Predicting and accurately classifying intentions for human hand gestures can be used not only for active prosthetic hands, rehabilitation robots and entertainment robots but also for artificial intelligence robots in general. In this paper, first of all, source data of three hand gestures of grasping and three hand gestures of sign language are acquired by using the armband combined with eight sEMG (surface Electromyography) sensors. To classify these hand gestures, basically simple CNN (convolutional neural network) models with raw data, short-time Fourier transform (STFT), wavelet transform (WT), and scale average wavelet transform (SAWT) are applied, and their performances are compared. Finally, it is shown that by using a CNN with SAWT images, the accuracy can be improved up to 94.6% for selected hand gestures with higher accuracy and lower computational burden than conventional multi-channel STFT or WT.

中文翻译：

---

基于多通道肌电图的尺度平均小波变换和卷积神经网络手势分类

预测和准确分类人类手势的意图不仅可以用于主动假手、康复机器人和娱乐机器人，还可以用于一般的人工智能机器人。本文首先利用臂带结合8个sEMG（表面肌电图）传感器获取三个抓握手势和三个手语手势的源数据。为了对这些手势进行分类，应用了具有原始数据、短时傅立叶变换 (STFT)、小波变换 (WT) 和尺度平均小波变换 (SAWT) 的简单 CNN（卷积神经网络）模型，并比较了它们的性能. 最后，结果表明，通过使用带有 SAWT 图像的 CNN，准确率可以提高到 94。