Page 162 - 2024-Vol20-Issue2
P. 162
158 | Murad & Alasadi
results, including supervised, unsupervised, semi-supervised, gestures.
reinforcement, transduction, and learning to learn [48]. The HMMs have high accuracy and can define several gestures but
chosen learning algorithm is largely determined by the hand require training and performance improvements. Each motion
gesture representation chosen. or posture must define the precise number of states. HMMs
Most available hand recognition technologies don’t gather all are the ideal choice when the types and quantity of gestures
the data or are static, making them inappropriate for many or postures are pre-defined. However, retraining may be more
circumstances. Different gesture detection methods rely on time-consuming when gestures and hand postures are deter-
strategies from modeling image processing, computer vision mined during system development. The hidden Markov model
science, and pattern recognition. Popular methods for dis- requires significant training time, making knowing what is
tinguishing between dynamic and static hand movements in- happening inside it difficult. Despite high precision, HMMs
clude: are considered the best method due to their high precision,
reaching over 90% [55].
A. Artificial Neural Networks
Artificial neural networks (ANNs) are non-linear algorithms C. Fuzzy Clustering Algorithms
that extract features through hidden layers and classify them Patterns can belong to various data sets due to fuzzy-clustering
by multilayer perceptrons [49]. These ANNs’ approximate algorithms’ flexible grouping of the data [56]. Xingyan [57]
functions receive large numbers of unknown inputs. The unveiled a device that recognizes hand motions using a fuzzy
nodes in ANNs are the core units of the brain, and links con- c-means clustering technique in a mobile remote. The sys-
nect nodes with a weight representing a way to store data. tem performs pre-processing procedures such as removal of
ANNs have been used to address problems in speech recogni- noise, removing undesired items, and threshold approaches.
tion, Visual-Interpretation, and Roberts-Control. Most ANNs It also extracts the shape of the hand from camera images and
are implemented in sequential machines but can work in par- converts them to HSV space color. The aspect ratio is the
allel machines, making them suitable for applications in these bounding box for the thirteen items that comprise the features
fields. Several researchers have used ANNs for gesture recog- vector, and grid cells are represented by grid cells. The FCM
nition, such as self-learning and big data processing. Some algorithm executes six motions with twenty language samples
have used ANNs in classification processes for gestures, while in various situations, including those with consistent lighting
others have used them to segment the hand. Maung [50] intro- and complex backgrounds. The recognition accuracy rate of
duced a system for tracking hands and recognizing Myanmar- the system is 85.83
Alphabet-Language (MAL) by NNs, using an Adobe Photo-
shop filter to find the edges of the hand image. A system D. K-Nearest Neighbor
for recognizing static hand motions was developed by Ster- K-NN is a statistical technique for classifying objects in fea-
giopoulou [51] utilizing SGONG Networks (Self-Growing ture space using the nearest training instances. It uses locally
and Self-Organized Neural Gas). Higher-level 2D features, approximating functions and postpones calculations to classi-
such as angles, palm positions, and the number of lifted fin- fication [58]. The object is categorized by a preponderance of
gers, are extracted from hand topology to recognize static votes from its neighbors, assigning it to the class with the most
hand gestures. common k nearest neighbors. K-NN is easy to implement and
The method Ismail et al [52] developed uses two recurrent understand, and its object property value can be assigned to
NNs to recognize Arabic Sign Language. Segmentation was the average of the k nearest neighbors’ values for regression.
carried out using the Elman network, partially utilized in- The neighbors are selected from the objects group for which
dependently of the recurrent and fully recurrent NNs. For the right classification is defined.
Elman-NN and fully recurrent NN, respectively, the recogni-
tion rate was 89.67% and 95.1%. E. Support Vector Machine
Vladimir Vapnik introduced the support vector machine (SVM),
B. Hidden Markov Models a nonlinear classifier that improves classification by mapping
In the mid-1990s, Hidden-Markov models (HMM) were intro- data to a high-dimensional space [59]. SVM classifiers are
duced as a solution to the segmentation problem in recognition similar to neural networks, with a two-layer perceptron neural
tasks. HMMs are non-deterministic and stochastic, with one network equivalent. SVM models represent examples as space
or more arcs carrying the same value [53]. They are useful points separated by the widest possible distance and classify
in speech and sign language recognition, as they can handle them according to their side of the gap. SVMs are designed
the limitations of Markov chains. Sarma and Bhuyan [54] as binary classifiers, but other methods treat multi-class prob-
introduced a hand gesture recognition system using real-time lems as one optimization problem. Using SVMs as binary
hand tracking and HMM for recognizing three-dimensional
