Page 136 - 2024-Vol20-Issue2

132 | Fathi & Aziz

Fig. 6. Positions of the 21 hand key points [18]

TABLE II.
PROPERTIES OF HAND GEOMETRY FEATURES

Property Description Number Fig. 7. (a – e) Geometric representation of feature set
Line slop slop of lines 9

Angles of knuckle between 21
two knuckle
Width first calculate the vector 3
proximal phalanx between any two knuckles 5
then find the angles 5
length between each two vectors.
Finger length proportion of palm width
to horizontal distance
between two knuckle
proportion of palm length
to proximal phalanx length
from each finger
proportion of palm length
to length of each finger

network model. A supervised artificial neural network that Fig. 8. Summary of the proposed ANN architecture
was described in Section D was trained and tested using 79
right palmar images from our own established dataset belong- proportion of correctly classified inputs, as depicted in (3).
ing to 14 families. All data were randomly split into two The confusion matrix of the results is shown in Fig. 10 and
groups: training data and testing data, where training data Table III. Our goal in this work is not to attain high classifi-
make up 80% of the dataset and testing data make up 20%(63 cation network accuracy, but to demonstrate the viability of
images for training, 16 for testing). 43 features were extracted kinship identification using human hands. We aspire, through
from each image (3397 features in total) and fed into the ANN our subsequent experiences, to improve result by identifying
for training and testing purposes. Evaluating the results of a effective features, using data augmentation techniques, train-
classification task using a neural network involves a variety ing the ANN model with features extracted from more than
of metrics and strategies to determine how well the model is one aspect of the hand and using deep learning algorithms.
performing. Accuracy is a common metric that calculates the