each 
layer. In the first two completely connected layers, 
dropout is used. 
Alex Krizhevsky and his team won the 2012 ImageNet 
Large Scale Visual Recognition Challenge with AlexNet, 
which used an 8-layer CNN [28]. This network showed for 
the first time that learning-based features can do better 
than manually designed ones, shattering the old paradigm 
in computer vision. 
5.4 VGG models 
VGG (Visual Geometry Group) is a convolutional 
neural network designed by K. Simonyan and A. 
Zisserman [29] of the University of Oxford that achieved 
recognition in 2014 after winning the ILSVRC (ImageNet 
Large Scale Visual Recognition Challenge). On Imagenet, 
the model achieved 92.7% accuracy, one of the greatest 
results ever. 
It was a step forward from earlier models in that it 
proposed convolution kernels with lower dimensions (33) 
in the convolution layers than had been done previously. 
Using cutting-edge graphics cards, the model was trained 
over several weeks. 
VGG16 [29] is a convolutional neural network trained on 
images from the ImageNet database. The network has an 
image input size of (224 × 224) and can categorise 
images into 1000 item categories with 16 layers. It has 
about 138 million parameters, making evaluation 
challenging and necessitating a large amount of memory. 
As a result, the network has learned a variety of rich 
feature representations for various images. 
The VGG19 model is a variant of the VGG model with 19 
layers (16 convolution layers, 5 MaxPool layers, 1 
SoftMax layer, and three fully connected layers). 
6 Matching and fusion 
Once each image's distinctive and relevant characteristics 
have been determined, the images are divided into 
numerous groups, and a decision is made to realise 
whether the user is a client or an impostor. In our system, 
the Deep Rule Based classifier (DRB) is devoted to 
performing the matching task. 
 
 
 
Figure 2: Diagram showing the required steps to conduct the OCF-based band selection process.    
 
Figure 2: The DRB classifier's general architecture 

68 Informatica 47 (2023) 63–80 M. Korichi et al. 
6.1 Deep Rule-Based classifier 
The Deep Rule-Based classifier (DRB) [30] is a Semi-
Supervised Multi-Layer Neuro-fuzzy Modeling 
(SSMNM) technique based on image prototypes that 
produce visible and human-comprehensible fuzzy rules 
(IF... THEN. Once new classes are added to the system, an 
incremental learning algorithm can be utilised in real-time 
applications to monitor new concepts, deal with 
uncertainty, self-evolve, and adapt their network and 
meta-parameters. The DRB's layers are as follows: 1) the 
scaling layer, 2) the normalisation layer, 3) the feature 
descriptor layer, 4) the FRB layer, and 5) the decision-
maker layer. 
A. Layer of Scaling (Optional): The Scaling Layer is in 
charge of shrinking the original image to reduce 
computational complexity, improve generalisation, 
etc. All images were resized to (128 x 128) as in the 
study, according to the requirements of the feature 
descriptor layer. 
B. Normalization Layer (A standard pre-processing 
step): The Normalisation Layer is in charge of 
translating image pixel values into the feature 
descriptor's specified range. This study's pixels are 
normalised between 0 and 1 because of the feature 
descriptor layer requirements. 
C. The Feature Descriptor Layer: The Feature 
Descriptor Layer extracts the global feature vectors 
from the images. The retrieved features are applied to 
the classification model's training and validation. 
D. Base layer using Fuzzy-Rule (Model engine): The 
DRB classifier's basic engine is the Fuzzy-Rule base 
(FRB) layer, composed of a massively parallel 
ensemble of type 0-order fuzzy rules. This layer is 
responsible for producing a highly accurate classifier 
while simultaneously providing interpretability and 
transparent models for human comprehension, in 
contrast to ordinary deep learning techniques, known 
as black boxes, since they do not offer insights into the 
network's structure. 
E. The Decision-Maker Layer: This layer is responsible 
for making the final decision by labelling the 
validation images depending on the degree of 
prototype similarity determined by the FRB layer. 
The essential layers of the DRB classifier architecture are 
shown in Figure 3 [30]. 
6.1.1 The DRB classifier algorithm 
First, we describe the important notations utilised in the 
DRB classifier algorithm in Table 1. 
 
 
Table 1: DRB classifier algorithm key notation 
description 
Notations Descriptions 
C The dataset's classes number 
D The feature vector's 
dimensionality 
K The observed training images 
number/instance of the current time  
I A single input image 
X The I corresponding feature vector  
NC The c
th
 class identified prototypes 
number  
𝝁 𝒄 The c
th
 class training images 
feature vector’s global average.  
Ic,k The  c
th
 class k
th
 training image  
𝝌 𝒄 ,𝒌 The I c,k corresponding feature 
vector  
Pc,i The c
th
 class i
th
 prototype 
pc,i The average of training images 
feature vectors associated with P c,i 
Sc,i The training images number 
associated with P c,i 
rc,i The data cloud influences area 
radius associated with  P c,i 
𝝀 𝒄 The confidence score is given by 
the local decision-maker of the c
th
 
fuzzy rule.  
The four main phases of the DRB classification algorithm 
are as follows [30]: 
1. Step 01: Pre-processing block- Most of the 
time, the DRB classifier uses simple and 
common pre-processing techniques, such as 
image normalisation, rotation, scaling, and image 
segmentation. 
2. Step 02: Feature extraction layer- Various 
types of feature descriptors can be used on this 
layer. 
3. Step 03: Massive parallel FRB layer- This step 
consists of four parts, which are as follows: 
initialization, preparation, system updating, and 
fuzzy rule creation, as stated below: 
Massive parallel FRB layer stages 
Stage 1: Initialization 
The identical DRB algorithm parameters are 
initialized by: 
𝑘 ← 1; 𝜇 𝑐 ← 𝜒 ̅ 𝑐 1
; Ν
𝑐 ← 1; Ρ
𝑐 ,Ν
𝑐 ← Ι
𝑐 ,1
; 𝑝 𝑐 ,Ν
𝑐 ← 𝜒 ̅ 𝑐 1
; 
𝑆 𝑐 ,Ν
𝑐 ← 1; 𝑟 𝑐 ,Ν
𝑐 ← 𝑟 0
  
