Few Sample Emotion Recognition Based On Multiple Physiological Signals

Emotion recognition is one of the key technologies for the further development of humancomputer interaction and is gradually becoming a hot spot in current AI research.Emotion recognition refers to the ability to recognize emotions based on the physiological characteristics of the body,expressions,words,sounds,and body postures.Among them,physiological signals,as the external expression of emotion,are not easily disguised and have objectivity.Emotion recognition based on physiological signals is one of the widely recognized directions.The main research content is emotion recognition based on EEG signals and telepathic signals in different emotional states.There is a disadvantage of insufficient sample size in emotion recognition based on physiological signals.Among them,the problems of category imbalance and intra-class imbalance are prevalent in emotion recognition.Based on these problems research is conducted around physiological signal processing,data enhancement,and enhanced data classification performance validation.(1)Physiological signal processing.An electroencephalogram(EEG)with valence arousal labels in DEAP(Database for Emotion Analysis using Physiological Signals)dataset was acquired and differential entropy features were extracted from the EEG as a validation dataset.Electrocardiogram(ECG)signals from the emotion induction experiment using the International Affective Picture System(IAPS)were collected,the validity of the SAM self-evaluation was used as a label,and the ECG was filtered and feature extracted as the second validation dataset.(2)Data enhancement.Target-Wasserstein Generative Adversarial Networks-gradient penalty(T-WGAN-GP)is proposed to generate samples for data augmentation to solve the problem of class imbalance and intra-class imbalance.We use the target factor detection to get the target samples in the samples and add a penalty term to the discriminator’s objective function to control the intra-class distribution of the generated samples so that the generated samples can achieve intra-class balance and thus obtain high-quality generated samples.(3)Validation of the classification performance of the augmented data.To verify the effectiveness of data augmentation with T-WGAN-GP.Both SVM(Support Vector Machines)and Deep Neural Networks(DNN)are used as classifiers.WGAN-GP(Wasserstein generative adversarial networks-gradient penalty)and T-WGAN-GP are used as data augmentation models,respectively.The performance metrics are compared before and after adding samples using different classifiers and using different data augmentation models.The results show that data augmentation using both augmentation models with a DNN classifier improves the accuracy and weighted F1 values than without data augmentation,where the accuracy and weighted F1 values are better when using T-WGAN-GP for data augmentation than using WGAN-GP,and the classification effect using SVM is better than that using DNN.