LIGHT-WEIGHTED CNN-ATTENTION BASED ARCHITECTURE FOR HAND GESTURE RECOGNITION VIA ELECTROMYOGRAPHY Soheil Zabihiy Elahe Rahimianz Amir Asify and Arash Mohammadiz

2025-05-03 0 0 511.42KB 5 页 10玖币

侵权投诉

LIGHT-WEIGHTED CNN-ATTENTION BASED ARCHITECTURE FOR HAND GESTURE

RECOGNITION VIA ELECTROMYOGRAPHY

Soheil Zabihi†, Elahe Rahimian‡, Amir Asif†, and Arash Mohammadi‡

†Electrical and Computer Engineering, Concordia University, Montreal, QC, Canada

‡Concordia Institute for Information System Engineering, Concordia University, Montreal, QC, Canada

ABSTRACT

Advancements in Biological Signal Processing (BSP) and Machine-

Learning (ML) models have paved the path for development of novel

immersive Human-Machine Interfaces (HMI). In this context, there

has been a surge of signiﬁcant interest in Hand Gesture Recogni-

tion (HGR) utilizing Surface-Electromyogram (sEMG) signals. This

is due to its unique potential for decoding wearable data to inter-

pret human intent for immersion in Mixed Reality (MR) environ-

ments. To achieve the highest possible accuracy, complicated and

heavy-weighted Deep Neural Networks (DNNs) are typically devel-

oped, which restricts their practical application in low-power and

resource-constrained wearable systems. In this work, we propose

a light-weighted hybrid architecture (HDCAM) based on Convo-

lutional Neural Network (CNN) and attention mechanism to effec-

tively extract local and global representations of the input. The pro-

posed HDCAM model with 58,441 parameters reached a new state-

of-the-art (SOTA) performance with 82.91% and 81.28% accuracy

on window sizes of 300 ms and 200 ms for classifying 17 hand ges-

tures. The number of parameters to train the proposed HDCAM ar-

chitecture is 18.87×less than its previous SOTA counterpart.

Index Terms—Attention Mechanism, Biological Signal Pro-

cessing (BSP), Mixed Reality (MR), surface Electromyogram.

1. INTRODUCTION

Surface Electromyogram (sEMG)-based Hand Gesture Recognition

(HGR) is regarded as a promising approach for a wide range of

applications, including myoelectric control prosthesis [1–4], virtual

reality technologies [5, 6], Human Computer Interactions (HCI) [7],

and rehabilitative gaming systems [8]. sEMG signals contain elec-

trical activities of the muscle ﬁbers that can be employed to decode

hand gestures and thereby enhance immersive HMI wearable sys-

tems for immersion in Mixed Reality (MR) environments [9, 10].

Consequently, there has been a surge of interest in the development

of Deep Neural Networks (DNNs) and Machine Learning (ML)

models to identify hand gestures using sEMG signals. Generally

speaking, sEMG datasets can be collected based on “sparse mul-

tichannel sEMG” or “High-Density sEMG (HD-sEMG)”. Despite

advantages of HD-sEMG, its utilization leads to structural complex-

ity [11, 12], while adoption of sparse multichannel sEMG signals

requires fewer electrodes making it the common modality of choice

for incorporation into wearable devices. Therefore, development of

DNNs based on sparse sEMG signals has gained signiﬁcant recent

importance.

Despite extensive research in this area and the fact that academic

researchers achieve high classiﬁcation accuracy in laboratory condi-

tions, there is still a gap between academic research in sEMG pattern

recognition and commercialized solutions [9]. In this context, one of

the objectives for reducing the gap is to focus on the development of

Fig. 1:Comparing different variants of our proposed HDCAM model with

SOTA designs for an input window size of 300 ms. The x-axis shows the

number of parameters and the y-axis displays the classiﬁcation accuracy on

the Ninapro DB2 dataset. Our HDCAM shows a better compute versus accu-

racy trade-off compared to recent approaches.

DNN-based models that not only have high recognition accuracy but

also have minimal processing complexity, allowing them to be em-

bedded in low-power devices such as wearable controllers [1, 13].

Furthermore, the designed DNN-based models should be based on

the minimum number of electrodes while estimating the desired ges-

tures within an acceptable delay time [9, 14]. Consequently, in this

paper, we develop the novel Hierarchical Depth-wise Convolution

along with the Attention Mechanism (HDCAM) model for HGR

based on sparse sEMG signals to ﬁll this gap by meeting criteria such

as improving the accuracy and reducing the number of parameters.

The HDCAM is developed based on the Ninapro [15, 16] database,

which is one of the most well-known sparse multi-channel sEMG

benchmark datasets.

Using Convolutional Neural Networks (CNN) [17–20] is a

common approach for hand movement classiﬁcation, where sEMG

signals are ﬁrst converted into images and then used as input for

CNN-based architectures. However, the nature of sEMG signals

is sequential, and CNN architectures only take into account the

spatial features of the sEMG signals. Therefore, in recent litera-

ture [12, 13, 21], authors proposed using recurrent-based architec-

tures such as Long Short Term Memory (LSTM) networks to exploit

the temporal features of sEMG signals. On the other hand, it is

suggested [22–24] to use hybrid models (CNN-LSTM architecture)

instead of using a single model to capture the temporal and spatial

characteristics of sEMG signals. Although recent academic re-

searchers are improving the performance by using Recurrent Neural

Networks (RNNs) or hybrid architectures, the sequence modeling

with recurrent-based architectures has several drawbacks such as

consuming high memory, lack of parallelism, and lack of stable

gradient during the training [4, 25]. It is demonstrated [26] that

sequence modeling using RNN-based models does not always out-

perform CNN-based designs. Speciﬁcally, CNN architectures have

several advantages over RNNs such as lower memory requirements

and faster training if designed properly [26]. Therefore, in the recent

literature [4, 25, 27, 28], the authors took advantage of 1-D Convolu-

tions developed based on the dilated causal convolutions, where the

arXiv:2210.15119v1 [cs.LG] 27 Oct 2022

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

LIGHT-WEIGHTEDCNN-ATTENTIONBASEDARCHITECTUREFORHANDGESTURERECOGNITIONVIAELECTROMYOGRAPHYSoheilZabihiy,ElaheRahimianz,AmirAsify,andArashMohammadizyElectricalandComputerEngineering,ConcordiaUniversity,Montreal,QC,CanadazConcordiaInstituteforInformationSystemEngineering,ConcordiaUniversity,Montreal,QC,...

展开>> 收起<<

LIGHT-WEIGHTED CNN-ATTENTION BASED ARCHITECTURE FOR HAND GESTURE RECOGNITION VIA ELECTROMYOGRAPHY Soheil Zabihiy Elahe Rahimianz Amir Asify and Arash Mohammadiz.pdf

共5页,预览1页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

LIGHT-WEIGHTED CNN-ATTENTION BASED ARCHITECTURE FOR HAND GESTURE RECOGNITION VIA ELECTROMYOGRAPHY Soheil Zabihiy Elahe Rahimianz Amir Asify and Arash Mohammadiz

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: