EEG Data Augmentation for Emotion Recognition Using Diffusion Model

Electroencephalogram (EEG) signals are playing an increasingly important role in affective computing, especially in emotion recognition. However, the process of collecting EEG signals is very complex, requiring subjects to conduct long-term experiments under ideal conditions, which has high requirements for both the subjects and the experimental environment. Therefore, how to quickly obtain a large number of high-quality EEG signals has become an issue. In recent years, the success of diffusion models in the field of image generation has attracted a large number of researchers' interest. Compared to traditional generative models, diffusion models have excellent properties such as better generation performance and a more stable training process. In this paper, we apply diffusion models to emotion recognition tasks for the first time. We optimize the sampling stage of the diffusion model to make it more suitable for generating high-quality EEG signals, and then enhance the original EEG signals and apply them to emotion recognition tasks. Additionally, we explore the impact of the quantity of generated data on task performance. We use three datasets in the experiment, SEED, SEED-IV, and DEAP datasets. The experimental results indicate that our data augmentation method can significantly improve emotion recognition.