Compact Deep Neural Networks for Computationally Efficient Gesture Classification From Electromyography Signals

Abstract

Machine learning classifiers using surface electromyography are important for human-machine interfacing and device control. Conventional classifiers such as support vector machines (SVMs) use manually extracted features based on e.g. wavelets. These features tend to be fixed and non-person specific, which is a key limitation due to high person-to-person variability of myography signals. Deep neural networks, by contrast, can automatically extract person specific features -an important advantage. However, deep neural networks typically have the drawback of large numbers of parameters, requiring large training data sets and powerful hardware not suited to embedded systems. This paper solves these problems by introducing a compact deep neural network architecture that is much smaller than existing counterparts. The performance of the compact deep net is benchmarked against an SVM and compared to other contemporary architectures across 10 human subjects, comparing Myo and Delsys Trigno electrode sets. The accuracy of the compact deep net was found to be 84.2 ± 6% versus 70.5 ± 7% for the SVM on the Myo, and 80.3 ± 7% versus 67.8 ± 9% for the Delsys system, demonstrating the superior effectiveness of the proposed compact network, which had just 5,889 parameters -orders of magnitude less than some contemporary alternatives in this domain while maintaining better performance.

Deep learning for processing electromyographic signals: A taxonomy-based survey

“…Hence, a deeply investigation of data augmentation techniques in EMG signal processing represents a completely open and unexplored field. At the moment, the majority of the authors primarily focused on a continuous improvement of the classification/regression performance [192], whereas, in the authors' opinion, a big effort should be put 1) in improving the generalization ability of the proposed models by introducing new data augmentation methods and 2) in investigating compact deep topologies to shorten both learning and execution time while maintaining high performance levels [94].…”

“…After a careful review of all papers within this class, CNNs revealed to be the most commonly used networks, followed by AEs, RNNs, and DBNs (see Figure 5). According to the increasing popularity of CNNs in several research fields due to their proven high performance, several authors proposed classifiers based on CNNs only [14,155,94,183,174,18,199,161,163,171,7,64,120,22,180,192,13,213,68,56,52,70,140], or on both CNNs-RNNs [81,200,198,191], or on CNN-AE [219]. Some authors have alternatively developed multi-class classifiers entirely based on deep AEs [222,110,128,164,163,3], RNNs [175,98,181] or DBNs [170,169].…”

An sEMG-Based Human-Exoskeleton Interface Fusing Convolutional Neural Networks With Hand-Crafted Features

“…Atzori et al (2016) applied convolutional neural network to sEMG data classification, and the proposed framework classification accuracy was higher than the average accuracy obtained by classical methods, with the highest accuracy reaching 87.8%. In the literature (Hartwell et al, 2018), a compact deep neural network architecture is used, which still achieves a classification accuracy of 84.2% even though the parameter values of other networks are several orders of magnitude less. Duan et al (2019) applied multi-channel convolutional neural network to sEMG dataset for gesture recognition, and the recognition accuracy was 90%.…”

“…In fact, deep convolutional neural networks (CNNs) can automatically extract appropriate features from the data. Some researchers have designed deep learning-based HRIs, which can achieve higher lower limb movement prediction performance than hand-crafted features (Atzori et al, 2016;Hartwell et al, 2018;Duan et al, 2019;Burns et al, 2020). However, due to the large amount of data required by deep learning, when predicting small datasets (such as the lower extremity motion dataset for a single hemiplegic patient), the lower-limb movement prediction method based on deep learning often suffers from overfitting.…”

A temporal-to-spatial deep convolutional neural network for classification of hand movements from multichannel electromyography data

“…We found here that on a high performance workstation GPU, an NVIDIA Geforce 1080Ti, the computation time was on the order of 1-4 ms to process a single forward pass through the TtS network, which is likely to be sufficient for realtime implementation. Of more relevance to embedded systems, CNNs of a similar design have been implemented by the authors in [63], on an NVIDIA Jetson Tx2 (embedded system), with times of around 20 ms for a single forward pass that can be reduced to ∼ 8 ms using network compression. This suggests that deep CNNs can be implemented currently at usable sample rates in both modern computational settings and embedded systems.…”