Machine-learning approaches for recognizing muscle activities involved in facial expressions captured by multi-channels surface electromyogram

Facial expression recognition plays an important role on mimicking and synchronizing person's mental activity. Various approaches have been employed for recognition on facial expression, like facial image and video analysis. Bio-signals as very important biological information reflect biological features. With the development of sensor technology and machine learning, collecting bio-signals to study facial expressions is practicable. Surface-Electromyogram is a way to collect EMG signal by sticking a signal collector on the surface of skin, much environment interference can be ignored and privacy can be protected. Inconvenience of collecting these kinds of bio-signal resulted in lacking of good publicly available datasets. In this paper, we have designed an facial expressions recognition system based on sEMG signals using Intel Edison board with advantages of high temporal resolution, potential flexibility of testing devices. The paper studies facial expressions comprehensively, abstracts 8 common types of facial expressions from 2 kinds of classes and establishes a new bio-EMG dataset with 1680 instances. The facial expressions are including three periodic expressions (chewing, speaking, gargling) and five transient expressions (sadness, surprise, happiness, pout, and angry). For recognition accuracy, by utilizing three different classifiers, Cubic SVM, Cubic KNN, and Gaussian SVM, each classifier has a good performance on expressions classification, the Cubic SVM classifier has the best performance of these three with accuracy as high as 99.52%. With the decreasing of sample amount for training model, the Cubic SVM classifier still performs well in classification with accuracy of 86.9%.