Deep Attention Network for Enhanced Hand Gesture Recognition System

Since technology is growing in all fields, researchers are developing more advanced technologies in human–computer communication and security systems. The flexible wrist hinge and crowded background make it difficult to distinguish hands in uncontrolled conditions. We proposed a deep attention network, that can automatically localize the hand and classify the gestures. Available datasets were collected to train the deep neural network architecture to create a gesture recognition system. The datasets used are 20-bn-jester and NTU-HD datasets. 20-bn-jester dataset is RGB dataset and NTU-HD dataset is RGB-D dataset. The features are extracted from the dataset and fed into the models for training and classification. Two models Resnet-50 and VGG-16 are used for training and classification. In this work, a special network used for localizing hand regions is a deep attention network. Using an attention network, gesture recognition system has been created for static RGB-D images. The proposed architecture yielded a better accuracy of 91.32% using Resnet-50, and 89% using VGG-16, respectively. The significance of these results shows that one can create a hand gesture recognition system feasibly and prominently.