Recognition of Micro-Expressions from Nonlinear Features of Surface Electromyography and Random Forest

Micro-Expression Recognition (MER) plays a pivotal role in comprehending concealed emotions which aids in identifying mental health issues. Recent advancements in MER use surface electromyography (sEMG) for detecting Micro-expressions (MEs), but is challenging due to low amplitude and its transient properties. In this study, an attempt has been made to classify micro-expressions using nonlinear features of sEMG and random forest. For this purpose, myoelectric signals and facial expressions are collected from eight participants while they are exposed to emotion eliciting videos, which focuses on happiness, sadness, fear, and anger. The signals are obtained from, four muscle regions: Mentalis (MT), Zygomaticus (ZM), Frontalis (FT), and Orbital (OT). These signals are pre-processed and segmented to micro-expressions using Facial Action Unit Coding (FACS) of facial expression. Nonlinear features, namely, Shannon, approximate, sample and permutation entropies are extracted and used to design random forest model. There are 2084 MEs obtained from our measurements: 94 happiness, 396 sadness, 915 fear, and 679 anger. All features except Shannon entropy show significant difference among the emotions. Synthetic Minority Over-sampling Technique is used to address the issue of data imbalance. The proposed model yields a sensitivity of 82%, specificity of 93%, accuracy of 81% and F1 score of 82%. It appears that nonlinear characteristics of sEMG could be useful in spotting micro-expressions.