Reserach On Emotion Recognition Of EEG Based On Convolutional Recurrent Neural Network

Affective computing is an important part of human-computer interaction.With the rapid development of artificial intelligence,many researchers have joined the research in the field of emotion recognition.Emotional recognition mainly focuses on physiological signals and non physiological signals.Among various physiological signals,EEG signals are closely related to emotions and can not be controlled subjectively,which can better reflect people’s emotions.This article focuses on the emotional classification of EEG signals based on convolutional recurrent neural networks.The specific research content is as follows:Firstly,the SKNet Inception Convolution Neutal Network(SKICNN)model is studied.The model is composed of SKNet module and Inception module.The input of the model is a three-dimensional signal constructed by differential entropy features of different frequency bands.Multiband combination is verified by experiment than less frequency combination accuracy is high,and use the baseline signal can improve the accuracy,at the same time to verify the steam used in 1 × 1,4 × 1,1 × 4 and 4 × 4 convolution kernel the highest accuracy.Compared with previous research results,the accuracy of SKICNN model is improved.In DEAP date set,the accuracy of arousal dimension is 94.36% and the accuracy of valence dimension is 93.41%.Secondly,the model based on the 4D EEGNet Recurrent Neural Network(4DEEGRNN).The model is composed of EEGNet module and Bi LSTM module.The 4D feature is the use of 0.5 second window segmentation EEG signal,and the three-dimensional signal composed of the differential entropy features of the four frequency bands is calculated as the input of the model.By constructing a comparison model,the influences of different inputs and different parameter Settings on the overall results are analyzed.Found by experiments with 9 × 9 × 4 as input,lost way for Spatial Dropout2 D presented.according to model for 0.1 K works best in all models.Compared with previous research results,the accuracy of 4D-EEGRNN model is improved.In DEAP date set,the accuracy of arousal dimension is 95.97% and the accuracy of valence dimension is 95.23%.Finally,a 2D-3D Convolution Neural Network(2-3DCNN)model is proposed.The model consists of two dimensional convolution,three dimensional convolution,SE-Res Net module,Xception network and deep residual contraction network.The input of the model is the original EEG signal.The influence of different positions of the deep residual shrinkage network on the final result is analyzed by experiments.It is found that the accuracy of the deep residual shrinkage network and SE-Res Net is the highest when they are in the same branch.Compared with previous research results,the accuracy of 2-3DCNN model is improved.In DEAP date set,the accuracy of arousal dimension is 97.19% and the accuracy of valence dimension is 97.58%.