En-Estrnet: Contrastive Learning for Cross-Subject Emotion Recognition with Temporal Depth Inception and Attention-Inspired Feature Enhancer

Recent research highlights the potential of using contrast learning to overcome electroencephalograph(EEG) cross-subject emotion recognition. Nevertheless, the potential to capture intricate emotional nuances remains an unexplored domain of research. This work proposes an Enhanced Emotional Spatio-Temporal Representation Network (En-ESTRNet) to address this deficiency. The temporal analysis module utilizes temporal depth inception, employing four convolutional layers with varying kernel sizes, to discern small nuances in time. The temporal and spatiotemporal analysis modules both include a feature enhancer that is motivated by attention. The similarity analysis module computed the similarity of EEG spatiotemporal features among participants who were subjected to the same emotional stimuli. Subsequently, inter-subject differences were minimized by maximizing inter-subject similarity through contrastive learning. En-ESTRNet has undergone extensive experimental validation on four public datasets (FACED, SEED, SEED-V, and DEAP), resulting in accuracy improvements of 2.94%, 4.31%, 6.67%, and 4.39% compared to the baseline model. The experimental results show that En-ESTRNet achieves state-of-the-art performance in cross-subject emotion recognition.