4 January 2024 in IEEE Sensors Journal
In recent years, gesture perception has become crucial to human-computer interaction technologies. Among various techniques, gesture recognition based on surface Electromyography (sEMG) signals has gained significant prominence, with deep learning methods playing a pivotal role in this domain. However, as the demand for accurate gesture recognition continues to rise, there is a growing inclination toward selecting complex deep neural network architectures. This trend, however, poses challenges in terms of performance and runtime requirements for computing devices. This paper introduces a novel gesture recognition method utilizing the MLP-Mixer framework combined with Stacking ensemble learning to address these challenges. The proposed method effectively captures the features of sEMG data by employing simple multi-layer perceptrons (MLP), achieving a level of accuracy comparable to complex networks while simultaneously reducing inference time. Experimental results demonstrate that the method performs a classification accuracy of 80.03% and 81.13% for 49 actions in the open-source dataset NinaPro DB2, using window lengths of 200ms and 300ms, respectively. Furthermore, the method achieves a single inference speed of 54.77ms with a window length of 200ms. In NinaPro DB5, with window lengths of 250ms and 300ms, the method presented in this paper achieves accuracy rates of 73.39% and 74.82%, respectively, completing inference within just 11.45ms using the 300ms window length. Notably, the technique also demonstrates its ability to mitigate the impact of individual differences in sEMG data on recognition accuracy.