Page 133 - 2024-Vol20-Issue2

129 |
accuracy rate [15]. Hand geometry systems are based solely
on their shape, not on their fingerprints, so the reader can
even read dirty and low-resolution hand images [5]. Hand
geometry-based biometric systems use the numerous features
retrieved from hand photographs to carry out tasks like identi-
fication or classification. Feature extraction can be performed
using manual; handcrafted methods [17] [18] [19] or auto-
mated; deep learning methods [20] [21]. By learning from
experience, ML enables computers to be programmed without
explicit experience [22]. A recent development in machine
learning is deep learning [23].Deep learning algorithms are
better suited for large datasets, require high-end machines,
feature engineering, a longer execution time, and are more
interpretable and problem-solving-oriented than traditional
methods [24]. The length, breadth, and aspect ratio of the
fingers are common geometrical properties, as well as the
width of the palm [4].
Kinship detection using computer vision is a new topic, and
it has been used and studied for the last several years [25].
Most research in this field checks if two people are from the
same family or not through facial images. The first attempt at
kinship verification using human traits was made by Fang et
al. by automatically classifying pairs of 150 pairs of parents
and children’s face images as ”related” or ”unrelated”, The
classification accuracy was 70.67% using KNN as a classi-
fier [26]. From that point forward, endeavours continued in
this field, which were all in light of facial images. From this
literature review, we concluded that there were no previous
works that have used the image of hands for kinship detection
(Family-based person classification). In addition to that, there
were no available datasets for this purpose. So, the research
questions were, Can the image of hands be used for kinship
detection? How? What are the useful features? and what are
the appropriate methods to achieve this goal? This cannot be
achieved without providing a suitable image dataset attached
by its ground truth labelling. Hand geometry features and ma-
chine learning techniques were adopted to validate the usage
of the created dataset.
III. METHODOLOGY
This paper proposes a new method to determine kinship re-
lations based on hand biometric features using handcraft fea-
tures extraction and supervised classifier. The proposed work
was to design and implement a persons’ kinship classification
system based on the images of their hands. This proposal
has not been addressed before. None of the available datasets
have ground truth relating to kinship relations, so we had
collected our own dataset. No previous work specified the
most beneficial region or features of the hand to be used for
that purpose therefore, many experiments have been done to
find out that. In this research we also studied and investigated involved in hand positioning during the acquisition process.
Fathi & Aziz
the most useful way for classification. In addition to that, a
suitable image acquisition setup, the type of imaging device
or devices and the pose of the hand had been investigated. Per-
son identification and verification is the way to find a unique
feature for each person, while gender classification is to sep-
arate the input images into two classes only. The proposed
research differs from both, it needs to find the membership
of a person to one of many groups or clusters. Framework of
ANN proposed model shown in Fig. 2.
A. Dataset Creation
The dataset collection and metadata file were prepared during
the image acquisition stage. Lighting, distance, direction for
image capturing were selected practically from many experi-
ences. Finally, following setup was used:
• A mobile phone application was prepared to be used for
data collection and saving.
• Image capture setup was sat up for image capturing
which consisted of mobile phone, holder, basement.
• The parameters of setup are empirically selected for
practical image photography.
• The parameters of setup are empirically selected for
practical image photography.
• Images are saved in jpg format.
For each individual 8 different images were captured (left
hand-dorsal, left hand-palm, left hand-dorsal with open fin-
gers, left hand-palm with open fingers) and the same for the
right hand. Fig. 3 shows example of the generated dataset.
Several experiments have been done, many images for many
families have been collected and some of them were rejected
for different reasons like (lighting problems, image noise,
clipped images, bad resolution, and others).The useful im-
ages were saved and marked using a systematic code which
contains relation of person to a family, age, hand side (right/
left , dorsal/palm), as characterized in Fig. 4. Dataset was
collected by the researcher and his colleagues by applying
the same procedure. Each image was accompanied using an
Excel file with ground truth and information about each family
(No. of person, No. of male, No. of female, age range, and
the relation to other families). The generated dataset consists
of 648 hand images of 81 subjects, they were (39 male and
42 female) belong to 14 families. The subjects have varying
ages between 3 - 70 years old. Fig. 5 presents the number
of families, the total number of people in each family, and
the gender breakdown for each family. The major challenges
are associated with non-uniform illumination and changes
B. MediaPipe
Google created the open-source MediaPipe framework. The
initial release took place in 2019 [30] is a computer vision