Stage 2: Preparation 
The 𝜒 ̅ 𝑐𝑘
 and 𝐷 (Ρ
𝑐 ,𝑖 ) for each image Ι
𝑐 ,𝑘 are calculated 
by: 
𝜒 ̅ 𝑐𝑘
=
𝜒 𝑐𝑘
‖𝜒 𝑐𝑘
‖
; 𝜇 𝑐 =
𝑘 −1
𝑘 𝜇 𝑐 + 𝜒 ̅ 𝑐𝑘
   
𝐷 (Ρ
𝑐 ,𝑖 ) =
1
1+‖𝐶 −𝜇 𝑐 ‖
2
/𝜎 2
  
Stage 3: Updating system 

An Effective Hyperspectral Palmprint Identification System Based… Informatica 47 (2023) 63–80 69 
 
The DRB algorithm verifies two conditions 
(Conditions 1 and 2) before updating the parameters of 
all images in stage 1. 
 
If condition 1 is true then add a new data cloud else 
find the nearest prototype 𝑃 𝑐 ,𝑛 corresponding to Ι
𝑐 ,𝑘 and 
go to the condition 2 
 
If condition 2 is true, Update 𝑃 𝑐 ,𝑛 , 𝑆 𝑐 ,𝑛 and 𝑟 𝑐 ,𝑛 2
 else 
add a new data cloud 
 
Stage 4: Fuzzy rules generation-Generate rule type 
Rulec 
𝒊𝒇 (𝑰 ~𝑃 𝑐 ,1
) 𝑂𝑅 ⋯ 𝑂𝑅 (𝑰 ~𝑃 𝑐 ,𝑁 𝑐 ) 𝑻𝒉𝒆𝒏 (𝑐𝑙𝑎𝑠𝑠𝑒 𝑐 )   
 
Condition 1  
 
𝒊𝒇 (𝐷 (Ι
𝑐 ,𝑘 ) >
𝑖 =1,2,3,⋯,𝑁 𝑐 𝑚𝑎𝑥
(𝐷 (Ρ
𝑐 ,𝑖 ))) 𝑂𝑅  
 (𝐷 (Ι
𝑐 ,𝑘 ) <
𝑖 =1,2,3,⋯,𝑁 𝑐 𝑚𝑖𝑛 ( 𝐷 (Ρ
𝑐 ,𝑖 )))   𝑻𝒉𝒆𝒏 Ι
𝑐 ,𝑘 is new 
prototype 
𝑁 𝑐 ← 𝑁 𝑐 + 1  
Ρ
𝑐 ,Ν
𝑐 ← Ι
𝑐 ,𝑘 ; 𝑝 𝑐 ,Ν
𝑐 ← 𝜒 ̅ 𝑐 ,𝑘 ; 𝑆 𝑐 ,Ν
𝑐 ← 1; 𝑟 𝑐 ,Ν
𝑐 ← 𝑟 0
  
Else  
Find 𝑃 𝑐 ,𝑛 by 𝑃 𝑐 ,𝑛 = 𝑎𝑟𝑔𝑚𝑖𝑛 (‖𝜒 ̅ 𝑐 ,𝑘 − 𝑃 𝑐 ,𝑗 ‖  
where 
𝑗 = 1,2, ⋯ , Ν
𝑐 
Condition 2 
𝑖𝑓 (‖𝜒 ̅ 𝑐 ,𝑘 − 𝑃 𝑐 ,𝑛 ‖ < 𝑟 𝑐 ,Ν
𝑐 )  𝑻𝒉𝒆𝒏 Ι
𝑐 ,𝑘 is assigned to 
 𝑃 𝑐 ,𝑛 
Else  
𝑁 𝑐 ← 𝑁 𝑐 + 1; Ρ
𝑐 ,Ν
𝑐 ← Ι
𝑐 ,𝑘 ; 𝑝 𝑐 ,Ν
𝑐 ← 𝜒 ̅ 𝑐 ,𝑘 ;  
𝑆𝑐 , Ν
𝑐 ← 1;  
 𝑟 𝑐 ,Ν
𝑐 ← 𝑟 0
    
 
Endif  
Endif  
 
4. Step 4. Decision maker layer: For each image 𝐼 
in the test data, find the confidence score 𝜆 𝑐 (𝐼 ) 
based on the image's feature vector by using the 
following formula:  
𝜆 𝑐 (𝐼 ) = 𝑎𝑟𝑔𝑚𝑎𝑥 (𝑒𝑥𝑝 (−‖𝜒 − 𝑃 𝑐 ,𝑗 ‖
2
), 𝑗 =
1,2, ⋯ , Ν
𝑐 .   
- the labels are identified by: 
𝑙𝑎𝑏𝑒𝑙 (𝐼 ) = 𝑎𝑟𝑔𝑚𝑎𝑥 (𝜆 𝑐 (𝐼 ) , 𝑐 = 1,2, ⋯ , 𝑐 . 
6.2 Fusion process 
It has been demonstrated in several studies that combining 
the results from different unimodal biometric 
identification systems generally leads to better system 
efficiency. Fusion at the matching-score level involves 
integrating the scores generated by each biometric system 
that describe the similarity between the biometrics 
acquired and their models [31]. We experimented with 
four different schemes during the system's design: 
▪ SUM-score (SUM):  
𝑆 = ∑ 𝑆 𝑖 𝑁 𝑖 =1
                     (1) 
 
▪ SUM-Weighting-Score (WSum): 
𝑆 = ∑ w
i
S
i
N
i=1
                   (2) 
 
▪ Production-score (Prod): 
𝑆 = ∏ 𝑆 𝑖 𝑖 =1,2,...,𝑁                          (3) 
 
▪ Production-Weighting- score (WProd): 
𝑺 = ∏ 𝒘 𝒊 𝑺 𝒊 𝒊 =𝟏 ,𝟐 ,...,𝑵                        (4) 
 
