Recognizing Bengali Sign Language Gestures for Digits in Real Time Using Convolutional Neural Network

Abstract—Recognizing sign language gestures for different
languages has been found as a promising field of research that
explores the possibility of communication by interpreting various
signs and translating them into text or speech. Establishing a
better communication way between deaf-mute people and ordinary
people is the prime objective of this research arena. There
are many existing Sign Language Recognition (SLR) systems
throughout the world and these SLR systems are implemented
using various methods, tools and techniques with a view to
achieving better recognition accuracy. This research work aims
at applying the concept of Convolutional Neural Network (CNN)
for recognizing Bengali Sign Language gesture images for digits
only in real time. Bengali sign language images for digits are
collected from different individuals and the CNN model is trained
with these images after performing several pre-processing tasks
i.e. resizing to a specific dimension, converting these RGB
images to the gray scale images, finding the equivalent binary
images and rotating the images into different degrees both in
left and right direction. The experiment is conducted using
two major techniques. Firstly, the model has been trained with
the dataset containing the equivalent binary images of the row
images collected directly from different individuals. Secondly,
the dataset is enriched by rotating all the images into 3°, 6°,
9°, 12°and 15°in both left and right directions. After applying
the rotation technique, the recognition accuracy is found to be
increased significantly. The maximum recognition accuracy of
the proposed CNN model with the dataset without image rotation
technique is 94.17% whereas the recognition accuracy is 99.75%
while including the rotated images in the dataset.
Index Terms—Bengali Sign Language Recognition, Convolutional
Neural Network, Image Recognition, Real Time Recognition