Improved sEMG signal classification using the Twin SVM

Locomo-Net: A Low -Complex Deep Learning Framework for sEMG-Based Hand Movement Recognition for Prosthetic Control

“…To ensure unbiased comparison, we have evaluated the performance based on the average accuracy of the able-bodied and amputated participants.The results are shown in Table IV, which show that the LoCoMo-Net has approximately 8.5% increase in classification accuracy compared to Twin-SVM for all the 15 movements for 250ms input window length. One of the obvious strength of this study is that it performs unsupervised feature extraction and learning while earlier comparable studies [15] used one feature, i.e., root mean square (RMS). The traditional methods of sEMG based movement classification rely on handcrafted feature selection followed by extraction before classifying the intended movements as shown in Fig 1. As a result, this process put a significant amount of time due to the human intervention and exertions in finding suitable features, and feature extraction increases computational time and complexity.…”

“…On the other hand, other approach like multi-modal where sEMG classifier's outputs are combined with inertial measurement unit (IMU) and locational identifiers [13][14] for better accuracy, but again their outcomes are limited to people performing specific activities in repetitive fashion. Thus, the desired system should be a myoelectric one with few electrodes and capability to identify several movements so that can provide the users In this context, many attempts have been made to improve the recognition of myoelectric based movements for the prosthetic hands [15][16][17][18][19][20][21]. These works can largely be categorized into two focal areas: (i) feature selection and (ii) classifier selection.…”

“…These works can largely be categorized into two focal areas: (i) feature selection and (ii) classifier selection. The work by Soman et al [15] used Twin-Support Vector Machine (SVM) for recognizing 15 wrist and finger flexion actions from the muscles of forearm using 4 channels of sEMG recording achieving average of 82% classification accuracy. Naik et al [16] performed experiments to select the appropriate SVM kernel with 4 channels with similar accuracy (82.0%) but with only for 7 movements.…”

“…In this context, many attempts have been made to improve the recognition of myoelectric based movements for the prosthetic hands [15] – [21] . These works can largely be categorized into two focal areas: (i) feature selection and (ii) classifier selection.…”

A Control System of SEMG Signal Based on Deep Learning

“…The model architecture of CNN is shown in Fig 4. The process uses 8-way electrode array as the input, after convolution and pooling processing, the convolution kernel size is 3 × 3, the pooling kernel is 2 × 2, and the ReLu activation function is selected. After the expansion process, the final step is to the full connection layer step to achieve classification output [24][25][26][27]. See Fig 5 for the way to match the output results to the manipulator.…”