Where 𝑤 𝑖 denotes the weight of the 𝑖 𝑡 ℎ
 biometric 
attribute of the 𝑘 𝑡 ℎ
 user's matching score, which is 
determined as follows: 
𝑤 𝑖 =
1
𝐸𝐸𝑅
𝑖 ∑
1
𝐸𝐸𝑅
𝑖 𝑁 𝑖 =1
                                (5) 
7 Experiment results and discussion 
As mentioned previously, this work's main objective is to 
design a biometric identification system using 
hyperspectral palmprint (HSP) imaging. A band selection 
scheme is first implemented before constructing our 
system and reducing the spectral band's number. When the 
representative bands have been established, two feature 
extraction methods are used to perform the feature 
extraction task: handcrafted and deep learning-based 
approaches. To test and evaluate the proposed system, 
experiments were conducted using the hyperspectral 
palmprint database collected by the HK PolyU Institute. 
7.1 Hyperspectral palmprint database 
description 
In the experiment test phase, the publicly available HK-
PolyU Hyperspectral palmprint database is used [32]. 
The Hong Kong Polytechnic University's Biometric 
Research Centre (UGC/CRC) has designed a 
hyperspectral palmprint capture device to collect images 
from 420 nm to 1100 nm. 
The Hong Kong PolyU hyperspectral palmprint database 
is a large database comprised of 190 volunteer images. 
However, the volunteers in this database range from 20 to 
60 years old. The samples were collected in two separate 
sessions. The patient was instructed to create seven 
palmprint images, with a size of 128 x 128 pixels, for each 
wavelength throughout each session. A set of 69 spectral 
bands with a step length of 10 nm might be used to image 
a palmprint over a range of 420 nm–1100 nm. With the 
first and second sessions separated by a month, there is 
typically a month between them. The database contains 
5,240 images from 380 different palms at one wavelength. 
7.2 OCF based hyperspectral palmprint 
bands selection 
The problem of huge processing complexity occurred 
with HSI's extensive dimensional data, and it remained a 

70 Informatica 47 (2023) 63–80 M. Korichi et al. 
significant challenge because of the strong correlations 
and dependencies among said data. The first step in 
designing our proposed hyperspectral palmprint system is 
to reduce the number of spectral bands and only use the 
most representative ones. This section investigates the 
implementation of the OCF framework to select HIS's 
most representative bands. However, three RCS were 
tested and evaluated to achieve the optimum results, 
including MVPCA, E. FDPC, and IE. Table 2 shows the 
results that came from using these methods. 
It should be noted that the PolyU hyperspectral palmprint 
database comprises two samples taken from the right and 
left hands. To accomplish the band selection process, the 
RCS procedures are applied independently to the two 
samples. While implementing the OCF framework, we set 
the number of the output selected bands to 4. It is possible 
to justify this by comparing it to the MSP palmprint 
database, which has four different modalities (Red, Green, 
Blue, and NIR). 
As previously stated, one of our objectives is to select the 
most representative HSI spectral bands. As a result, 
various experiments were carried out and evaluated using 
three RCS methods. When looking at and analysing Table 
1, If the left-hand spectral bands are employed, the 
MVPCA RCS techniques generate bands of wavelength 
(rank) as follows: 610 nm (20), 710 nm (30), 810 nm (40), 
and 960 nm (55). When the right-hand spectral bands are 
employed, the selectable bands are 610 nm (20), 710 nm 
(30), 870 nm (46) and 940 nm (940 nm) (53). However, 
by employing the left-hand spectral bands and the E FDPC 
techniques as an RCS methodology, the resulting OCF-
based systems are 470 nm (06), 690 nm (28), 960 nm (55), 
and 990 nm (58). The system-selected bands for the right-
hand spectral bands are 480 nm (07), 590 nm (18), 850 nm 
(44) and 990 nm (58). Finally, if the IE-based RCS 
technique is employed in conjunction with the left-hand 
spectral bands, the selected bands are 610 nm (20), 690 
nm (28), 850 nm (44) and 940 nm (53). The IE selectable 
bands for the right-hand spectral bands are 610 nm (20), 
760 nm (35), 850 nm (44) and 940 nm (53). 
We noticed through manual visual analysis that the first 
two bands chosen by E_FDPC-based RCS techniques (the 
06th spectral band (470 nm) for the left hand and the 07th 
(480 nm) with the 18th (590 nm) spectral bands for the 
right hand) carry little to no information, making them 
useless for our proposed system. 
As a result, we concluded that it would produce 
undesirable effects on the proposed system's performance. 
This leaves us with the two remaining techniques that 
produce relatively similar results, compelling us to 
experiment with the selected bands. In conclusion, the 
entropy bands produced a more desired effect than the 
MVPCA bands. 
An example of the HSI-selected bands using the IE-based 
RCS techniques is shown in Figure 4. Hence, the selected 
bands for both the right and left hand are then used to 
construct a unimodal/multimodal biometric identification 
system where two feature extraction approaches are used 
Table 2: Hyperspectral palmprint band selection using OCF framework 
Modality RCS method 1
st
 band 2
nd
 3
rd
 band 4
th
 band 
 
LEFT Hand 
MVPCA 610 nm (20) 710 nm (30) 810 nm (40) 960 nm (55) 
E_FDPC 470 mm (06) 690 nm (28) 960 nm (55) 990 nm (58) 
IE 610 nm (20) 690 nm (28) 850 nm (44) 940 nm (53) 
 
RIGHT Hand 
MVPCA 610 nm (20) 710 nm (30) 870 nm (46) 940 nm (53) 
E_FDPC 480 nm (07) 590 nm (18) 850 nm (44) 990 nm (58) 
IE 610 nm (20) 760 nm (35) 850 nm (44) 940 nm (53) 
 
 
 
