Air-Writing Recognition on GPU accelerated Neural Network Architecture

Abstract

Air-writing refers to the idea of writing the characters in free-space by the movement of hand-palm or fingers, or any writing device like a pen. In this paper, we use a generic web-camera to recognize isolated characters scripted in the air with a neural network approach. Air-Scripting is performed using a circular-like object that acts as a marker of a fixed color in front of a camera. Color-based segmentation is applied to identify the center and trace its trajectory that forms a character. The recognition is carried out by Multilayer Perceptron and Convolution Neural Network approaches. The performance depends on various illumination conditions owing to color-based segmentation. In a less fluctuating lightning condition, the system is able to recognize isolated upper-case letters with a recognition rate of 91.53% and 96.66% and lower-case letters with a recognition rate of 87.42% and 89.98% for MLP and CNN respectively.