Pose invariant hand gesture recognition using two stream transfer learning architecture

Abstract

The hand gesture detection problem is one of the most prominent problems in machine learning and computer vision applications. Many machine learning techniques have been employed to solve the hand gesture recognition. These techniques find applications in sign language recognition, virtual reality, human machine interaction, autonomous vehicles, driver assistive systems etc. In this paper, the goal is to design a system to correctly identify hand gestures from a dataset of hundreds of hand gesture images. In order to incorporate this, decision fusion based system using the transfer learning architectures is proposed to achieve the said task. Two pretrained models namely ‘MobileNet’ and ‘Inception V3’ are used for this purpose. To find the region of interest (ROI) in the image, YOLO (You Only Look Once) architecture is used which also decides the type of model. Edge map images and the spatial images are trained using two separate versions of the MobileNet based transfer learning architecture and then the final probabilities are combined to decide upon the hand sign of the image. The simulation results using classification accuracy indicate the superiority of the approach of this paper against the already researched approaches using different quantitative techniques such as classification accuracy.