Pruning CapsNet for Hand Gesture Recognition with sEMG Signal Based on Two-Dimensional Transformation.

The sEMG signal-based recognition is broadly used in the field of human-computer interaction. To improve the signal classification accuracy, this paper proposes a two-dimensional transformation pruning capsule network (TDPCAPS) to recognize different hand gestures. To apply deep learning methods to signal classification, a two-dimensional transformation method is proposed, which converts feature vectors into two-dimensional feature data. Moreover, using the capsule network to explore the characteristics of the sEMG signal, as this model overcomes the defect that the convolution neural network fails to capture the correlation among features. However, the capsule network requires a lot of computing resources, so this paper adopts a pruning mechanism to reduce the number of coupling coefficients and speed up the calculating process. In the experiments of electrode displacement and several subjects, the recognition accuracy of TDPCAPS reaches 84.92% and 80.31%, respectively. Meanwhile, the classification time for a window is reduced by 11.39%. The experimental results show that the proposed method can ensure recognition accuracy and improve computational efficiency at the same time.