Exploring Enhanced Recognition in Gesture Language Videos Through Unsupervised Learning of Deep Autoencoder

The primary objective of a neural network is to achieve generalization, enabling it to recognize previously unseen data from the same category. This concept facilitates the transfer of knowledge between neural networks trained for different purposes. In this study, we investigate the utilization of deep Autoencoder neural networks for unsupervised learning as a regression network to improve the performance of a categorization network. We propose a novel deep Autoencoder network designed to accomplish this goal. Our proposed architecture is applied to video data, specifically focusing on hand gesture recognition using the Chalearn 2014 and IsoGD datasets. Hand gesture recognition plays a crucial role in human-machine interactions, despite challenges posed by temporal information, gesture overlap in videos, and variations in hand gesture orientation. Through our research, we have successfully enhanced the recognition accuracy from 46.1% to 76.4% by employing transfer learning techniques within the same dataset and across different datasets.