Real-Time EEG-Based Facial Expression Recognition Using Laplacian Energy Features on FPGA

Electroencephalogram (EEG)-based facial expression recognition plays an important role in affective computing and brain–computer interface (BCI) systems; however, it is conceptually distinct from general emotion recognition. EEG signals recorded during facial expressions are strongly influenced by motor and sensorimotor activity associated with facial movements, whereas emotion recognition aims to decode internally generated affective states arising from distributed limbic–cortical networks. At sensor level, EEG signals inherently exhibit low signal-to-noise ratio (SNR), high temporal variability, and nonlinear spatial dependencies across electrodes, which further degrade the reliability of affective decoding. Conventional feature extraction techniques capture local time–frequency information but fail to preserve inter-electrode spatial topology, leading to poor generalization across subjects and sessions and limiting real-time embedded deployment. To address these challenges, this work introduces a Graph Signal Processing (GSP)-based Laplacian energy (LE) feature extraction framework integrated with lightweight machine learning classifiers, explicitly modeling spatial–topological dependencies among EEG channels and enabling efficient, interpretable, and real-time affective state recognition across multiple frequency bands. EEG features are classified using Random Forest (RF), Support Vector Machine (SVM), Decision Tree (DT), Logistic Regression (LR), K-Nearest Neighbors (KNN), and Light Gradient Boosting Machine (LightGBM), achieving 100% accuracy with cross-validation mean accuracy above 99.89%. Implemented on the FPGA (PYNQ-ZU) platform, the system demonstrates 6–8 mW power consumption and sub-millisecond latency. In contrast, the proposed LE + LR model achieves 100% accuracy with only 0.007 W power and 0.08 ms latency—representing a 20×–500× gain in power efficiency and a 10×–2000× latency reduction over existing FPGA-based methods.