Abstract:
Sign Language is the essential language for deaf and dumb people. It is a non verbal language used by people with speech and hearing disabilities for communication.
Only a few people can use sign language proficiently and most of the people do not
know how to communicate the deaf people. Therefore, sign language became a boon
for the physically challenged people to express their thoughts and emotion. A system
that can translate is needed when a normal person wants to talk with a dumb or deaf
person. Our proposed system was built to classify 11 static number signs and 30
consonant signs expect 3 dynamic signs for Myanmar Language.
In our proposed system, there are three main processes, namely, preprocessing,
features extraction and classification for those extracted features. In the preprocessing
stage, the input images are cropped manually for only hand regions and resized them.
And then, they are converted into grayscale images. One of the feature extraction
methods in image processing, namely, Kernel Principal Component Analysis (KPCA)
is combined with Supportive Vector Machine (SVM) to implement a Myanmar sign
language recognition system. The main concept of Kernel Principal Component
Analysis is to extract non-linear features. Conventional Principal Component Analysis
(PCA) is effective if data are in the form of linear structure. But it can fail to reduction
data dimensions if data belong to a nonlinear low-dimensional manifold.
For classification of the extracted features, SVM is used. Its goal is to determine
the best decision boundary that can separate classes with less error. Among many
kernels, Gaussian Kernel Radial Basis Function (RBF) is used together with SVM to
classify non-linear data in higher dimension. Data collected from 30 different people
are used as dataset. As a result, KPCA with SVM have the highest accuracy (82%)
compared with Principal Component Analysis.