Figure 4: Information Entropy (IE) based RCS techniques most representative 

An Effective Hyperspectral Palmprint Identification System Based… Informatica 47 (2023) 63–80 71 
 
along with the DRB classifier. The next section will 
describe and assess the proposed identification system's 
performance. 
7.3 Unimodal systems test results 
After extracting the most representative HSI palmprint 
bands based on the OCF framework, our experiment's next 
step is to deploy them to design a reliable biometric 
identification system. It was necessary to conduct three 
separate experiments in order to evaluate the performance 
of our proposed hyperspectral palmprint identification 
system. In the first experiment, we used only one spectral 
band. In contrast, we will integrate multiple spectral bands 
in the second and third experiments to construct a 
multimodal biometric system. Note that in all 
experiments, two approaches for feature extraction are 
used: handcrafted approaches (LPQ and BSIF) and deep 
learning techniques with three pre-trained networks 
named VGG16, VGG19, and the AlexNet network. 
As we reported earlier, the palmprint image database 
contains 190 persons. Each person has 12 images of the 
left hand and 12 of the right hand, which were taken in 4 
bands. We chose six images from 12 for training 
[𝟏 𝟑 𝟓 𝟕 𝟗 𝟏𝟏 ] and the other six [𝟐 𝟒 𝟔 𝟖 𝟏𝟎 𝟏𝟐 ] for the 
test (this applies to both the left-hand and right-hand 
palmprint images). 
Table 3 shows the performance of the proposed unimodal 
system by using palmprint images captured at the selected 
band wavelength. 
For the best biometric identification system performance, 
which means minimising the Error Equal Rate (EER) 
for the open-set identification mode and maximising the 
Recognition of One Rate (ROR) for the closed-set 
identification mode, our experiments aim to find the best 
selected spectral bands and the best feature descriptor 
(handcrafted or deep learning-based) to represent the 
palmprint features. 
According to the results shown in Table 3 and Figures 5, 
6, 7, and 8, one can observe that: 
The first remark is that the band 610 nm-based system 
performance achieves the lowest performance for both 
identification modes and feature extraction approaches. 
Also, one can observe that the performance of the deep 
learning feature-based feature extraction system 
 
 
Figure 5: Proposed open-set identification system based on Handcrafted feature extraction approach. 
 
 
Figure 6: Proposed closed-set identification system based on Handcrafted feature extraction approach. 
 
 
 
 
 
 

72 Informatica 47 (2023) 63–80 M. Korichi et al. 
outperforms the system performance when using 
handcrafted feature ex traction techniques. Hereafter, we 
will discuss and analyse the proposed unimodal system's 
performance. In the case of the open-set identification 
mode, handcrafted feature descriptors are used. Our 
results show that the system performs best when utilising 
the left-hand-based spectral bands, except when the 850 
nm right-hand-based spectral band is employed, where the 
system achieves its optimum performance, which results 
in an EER of 0.0318% at the threshold T0 = 0.8173 when 
the LPQ is used as a feature descriptor. 
Furthermore, if the LPQ is deployed, the EER ranges 
between [0.3508%–5.2631%] for the left-hand spectral 
bands and [0.3911%–6.4035% for the right-hand spectral 
bands. Whenever the proposed system uses the BSIF as a 
feature descriptor, the results are noticeably better than 
when the LPQ features are used. The EER of left-hand 
selectable bands based on BSIF features ranged from 
[0.1754%–4.2982%], whereas it ranged from [0.2154%–
5,4385%] for the right-hand selected bands. 
The second part of this section's experiments will evaluate 
the proposed system's performance when some well-
known deep pre-trained networks are used to extract deep 
features from the selected bands. Once the AlexNet model 
is used as a feature descriptor, it is possible to see that the 
proposed system achieves its best performance by 
providing the lowest value of EER (0.0129%) at a 
threshold value of T0 = 0.8701. This is the case when 
using the 850 nm right-hand spectral bands. However, 
through this series of experiments, we are convinced that 
the right-hand band of the 850 nm spectrum contains the 
most discriminative information. Now, turning our 
attention to the system's performance when deep learning 
feature descriptors were applied, what stands out in Table 
3 is that the proposed deep feature descriptors gave perfect 
results in certain bands for both the left and right hand. As 
can be seen, when the VGG16 is used with the right-hand-
            
 
Figure 7: Proposed open-set identification system based on deep learning deep features approach 
 
  
Figure 8: Proposed closed-set identification system based on deep learning deep features approach. 

An Effective Hyperspectral Palmprint Identification System Based… Informatica 47 (2023) 63–80 73 
 
