Constructing an image vision transformer for recognizing hand gestures using surface electromyography signals

Abstract

Purpose Interest is growing in developing techniques and methodologies for acquiring and decoding biological signals. Different applications particularly in the area of prosthetic controlling and rehabilitation where an accurate recognition of the hand gesture using surface electromyography (sEMG) signals; are highly needed. sEMG signals are characterized by complex and high-variable information, accordingly extracting useful information from the sEMG signals requires advanced signal processing and data analysis techniques. Methods The present work uses the NinaPro Database 1, as it is an open-source database and used for benchmarking sEMG classifiers. Hand gesture recognition using sEMG signals is commonly approached using algorithms similar to those employed for image classification, which was the motive to build a Vision Transformer (ViT). Three different techniques have been studied in this work for extracting features from the sEMG signals prior classification, these techniques are: extracting the Fast Fourier Transform (FFT), extracting wavelets, and using a pre-trained convolutional neural network (CNN) for feature extraction. Results The findings revealed that the performance of the Vision Transformer (ViT) exceeded that of the majority of CNN architectures employed for classifying sEMG signals. The results also show improvements in the training accuracy when using the wavelet than using the other two feature extractors. The Vision Transformer demonstrated its capability to capture fine structures and integrate global information in imaged signals through plausible basis functions. Additionally, the attention mechanism employed by the Vision Transformer shed light on the significance of electrode readings, offering valuable insights for accurate classification. Conclusion While the majority of research in hand gesture prediction using sEMG signals primarily emphasizes convolutional neural networks, it is crucial to shift attention towards exploring the potential of transformer architecture. This study further highlights the importance of feature extraction and reveals the substantial impact it has on classification accuracy.