Abstract
Emotion is a generic term for a set of subjective cognitive experiences, a mental state and a physiological state resulting from a combination of multiple sensations, thoughts and behaviours. Emotion recognition has a wide range of applications in the medical field, distance education, security and health detection, healthcare, and human-robot interaction. We use ECG signals for emotion recognition, but the difficulties are that it is difficult to obtain high quality physiological signals about emotions and the small sample data make it impossible to train a classifier with high accuracy. To address these problems, we propose to use data augmentation to solve the problem of small samples by adding target detection and target loss to WGAN-GP to control the intra-class distribution of the generated data to achieve intra-class balance in the training set, i.e., T-WGAN-GP network. We validated the effectiveness of our proposed model using ECG signals from the DEAP dataset in terms of two evaluation criteria, Accuracy (ACC) and Weighted F1 (WF1).