based bands, the system produces an EER between 
[0.3508%-5.5263%], whereas when the left-hand-based 
bands are utilized, the system produces an EER between 
[0.0877%-4.1228%]. For the VGG19-based proposed 
system performance, one can notice that the system 
achieves approximately similar results to those obtained 
when using VGG16. Finally, Figures 5 and 7 are plotted 
to show how the open-set unimodal system performs when 
using the two feature approaches. 
In the case of closed-set identification, we notice from 
Table 3 that the previous observations made in the open-
set mode still apply to the closed-set mode. The proposed 
system achieves its best performance when the 850 nm 
spectrum right-hand band is deployed along with 
AlexNet-based deep learning techniques, where the 
system gives a ROR of 99.9123% with an RPR equal 
to 4. In addition, if the system works under a handcrafted-
based features descriptor, the best performance is obtained 
using the same spectrum and an LPQ descriptor (ROR = 
99.5614%, RPR = 9). However, it can be shown that both 
hands' two wavelengths, 850 nm and 940 nm, performed 
similarly. Compared to the other networks, the system's 
efficiency is clearly superior when the AlexNet network is 
utilized. To give a concise summary of the performance of 
the proposed closed-set identification mode for both 
feature extraction approaches, Figures 6 and 8 have been 
plotted. 
7.4 Multimodal systems test results 
This stage of experiments aims to improve the results 
given by the unimodal biometric identification system by 
using information from various bands and instances (left 
and right palmprints). Two fusion scenarios were used to 
accomplish this. The first is based on the fusion of four 
bands from the same instance (left or right hand). The 
second fusion scheme employs all bands from the two 
instances. The fusion process is carried out at the score 
level in our study work. When using matching score level 
fusion, the two sub-systems' matching scores are blended 
to produce a single score. The ultimate decision is then 
based on the outcome. To accomplish this, we experiment 
with rules-based fusion, which employs four distinct rules 
(described in Section 6.2). 
Therefore, several tests were carried out to choose the best 
fusion scenario and the fusion rule. The results are 
displayed in Tables 4 and 5. Once again, the EER is used 
to test the performance of the multimodal system in open-
set identification mode. From Table 4, where the first 
fusion scenario is applied, one can see that the basis of 
fusion for both feature extraction approaches enhance the 
efficiency of the unimodal system. This can be justified by 
the results obtained in Table 4 compared to those in Table 
3. As an example, one can observe that the EER is 
reduced to 0.000% at a threshold of 0.8440 when the 
right-hand spectral bands are fused using the PROD 
fusion rule and the AlexNet as a feature descriptor. This 
table also makes it clear how much the performance of the 
handcrafted feature-based approach has improved. For 
instance, in the right-hand fused bands, the maximum 
value of EER when utilising the LPQ descriptor was 
6.4035%, but after the fusion, it was only 0.5263%. This 
observation was also confirmed when the pre-trained deep 
learning networks were used, which served to justify the 
usefulness of incorporating the fusion process. By fusing 
all selected bands from the left and right hand (second 
fusion scenario), the system achieves its optimum 
performances under both feature approaches, where the 
EER is reduced to 0.000% for all feature descriptors (see 
Table 5) under the use of SUM and WSUM-based fusion 
rules (in the case of AlexNet-based system performances, 
the system achieves the same performances for all fusion 
rules). 
To validate our concepts, we have to run other 
experiments for closed-set identification. For the first 
scenario, we can also see from Table 4 that the AlexNet-
based system performs well where the system achieves a 
ROR of 100% with an RPR equal to one under the fuse 
of right-hand bands with the PROD fusion rules. Contrary 
to the second fusion scenario, the system, like the open-set 
experiment tests, achieves optimal performance under all 
feature descriptors when using SUM and WSUM fusion 
rules. The system produces a ROR of 100% and an RPR 
of 1. 
7.5 Comparison with state of the art 
To show the proposed protocol's efficiency, a comparison 
study with the most relevant and related research works is 
performed. 
It is necessary to mention that the comparison is 
performed to the research works where the protocols are 
evaluated using the PolyU hyperspectral palmprint 
database. 
Although, as mentioned in Section 2, the literature on 
hyperspectral palmprint identification is very limited, 
some works concentrate on extracting 2D information 
from each band image and fusing them for recognition. In 
contrast, other works use clustering and band selection 
approaches to select the most informative bands.  
However, the main objective in designing such a 
hyperspectral palmprint system is to obtain the optimum 
system performance using minimum features. For 
example,[21] achieves an EER of 0.0780% by increasing 
the number of clusters to 4. Also, [19] research is based on 
the clustering technique (3 clusters), where the obtained 
EER was 0.17325%. In addition, [12] designed a 
hyperspectral palmprint system and a clustering technique 
(10 clusters), and he obtained an EER=3.26%. Other 
works also extract information from the hyperspectral 
cube. For example, [11] obtained an EER= 4.00% using 
a hyperspectral palmprint of 52 bands, whereas [23] and 
[22] used a cube of 53 bands and obtained (EER=0.11%, 
ROR=99.76%) and (EER=0.015%, ROR=99.625%) 
respectively.  
By manually evaluating all 69 hyperspectral bands, [14] 
made a simple hyperspectral palmprint system with a 
ROR of 99.92%. This method is not preferred because it 
takes more processing time. 
The results in Table 6 clearly show that our proposed 
hyperspectral palmprint system exceed most of the cited 
relevant work in term of EER and ROR by achieving a 

74 Informatica 47 (2023) 63–80 M. Korichi et al. 
(EER = 0.0123%, ROR=99.9123%) in the case of 
unimodal system conception and a (EER=0.000%, 
ROR=100%) in the case of multimodal system. 
8 Conclusion 
Through technological improvements, we are becoming 
highly dependent on data, which is increasing in size and 
becoming vital and influential in our lives. Recently, 
palmprint-based biometric recognition systems have 
emerged as a promising field with a high ability to 
distinguish between people. Among the palmprint 
recognition systems, one can find systems that use 
hyperspectral palmprint images as a biometric modality. 
Due to the larger system database size and longer 
processing times, the most challenging part of designing 
such a hyperspectral palmprint system is utilizing all the 
information these images may offer. As a solution, we 
proposed optimal framework clustering (OCF) that 
automatically extracts the most representative 
hyperspectral bands to reduce the system database size 
and eliminate redundant information. After that, a 
biometric identification system based on the selected 
bands is constructed, and two feature descriptors are 
applied. The handcrafted features are extracted using LPQ 
and BSIF descriptors, and the AlexNet, VGG16, and 
VGG19 pre-trained networks are used to extract the deep-
learning features. The DRB classifier is devoted to 
performing the classification tasks. The proposed system 
is evaluated and tested using the PolyU hyperspectral 
palmprint database. The obtained unimodal experimental 
results showed that the bands of 850 nm and 940 nm can 
be considered the most representative’ bands according to 
the proposed protocol. Also, the experiments showed that 
the deep learning-based approach outperformed the 
handcrafted methods. However, the multimodal biometric 
system was introduced in order to overcome the 
limitations of the unimodal systems. As a result, according 
to the proposed fusion scenario, one can observe that the 
EER is reduced to 0.000% in the open-set identification 
mode and obtains a ROR equal to 100% in the closed-set 
mode. 
For future work directions, we will extend our proposed 
framework for band selection to other hyperspectral traits, 
such as face or iris hyperspectral images. We will also 
look into combining deep learning or handcrafted features 
from different bands to make a better biometric system. 
9 Acknowledgment 
The authors gratefully acknowledge the Directorate 
General for Scientific Research and Technological 
Development (DGRSDT) of Algeria for the financial 
support for this work. The authors also acknowledge 
resources and support from the electrical engineering 
laboratory (laboratoire de génie électrique (LAGE)) of 
KASDI Merbah University, Ouargla, Algeria. 
 
