Electrocardiograph Based Emotion Recognition via WGAN-GP Data Enhancement and Improved CNN

Abstract

Emotion recognition is one of the key technologies for the further development of human-computer interaction, and is gradually becoming a hot spot in current AI research. At the same time, physiological signals are objective external manifestations of emotions, and emotion recognition based on physiological signals often lacks high-quality training samples and suffers from inter-sample category imbalance. In this paper, 140 samples of electrocardiogram (ECG) signals triggered by Self-Assessment Manikin (SAM) emotion self-assessment experiments were collected using International Affective Picture System (IAPS). To cope with the problem of small data size and data imbalance between classes, Wasserstein Generative Adversarial Network-Gradient Penalty (WGAN-GP) was used to add different number samples for different classes to achieve class balance in the training set, and by continuously adding -samples, the training set to different sizes, using three classifiers to train different sizes of training set samples separately. The results show that the accuracy and weighted F1 values of all three classifiers improve after increasing the data, where higher accuracy and F1 values can be obtained using the proposed Multi Attention-CNN(MA-CNN) as a classifier before and after increasing the samples.