Sri Lankan Sign Language Tutor
No Thumbnail Available
Date
2013
Journal Title
Journal ISSN
Volume Title
Publisher
Uva Wellassa University of Sri Lanka
Abstract
Sign Language Recognition is one of the major research areas of Human Computer Interaction
(HCI). A large number of researches had been done in this area for American Sign Language,
Indian sign language, Chinese sign language, Thai sign language, etc. (Rajathi et al., 2013). But,
less amount of researches are done related to Sinhala sign language recognition. Specially no
research work found in developing a tutor for Sinhala sign language.
There may have problems when teaching sign language to disabled children due to lack of
teachers, less/no attention to every child at every moment due to lack of resources, parents of
these disabled children may be too busy, less interest of children to study, etc. Therefore, this
research was carried out to develop an automated tutor to the Sri Lankan deaf community for
Sinhala sign language to practice & check their knowledge.
Methodology
This is an Image based Sign Recognition System which uses Fourier Descriptors for feature
Extraction (Nixon et al., 2002). The system architecture can be divided into five modules as
shown in Figure 1.
In Image Acquisition Module, images of 200*200dpi (dots per pixel) resolution which contains
only the hand was captured in a black background by wearing a black long sleeve top with a
white glove. In Image Processing Module, image processing techniques will be applied to
manipulate the image. It results the pure contour of the shape as the final output.
In Feature Extraction Module, Fourier Transformation is applied on closed contour. For that the
shape contour is sampled to 64 points by equal arc length method and Centroid distance
function is applied. Discrete Fourier Transformation is applied to the Centroid distance shape
signature which result 32 Fourier Descriptors. The Fourier Descriptors were modified in order
to preserve scale and rotation invariance by dividing them by the first Fourier Descriptor. The
first 10 Fourier Descriptors except the first Fourier Descriptor are taken into account.
In Neural Network Module, a neural network of 10 input layers, 6 output layers and 8 output
layers is used. The 10 Fourier Descriptors of each sign are fed as the input and a pattern for each
sign was fed as the output. It will output the results as a neuron weight file. This weight file is
used in Result Module to determine the correctness (right/wrong) of each sign to the user. The
result will be displayed to the user by a two visual indicators.
Result and Discussion
Sri Lankan Sign Language Tutor was implemented to recognize the correctness of 8 static signs
of Sinhala Alphabet (Figure 2). The system is tested with 800 images of signs including 100
images from each sign for scale, rotation, translation and starting point invariance and the
obtained accuracy level for each letter is listed in Table 1.
The system is trained by using 200 images of signs including 25 images from each sign. The
obtained accuracy level can be increased by increasing the training set. But it requires more
time.
Description
Keywords
Computer Science, Science and Technology, Technology, Language, Software Developing