Periodic Shifted Patching on Vision Transformer for Electroencephalogram Signal-Based Emotion Recognition

Recognizing emotions from electroencephalogram (EEG) signals is vital for applications ranging from clinical diagnostics to human-computer interaction. Despite extensive research, current methods often rely on expensive handcrafted features or computationally intensive signal-domain transformations that struggle with the inherent noise and non-stationarity of EEG data. Furthermore, achieving high accuracy on limited datasets remains a challenge due to the overfitting tendencies of deep learning models. To address these issues, we propose PeViT, an efficient and effective model designed for raw EEG-based emotion recognition. We introduce a Periodic Shifted Patching (PSP) mechanism that provides on-the-fly data augmentation while preserving critical spatio-temporal relationships without the need for prior signal processing. This is integrated with a modified Vision Transformer architecture tailored to capture global contextual relations and long-term dependencies in EEG signals. We validated PeViT on four datasets, including the recent and large-scale HBUED (50 participants), alongside DEAP, DREAMER, and SEED. The results demonstrate that PeViT scales effectively to larger cohorts, achieving over 90% accuracy on HBUED while maintaining state-of-the-art performance and a 30% speedup on established benchmarks. This underscores the model's robustness in capturing spatio-temporal relationships across diverse subjects with the lowest parameter count among the compared models The source code for this work is available at: https://github.com/w4w4n78/PeViT.