Digital Gesture Recognition based on Deep Capsule Network

Abstract

Because gesture recognition is more and more useful in life, gesture recognition should not only consider the changes of gestures, but also be affected by the environment. Therefore, there are high requirements for the gesture recognition algorithms. Previously, image processing methods were used for gesture recognition, but the recognition accuracy of traditional methods was not high. The Capsule Network Model (CapsNet) proposed by Hinton has better recognition performance than traditional networks due to the use of dynamic routing algorithms. Based on CapsN et, this paper designs and realizes a deep capsule network model (DCNet), change the size of the convolution kernel to reduce the amount of parameters, and adds a convolution to the original capsule network structure to extract more feature information, and increase the dimension of the capsule from 8 to 16 to improve the decision-making ability of the capsule. Make the recognition accuracy of the network on a SIGN digital gesture data set higher. Using DCNet to test on the SIGN data set, the detection accuracy of training learning can reach 99.62%. The experimental show that the test accuracy of DCN et is higher than CapsN et for testing with digital gesture data set.