 
References 
[1] Altarawneh, Mokhled Sueliman. Cancelable 
fingerprint features using chaff points 
encapsulation. Informatica, 2018, vol. 42, no 3..  
https://doi.org/10.31449/inf.v42i3.1855   
[2] Sharif, M., Raza, M., Shah, J.H., Yasmin, M., 
Fernandes, S.L. (2019). An Overview of 
Biometrics Methods. In: Singh, A., Mohan, A. 
(eds) Handbook of Multimedia Information 
Security: Techniques and Applications. Springer, 
Cham. https://doi.org/10.1007/978-3-030-15887-
3_2 
[3] A. Benlamoudi, S. E. Bekhouche, M. Korichi, K. 
Bensid, and A. H. and A. T.-A. Abdeldjalil 
Ouahabi, “Face Presentation Attack Detection 
Using Deep Background Subtraction,” sensors, 
vol. 22, no. 10, pp. 1–17, 2022. 
https://doi.org/10.3390/s22103760 
 [4] S. A. Abdulrahman and B. Alhayani, “A 
comprehensive survey on the biometric systems 
based on physiological and behavioural 
characteristics,” Materials Today: Proceedings, 
2021. 
https://doi.org/10.1016/j.matpr.2021.07.005 
[5] Jia, W., Xia, W., Zhao, Y. et al. 2D and 3D 
Palmprint and Palm Vein Recognition Based on 
Neural Architecture Search. Int. J. Autom. 
Comput. 18, 377–409 (2021). 
https://doi.org/10.1007/s11633-021-1292-1 
[6] Raouia Mokni, Hassen Drira, Monji Kherallah. 
Deep-Analysis of Palmprint Representation based 
on Correlation Concept for Human Biometrics 
identification. International Journal of Digital 
Crime and Forensics, vol. 12, no. 2, pp. 40–58, 
2020.      
https://doi.org/10.4018/ijdcf.2020040103  
[7] Y. Aberni, L. Boubchir and B. Daachi, 
"Multispectral palmprint recognition: A state-of-
the-art review," 2017 40th International 
Conference on Telecommunications and Signal 
Processing (TSP), 2017, pp. 793-797,                    
https://doi.org/10.1109/tsp.2017.8076097 . 
[8] THAMRI, Essia, ALOUI, Kamel, et NACEUR, 
Mohamed Saber. Selection of hyperspectral bands 
by adopting a dimension reduction strategy for 
recognition of multispectral palmprint. In : 2017 
International Conference on Advanced Systems 
and Electric Technologies (IC_ASET). IEEE, 
2017. p. 91-96. 
https://doi.org/10.1109/aset.2017.7983672  
[9] Q. Wang, F. Zhang and X. Li, "Optimal Clustering 
Framework for Hyperspectral Band Selection," in 
IEEE Transactions on Geoscience and Remote 
Sensing, vol. 56, no. 10, pp. 5910-5922, Oct. 
2018,  
https://doi.org/10.1109/tgrs.2018.2828161 . 
[10] Trabelsi, S., Samai, D., Dornaika, F. et al. 
Efficient palmprint biometric identification 
systems using deep learning and feature selection 
methods. Neural Comput & Applic 34, 12119–

An Effective Hyperspectral Palmprint Identification System Based… Informatica 47 (2023) 63–80 75 
 
12141,(2022).                
https://doi.org/10.1007/s00521-022-07098-4 
[11] L. Shen, W. Wu, S. Jia and Z. Guo, "Coding 3D 
Gabor Features for Hyperspectral Palmprint 
Recognition," 2014 International Conference on 
Medical Biometrics, 2014, pp. 169-173, 
https://doi.org/10.1109/icmb.2014.36 . 
[12] L. Shen, Z. Dai, S. Jia, M. Yang, Z. Lai and S. Yu, 
"Band selection for Gabor feature based 
hyperspectral palmprint recognition," 2015 
International Conference on Biometrics (ICB), 
2015, pp. 416-421, 
https://doi.org/10.1109/icb.2015.7139104  
[13] Jie Zhou· Yunhong Wang· Zhenan Sun · Yong Xu 
Linlin Shen · Jianjiang Feng Shiguang Shan · Yu 
Qiao Zhenhua Guo · Shiqi Yu, Biometric 
recognition, Proceedings:12th Chinese 
Conference, CCBR 2017, Shenzhen, China, 
October 28-29, vol. 449, no. 7158,2017. 
https://doi.org/10.1007/978-3-319-69923-3  
[14] Khandizod, A.G., Deshmukh, R.R. (2019). 
Optimal Band Selection for Improvement of 
Hyperspectral Palmprint Recognition System by 
Using SVM and KNN Classifier. In: Santosh, K., 
Hegadi, R. (eds) Recent Trends in Image 
Processing and Pattern Recognition. RTIP2R 
2018. Communications in Computer and 
Information Science, vol 1036. Springer, 
Singapore. https://doi.org/10.1007/978-981-13-
9184-2_38  
[15] M. Korichi and A. Meraoumia, “Improved 
biometric identification system using a new 
scheme of 3D local binary pattern,” International 
Journal of Information and Communication 
Technology, Vol. 14, No. 4, p. 439-455, 2019.  
https://doi.org/10.1504/ijict.2019.101863  
[16] Z. Guo, L. Zhang and D. Zhang, "Feature Band 
Selection for Multispectral Palmprint 
Recognition," 2010 20th International Conference 
on Pattern Recognition, 2010, pp. 1136-1139, 
https://doi.org/10.1109/icpr.2010.284 . 
[17] A. Meraoumia, S. Chitroub, and A. Bouridane, 
“An efficient palmprint identification system 
using multispectral and hyperspectral imaging,” 
In: Modeling Approaches and Algorithms for 
Advanced Computer Applications. Studies in 
Computational Intelligence, vol 488. Springer, 
Cham.                                             
https://doi.org/10.1007/978-3-319-00560-7_20 
[18] V. Roşca and A. Ignat, "Quality of Pre-trained 
Deep-Learning Models for Palmprint 
Recognition," 2020 22nd International 
Symposium on Symbolic and Numeric 
Algorithms for Scientific Computing (SYNASC), 
2020, pp. 202-209, 
https://doi.org/10.1109/synasc51798.2020.00041  
[19] Junwen Sun, Waleed Abdulla, Weiming Wang, 
Qiong Wang, and Hai Zhang, "Band Selection for 
Palmprint Recognition," Journal of Advances in 
Information Technology Vol. 7, No. 4, pp. 287-
290, November, 2016. 
https://doi.org/10.12720/jait.7.4.287-290  
[20] R. Chlaoua, A. Meraoumia, M. Korichi and K. 
Aiadi, "Visible spectrum bands of palmprint 
image for a robust biometric identification 
system," 2016 International Conference on 
Information Technology for Organizations 
Development (IT4OD), 2016, pp. 1-4, 
https://doi.org/10.1109/it4od.2016.7479292  
[21] Z. Guo, D. Zhang, L. Zhang and W. Liu, "Feature 
Band Selection for Online Multispectral 
Palmprint Recognition," in IEEE Transactions on 
Information Forensics and Security, vol. 7, no. 3, 
pp. 1094-1099, June 2012, 
https://doi.org/10.1109/tifs.2012.2189206  
[22] S. Zhao, B. Zhang, and C. L. Philip Chen, “Joint 
deep convolutional feature representation for 
hyperspectral palmprint recognition,” Inf. Sci. 
(Ny)., vol. 489, pp. 167–181, 2019.  
https://doi.org/10.1016/j.ins.2019.03.027. 
[23] S. Zhao, W. Nie and B. Zhang, "Multi-Feature 
Fusion Using Collaborative Residual for 
Hyperspectral Palmprint Recognition," 2018 
IEEE 4th International Conference on Computer 
and Communications (ICCC), 2018, pp. 1402-
1406, 
https://doi.org/10.1109/compcomm.2018.878074
8 
[24] Meraoumia, A., Kadri, F., Bendjenna, H., 
Chitroub, S., Bouridane, A. (2017). Improving 
Biometric Identification Performance Using 
PCANet Deep Learning and Multispectral 
Palmprint. In: Biometric Security and Privacy. 
Signal Processing for Security Technologies. 
Springer, Cham. 
             https://doi.org/10.1007/978-3-319-47301-7_3 
[25] BENSID, Khaled, SAMAI, Djamel, LAALLAM, 
Fatima Zohra, et al. Deep learning feature 
extraction for multispectral palmprint 
identification. Journal of Electronic Imaging, 
2018, vol. 27, no 3, p. 033018. 
https://doi.org/10.1117/1.JEI.27.3.033018 
[26] J. Heikkila and V. Ojansivu, "Methods for local 
phase quantization in blur-insensitive image 
analysis," 2009 International Workshop on Local 
and Non-Local Approximation in Image 
Processing, 2009, pp. 104-111, 
https://doi.org/10.1109/lnla.2009.5278397  
[27] ADJIMI, Ahlem, HACINE-GHARBI, Abdenour, 
RAVIER, Philippe, et al. Extraction and selection 
of binarised statistical image features for 
fingerprint recognition. International Journal of 
Biometrics , vol. 9, no 1, p. 67-80, 2017. 
https://doi.org/10.1504/ijbm.2017.10005054  
[28] Alex Krizhevsky, “ImageNet Classification with 
Deep Convolutional Neural Networks Alex,” 
Handb. Approx. Algorithms Metaheuristics, pp. 
1–9, 2017. 
             https://doi.org/10.1145/3065386  
[29] K. Simonyan and A. Zisserman, “Very deep 
convolutional networks for large-scale image 
recognition,” 3rd Int. Conf. Learn. Represent. 

