MGMS: A Modality-General and Modality-Specific Learning Model Using EEG and Facial Expression for Hearing-Impaired Emotion Recognition

Currently, most research on emotion recognition primarily relies on single-modal physiological or nonphysiological methods, overlooking the complementarity of emotion representation across different modalities. Individuals with hearing impairments may experience emotional cognitive biases due to the loss of the emotional acquisition pathway associated with hearing. Therefore, this study introduces the modality-general and modality-specific (MGMS) learning model, which aims to examine the emotions of hearing-impaired individuals in four categories (fear, happy, neutral, and sad) through the fusion of electroencephalogram (EEG) and facial expression. Specifically, the differential entropy (DE) features are manually extracted from each EEG channel by different brain regions, and then the spatial information is captured by a long short-term memory (LSTM) network. In terms of facial expression, texture features and geometric features are combined which are extracted by the ResNet network and 68 facial key points, respectively. By constructing a general-specific discriminator, the MGMS features are separated from the two modes. Furthermore, a Transformer encoder is employed to classify the four features using a cross-entropy loss function. Experimental results demonstrate that the proposed methods achieve an average classification accuracy of 86.01% for subject-dependent classification, surpassing the respective accuracies of 65.12% for EEG and 59.86% for facial expressions.