Research on Hearing-Impaired EEG Emotion Recognition Based on Deep Residual Shrinkage Networks

With the development of affective computing and computer science, electroencephalogram (EEG) based emotion recognition has attracted much more attention. In this paper, we collected EEG signals from fifteen healthy people and fifteen hearing-impaired people when they were watching five kinds of emotional pictures (fear, anger, sadness, happiness and neutral). The collected EEG signals are preprocessed to remove artifacts by independent component analysis (ICA). Then the differential power spectral density, entropy, and wavelet entropy features were extracted (PSD, DE, WE). The Deep Residual Shrinkage Networks (DRSNs) were composed of residual building units (RBU s) and an attention mechanism was used to capture the representative features, and then made a classification. The classification results prove that using DE as a feature for classification is better than other features, with an average accuracy being 76.35% and 80.47% for the hearing-impaired group and the normal group, respectively. Moreover, from the brain topography, we found that there are certain differences in brain function between hearing-impaired and healthy people.