76 Informatica 47 (2023) 63–80 M. Korichi et al. 
ICLR 2015 - Conf. Track Proc., pp. 1–14, 2015. 
             https://doi.org/10.48550/arXiv.1409.1556 
[30] X. Gu, P. P. Angelov, C. Zhang and P. M. 
Atkinson, "A Massively Parallel Deep Rule-
Based Ensemble Classifier for Remote Sensing 
Scenes," in IEEE Geoscience and Remote Sensing 
Letters, vol. 15, no. 3, pp. 345-349, March 2018, 
https://doi.org/10.1109/lgrs.2017.2787421 . 
[31] K. Nandakumar, Yi Chen, A. K. Jain and S. C. 
Dass, "Quality-based Score Level Fusion in 
Multibiometric Systems," 18th International 
Conference on Pattern Recognition (ICPR'06), 
2006, pp. 473-476, 
https://doi.org/10.1109/icpr.2006.951  
[32] Department of Computing, the Hong Kong 
Polytechnic University (PolyU), Hyperspectral 
Palmprint database, PolyU, available at: 
http://www4.comp.polyu.edu.hk/~biometrics 
/Hyperspectral Palmprrint /HSP .htm. 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

An Effective Hyperspectral Palmprint Identification System Based… Informatica 47 (2023) 63–80 77 
 
 
Table 3: proposed unimodal system performance (handcrafted & deep learning approach)  
Method Modality Band 
EER 
(%) 
T0 
ROR 
(%) 
RPR 
LPQ 27 
Left 
610 (20) 5.2631 0.9683 89.4737 188 
690 (28) 0 . 8 1 7 0 0 . 7 8 1 3 96.2281 171 
850 (44) 0.0877 0.7698 99.5614 17 
940 (53) 0.3508 0.7210 99.1228 92 
Right 
610 (20) 6.4035 0.7194 88.0702 188 
760 (35) 2.2807 0.7539 94.7368 188 
850 (44) 0.0318 0.8173 99.5614 9 
940 (53) 0.3911 0.7444 99.2105 88 
BSIF 
Left 
610 (20) 4.2982 0.6489 90.5263 190 
690 (28) 0.7894 0.7895 96.4912 143 
850 (44) 0.1619 0.7158 99.6491 19 
940 (53) 0.1754 0.7381 99.2982 71 
Right 
610 (20) 5.4385 0.6744 89.2982 188 
760 (35) 1.2595 0.7753 95.8772 187 
850 (44) 0.0877 0.7515 99.5614 6 
940 (53) 0.2154 0.7124 99.5614 91 
CNN-
VGG16 
Left 
610 (20) 4.1228 0.7972 88.1579 180 
690 (28) 1.0337 0.8197 96.1404 98 
850 (44) 0.0877 0.8553 99.3860 147 
940 (53) 0.0877 0.8465 99.4737 108 
Right 
610 (20) 5.5263 0.8395 86.2281 190 
760 (35) 0.4385 0.8436 97.2807 67 
850 (44) 0.1896 0.8371 98.9474 12 
940 (53) 0.3508 0.8175 98.8596 32 
CNN-
VGG19 
Left 
610 (20) 3. 8 1 7 9 0. 8 2 6 5 87.6316 175 
690 (28) 1 . 3 1 5 7 0 . 8 0 5 3 96.0526 123 
850 (44) 0.1754 0.8317 99.6491 93 
940 (53) 0.0877 0.8742 99.3860 69 
Right 
610 (20) 5.4321 0.8389 86.0526 189 
760 (35) 0.5047 0.8267 97.3684 29 
850 (44) 0 . 1 7 5 4 0 . 8 4 2 1 99.1228 21 
940 (53) 0 . 9 4 30 0.8670 99.9123 3 
CNN-
AlexNet 
Left 
610 (20) 1.9298 0.7842 93.5088 113 
690 (28) 0.7017 0.7851 97.8070 165 
850 (44) 0.0877 0.7642 99.9123 55 
940 (53) 0.0877 0.7687 99.7368 
31 
 
