Multiclass Least-Square Support Vector Machine for Myoelectric-Based Facial Gesture Recognition

Abstract

Facial gesture recognition (FGR) is considered as a state-of-the-art which has drawn the researchers’ attention in numerous fields of study due to its high potential in different applications. Recognizing the gestures through bio-signals generated from facial muscle movements has been recently proposed as an accurate and reliable pathway. The performance of gesture recognition-based systems directly depends on the effectiveness of classification techniques. Besides, a reasonable trade-off between recognition accuracy and computational cost is counted as the most significant factor for designing such systems. The aim of this paper was the classification of facial gestures electromyogram (EMG) signals by means of a least square support vector machine (LS-SVM) algorithm. Ten predefined facial gestures EMGs were recorded from ten participants through three bi-polar channels. Acquired signals were preprocessed using a band-pass filter and a segmentation technique. Then, time-domain features mean absolute value (MAV) and root mean square (RMS) were extracted from each segment. In order to classify the features, LS-SVM was implemented by considering radial basis function kernel and two multiclass encoding schemes, one-versus-one (OVO) and oneversus- all (OVA). This research showed that LS-SVM was a robust method for classification of facial gestures with 97.1% classification accuracy and 1.37 seconds training time when utilizing the feature combination MAV+RMS and the encoding technique OVA. It was also concluded that LS-SVM outperformed SVM and fuzzy c-means classifiers in this field of study. The results of this paper can be used as efficient processing tools in designing reliable interfaces for FGR systems.