sEMG based hand gesture recognition with deformable convolutional network

There is a growing interest in human machine interface and their applications using surface electromyography (sEMG). sEMG based gesture recognition plays a crucial role in interfacing with peripheral devices such as prosthetic hands. Give the challenges in the state of the art of sEMG based gesture recognition using deep learning, we propose a deformable convolutional network (DCN) to optimise the conventional convolution kernels with a goal of achieving better performance of sEMG based gesture recognition. The DCN first apply traditional convolutional layer to obtain low-dimensional feature maps, then use deformable convolutional layer to get high-dimensional feature maps. Moreover, we propose and compare two new image representation methods based on traditional feature extraction, which enable deep learning architectures to extract implicit correlations between different channels from the sparse multichannel sEMG signals. The experiments are conducted to evaluate the proposed methods on three groups of different types and numbers of gestures on the Ninapro-DB1 data set, the proposed DCN has an improvement of 1.1%, 2.6%, and 4.9% compared with traditional CNN, respectively. In addition, the results of experiments indicate that the DCN shows robustness and feasibility in both feature extraction and classification recognition for the sEMG based gesture recognition.