Right 
610 (20) 3.0701 0.7839 90.2632 145 
760 (35) 0.2631 0.8091 99.2982 107 
850 (44) 0.0129 0.8701 99.9123 4 
940 (53) 0.0349 0.8513 99.8246 5 
 

78 Informatica 47 (2023) 63–80 M. Korichi et al. 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Table 4: Multimodal open/closed-set test results (first scenario) 
Method Modality Band EER (%) T0 
ROR 
(%) 
RPR 
LPQ 27 
Left 
Bands 
Sum 0.0877 0.7535 99.6491 26 
WSum 0.0877 0.7819 99.6491 15 
Prod 0.7260 0.2372 97.8070 135 
WProd 0.0877 0.7449 99.6491 15 
Right 
Bands 
Sum 0.0352 0.8184 99.6491 10 
WSum 0.2190 0.7505 99.0351 83 
Prod 0.5358 0.3085 98.3333 171 
WProd 8.2667.10
-3
 0.8820 99.5614 7 
BSIF 
Left 
Bands 
Sum 0.0647 0.7137 99.8246 17 
WSum 0.0536 0.7446 99.7368 11 
Prod 0.5263 0.2070 98.5088 187 
WProd 0.2631 0.6821 99.6491 187 
Right 
Bands 
Sum 0.0129 0.8351 99.8246 13 
WSum 0.1754 0.7570 99.5614 75 
Prod 0.2227 0.3485 99.0351 88 
WProd 0.0445 0.7715 99.8246 3 
CNN-
VGG16 
Left 
Bands 
Sum 0.0877 0.7973 99.8246 80 
WSum 0.0877 0.8177 99.7368 36 
Prod 0.0877 0.5151 99.7368 63 
WProd 0.0877 0.8168 99.8246 117 
Right 
Bands 
Sum 0.0459 0.8435 99.2982 5 
WSum 0.0877 0.8288 99.2982 16 
Prod 0.0877 0.5232 99.2982 188 
WProd 0.1754 0.7980 99.0351 188 
CNN-
VGG19 
Left 
Bands 
Sum 0.0877 0.7935 99.9123 41 
WSum 0.0536 0.8296 99.7368 15 
Prod 0.0877 0.4917 99.7368 29 
WProd 0.0877 0.8242 99.9123 69 
Right 
Bands 
Sum 0 . 0 1 2 0 0 . 9 1 3 1 99.4737 2 
WSum 0 . 0 2 9 7 0 . 8 8 8 3 99.1228 6 
Prod 0 . 0 8 7 7 0 . 5 0 0 5 99.6491 188 
WProd 0.1754 0.8036 99.2982 188 
CNN-
AlexNet 
Left 
Bands 
Sum 7.4259.10
-3
 0.8030 99.9123 13 
WSum 5.5111.10
-3
 0.8930 99.7368 5 
Prod 6.4977.10
-3
 0.6360 99.8246 6 
WProd 0.0877 0.7267 99.9123 30 
Right 
Bands 
Sum 1.8564.10
-3
 0.9760 99.9123 2 
WSum 0.0102 0.9431 99.7368 9 
Prod 0 0.8440 100 1 
WProd 3.6741.10
-3
 0.9110 99.8246 3 
 

An Effective Hyperspectral Palmprint Identification System Based… Informatica 47 (2023) 63–80 79 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Table 5: Multimodal open/closed-set test results (second scenario) 
Method Band EER (%) T0 
ROR 
(%) 
RPR 
LPQ 27 
Sum 0 0.9130 100 1 
WSum 0 0.9850 100 1 
Prod 0.0891 0.2130 99.3860 62 
WProd 9.2824x10-4 0.9740 99.9123 2 
BSIF 
Sum 0 0.9600 100 1 
WSum 0 0.9590 100 1 
Prod 0.1754 0.0725 99.7368 184 
WProd 0.1754 0.4195 99.8246 184 
CNN-
VGG16 
Sum 0 0.8210 100 1 
WSum 0 0.8770 100 1 
Prod 0.0877 0.1327 99.9123 186 
WProd 0.0877 0.5821 99.9123 186 
CNN-
VGG19 
Sum 0 0.8050 100 1 
WSum 0 0.8950 100 1 
Prod 0.0877 0.1221 99.9123 184 
WProd 0.0877 0.5761 99.9123 184 
CNN-
AlexNet 
Sum 0 0.7930 100 1 
WSum 0 0.8260 100 1 
Prod 0 0.1510 100 1 
WProd 0 0.7450 100 1 
 
Table 6: proposed system performance comparison with the state-of-the-art 
Method 
Band selection 
evaluation way 
Number of 
selected bands 
Feature 
extraction approach 
EER ROR 
[21] 
 
Clustering (4 
clusters) 
4 Handcrafted 0.0780% / 
[11] / 52 Handcrafted 4.00% / 
[12] 
Clustering (10 
cluster) 
/ Handcrafted 3.26% / 
[19] 
Clustering (3 
clusters) 
3 Handcrafted 0.17325% / 
[13] / All Handcrafted / 99.43% 
[23] / 53 Deep learning 0.11% 99.76% 
[22] / 53 Deep learning 0.015% 99.62% 
[14] Manual 2 Handcrafted / 99.92% 
Proposed 
approach 
Automatically 4 
Handcrafted/deep   
learning 
0.000% 100% 
 

80 Informatica 47 (2023) 63–80 M. Korichi et al.