Machine learning algorithms for facial gesture recognition: a first analysis based on event-driven sEMG acquisition

Facial gesture recognition has wide application in Human-Machine Interaction (HMI), which, in the medical area, can be identified with behavioral and emotional analyses, as well as rehabilitative procedures. Although historical approaches for facial expression recognition rely on videos and images data, in recent years, with the progress of the sensors technology and Machine Learning (ML) algorithms, recognition is also being achieved using biological signals, as surface ElectroMyoGraphic (sEMG) signal. The thesis focuses on recognizing and classifying jaw movements and facial expressions from sEMG signals recorded by face muscles during the execution of such actions. The innovative event-driven technique, named Average Threshold Crossing (ATC), is applied to the amplified and filtered sEMG signal to extract the ATC feature. This feature is computed by averaging the events generated when an sEMG signal exceeds a voltage threshold on a predefined time window. Past works demonstrated the benefits of the ATC technique in terms of reduction of data processing, transmission and related power consumption, allowing it to be an optimal solution in the development of wearable, miniaturized and energy-efficient data acquisition system. With the aim to develop an ATC-based facial network, the thesis’goal is to understand whether the ATC approach is suitable for the recognition of facial gestures. A first step towards this direction was to define which were the movements to be recognized, the corresponding musculature and, consequently, the electrodes position for proper signals detection. A preliminary test, beside confirming the feasibility of this idea, was needed to organized the sensors location of the facial network better. At this point, with the goal to train facial expression classifiers, a data collection was launched, involving 21 subjects. Each subject performed a list of eight gestures for different session in order to obtain a robust dataset. The raw sEMG signals have been recorded using the g.HIamp-Research amplifier; then, data were processed to extract the ATC parameter, used as input of a classifier. In fact, several Machine Learning algorithms have been implemented to recognize facial movements: Random Forest (RF), Support Vector Machine (SVM), k-Nearest Neighbour (k-NN), and Artificial Neural Networks (ANN). All the four classifiers perform the recognition obtaining similar accuracies. In particular, they reached an overall percentage of success greater than 60% when recognizing eight expressions. In comparison, they improve their average recognition rate up to 75% when two not well-defined expressions are removed from the dataset. These percentages will pave the way to the application of the ATC technique to facial gesture recognition.