Realization and Validation of a Wearable Prototype for sEMG-based Facial Expressions Recognition

Facial expressions form a rich communication code, enabling the conveyance of emotions and intentions. Their classification has garnered increasing attention in research, as it has yielded promising results in various applications, such as psychological state assessment, control of human-machine interfaces, and rehabilitation of facial muscle impairments. Facial Expression Recognition (FER) can be performed through several techniques. The approach based on the surface ElectroMyoGraphic (sEMG) signal, distinct from the more commonly used computer vision strategies, offers direct insights into muscle activity, holding great potential for rehabilitative and diagnostic purposes. Preliminary studies, from a research team’s prior thesis work, have shown high average accuracy in real-time recognition of 11 facial expressions, thanks to the implementation of an Artificial Neural Network (ANN). The classifier receives 5 inputs corresponding to the Average Threshold Crossing (ATC) parameter, which is computed for the selected muscles and is proportional to their activation. This thesis project has aimed to propose an evolution of the previous prototype by implementing a low-consumption, battery-powered wearable system, inspired by the acquisition devices developed by the research team. The initial phase of the work focused on conducting preliminary tests to assess two primary aspects. Firstly, it involved testing various configurations of gold-coated dry electrodes to find a setup that offered an optimal balance between electrode size and signal quality, which is crucial for achieving high classification accuracy. Secondly, it included evaluating the most appropriate amplification levels for the programmable gain amplifier integrated into the system. The core of the work involved the realization of two Printed Circuit Boards (PCBs): one for active probes, intended to be placed directly on the skin in correspondence of the selected 5 facial muscles, and the other for a main board, integrated into a mechanical support and worn on the head. The active probes were realized with a rigid-flex stack-up and provide initial conditioning of the sEMG signal, which is acquired by the gold-coated electrodes. The main PCB, which is based on the AmbiqMicro Apollo3 Blue microcontroller, was designed to include further signal conditioning, power management, and serial communication circuits, as well as a Flash memory and an Inertial Measurement Unit (IMU), for further research applications. Subsequently, the mechanical support, a key element to ensure the wearability of the system, was designed and 3D printed, to offer a good fit for different subjects while ensuring proper allocation for the main PCB, the battery, and all the system components. Finally, the work has focused on adapting firmware and software for user-friendly system operation, allowing visualization of sEMG signals, ATC values, and expression prediction. The evaluation of the sEMG Signal to Noise Ratio (SNR), acquired from facial muscles, yielded an average value of 25 dB. Ultimately, the performance of the classifier was assessed, showing average accuracy exceeding 90% in the classification of 8 facial expressions. These results reinforce the prototype suitability for future rehabilitative applications, leveraging biofeedback and serious games for facial muscle impairment treatment.