Fig 1 - uploaded by Jingwei Too
Content may be subject to copyright.
Source publication
In recent days, electromyography (EMG) pattern recognition has becoming one of the major interests in rehabilitation area. However, EMG feature set normally consists of relevant, redundant and irrelevant features. To achieve high classification performance, the selection of potential features is critically important. Thus, this paper employs two re...
Similar publications
Citations
... A dimensionality reduction technique was applied after feature selection, which did not present relevant difference in classifier accuracy and a fair comparison between the approaches. Others works as compared some RD and FS techniques but outnumbered or make qualitative analysis of feature combination [16,[23][24][25][26]. ...
Gesture recognition by surface electromyography (sEMG) signals is used for several applications as prosthesis control and human-machines interfaces. One of trending approaches to sEMG acquisition is the multiple-channel armband with equidistant electrodes. Several efforts are made to improve the performance of these devices in gestures recognition applications, especially in feature selection. It is necessary choose one approach for feature selection due to the great number of electromyography features available to use. In this work, an extensive comparison of feature reduction techniques and their influences in the classification process is presented. Unlike other works, we presents the comparison between methods of feature selection and classification; this main contribution is show how the feature reduction process can aid and increase the performance of classification of sEMG in armband acquisition approach. Two general methods were employed, feature selection by wrapper forward stepwise and dimensionality reduction, resulting in eight different techniques. The following dimensionality reduction techniques were used: Principal Component Analysis, Linear Discriminant Analysis, Isomap, Manifold Charting, Autoencoder, t-distributed Stochastic Neighbor Embedding, and Large Margin Nearest Neighbor (LMNN). Seven classifiers were used, aiming at recognize six gestures acquired from an 8-channel armband of 13 subjects. An average accuracy of 89.4 % was obtained with 5 features and an Extreme Learning Machine classifier, in the feature selection approach. On other hand, considering 40 dimensions, an average accuracy of 94 % was obtained, regarding a combination between Support Vector Machine with Gaussian kernel and a LMNN technique. These results showed significant differences in statistical tests.
