Facial neuromuscular signal classification by means of least square support vector machine for MuCI

•Recognizing ten facial gestures through analyzing facial neuromuscular signals.•EMG analysis including filtering, segmentation, feature extraction (RMS).•Classification of facial gestures by multi-class LS-SVM.•Tuning the kernel parameters automatically and manually.•Constructing different LS-SVM models.•48 automatic and 8 manual LS-SVM models were constructed.•The models were compared in terms of classification accuracy and complexity.•The best performance was gained by the model tuned manually including RBF and OVA.•LS-SVM was compared with popular classifiers FCM, SVM and FGGC.

Facial neuromuscular signal has recently drawn the researchers’ attention to its outstanding potential as an efficient medium for Muscle Computer Interface (MuCI) applications. The proper analysis of such electromyogram (EMG) signals is essential in designing the interfaces. In this article, a multiclass least-square support vector machine (LS-SVM) is proposed for classification of different facial gestures EMG signals. EMG signals were captured through three bi-polar electrodes from ten participants while gesturing ten different facial states. EMGs were filtered and segmented into non-overlapped windows from which root mean square (RMS) features were extracted and then fed to the classifier. For the purpose of classification, different models of LS-SVM were constructed while tuning the kernel parameters automatically and manually. In the automatic mode, 48 models were formed while parameters of linear and radial basis function (RBF) kernels were tuned using different optimization techniques, cost functions and encoding schemes. In the manual mode, 8 models were shaped by means of the considered kernel functions and encoding schemes. In order to find the best model with a reliable performance, constructed models were evaluated and compared in terms of classification accuracy and computational cost. Results reported that the model including RBF kernel which was tuned manually and encoded by one-versus-all scheme provided the highest classification accuracy (93.10%) and consumed 0.98 s for training. It was indicated that automatic models were outperformed since they required too much time for tuning the parameters without any meaningful improvement in the final classification accuracy. The robustness of the selected LS-SVM model was evaluated through comparison with Support Vector Machine, fuzzy C-Means and fuzzy Gath-Geva clustering techniques.

Graphical abstractFigure optionsDownload full-size imageDownload as PowerPoint slide