Deep Convolutional Neural Network For Featureless Electromyogram Pattern Recognition Using Time-Frequency Distribution
Feature extraction is an essential step to extract useful information from electromyogram (EMG) signal in the classification of upper limb movements. However, the process of feature extraction and selection require expert knowledge and experience. Therefore, this paper proposed a new approach for an...
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | en |
| Published: |
American Scientific Publishers
2018
|
| Subjects: | |
| Online Access: | http://eprints.utem.edu.my/id/eprint/23005/2/Deep%20Convolutional%20Neural%20Network%20for%20Featureless%20EMG%20Pattern%20Recognition%20Using%20Time-Frequency%20Distribution.pdf http://eprints.utem.edu.my/id/eprint/23005/ https://www.ingentaconnect.com/content/asp/senlet/2018/00000016/00000002/art00002 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Feature extraction is an essential step to extract useful information from electromyogram (EMG) signal in the classification of upper limb movements. However, the process of feature extraction and selection require expert knowledge and experience. Therefore, this paper proposed a new approach for an efficient classification of hand movements without the need of manual feature extraction. A deep convolutional neural network (CNN) architecture was developed to learn the feature automatically from the data. Two time-frequency distributions, Spectrogram and Gabor Transform (GT) were employed to transform the signal into time-frequency images. The time-frequency images were then directly fed into the CNN for classification. From the analysis, spectrogram and GT images were able to successfully recognize the ten different hand movements. Our results showed that GT images achieved the highest mean classification accuracy of 97.59%, followed by spectrogram, 97.13%. In order to validate our results, the performance of proposed CNN was compared with a feature based learning method. When compared to support vector machine (SVM), our recognition system consistently showed a higher performance. Moreover, the proposed CNN was tested using the EMG data of amputee subjects obtained from Nina Pro database. The results suggest that CNN is more appropriate to be applied in the rehabilitation and engineering application. |
|---|