EEG-Based Emotion Recognition Using Convolutional Neural Networks

Brain-computer interface(BCI)is a technology which can build communication between a human brain and an external device without neuromuscular muscles.It realizes the human-computer interaction by recognizing the Electroencephalography(EEG)signals.This technology has a good application prospect in many fields,such as military,medical and artificial intelligence.Many studies have achieved good results in BCI field,emotional EEG recognition is one of them.The emotions of humans can always be easily hidden,because the external features such as facial expressions and hand movements are easily suppressed.But the physiological signals like EEG signals and heart rate are difficult to disguise.Therefore,the study of emotional EEG is recognization of different emotions by analyzing the EEG signals of the subjects.This technology can be widely used in business,polygraph,medicine and other fields.How to improve the recognition accuracy of different emotions becomes an important problem of BCI field,and it is worthy to investigate.Our study mainly includes:1.The BCI and emotional EEG signals is introduced in detail.The traditional emotional EEG recognition methods and feature extraction methods are briefly summarized.In order to compare the advantages and disadvantages of deep learning and traditional methods,we design two Convolutional Neural Net(CNN)structures,called Shallow ConvNet(SCN)and Deep ConvNet(DCN).We used the EEG signals as the input of the network directly,and compare to the classification results of different artificial extraction features with SVM.The results show that the classification of CNNis higher than most of the feature extractions,near to the widely used Differential Entropy(DE)features.2.Two experiments are designed in this paper,the single-subject and cross-subjects experiments.In these experiments,the CNN classify the three kinds of emotional EEG signals(positive,neutral,negative).Then,we also use EEG samples of different frequency bands and different time lengths as CNN inputs,and analyze the experimental results.The results of high frequency band inputs are better than the low frequency band,and the full frequency band has the best performance.In response to the specific differences between the three emotions,we also designed several two-category experiment of emotional EEG to compare the recognition of different emotions.The positive emotion can be easily recognized and the negative emotion is the worst.To verify the performance of the proposed CNN,two emotional EEG datasets,SEED and DEAP,are used for our study.Comparing to the most of traditional methods,the proposed CNN has a good improvement in classification performance in single-subject and cross-subjects experiments.It means that CNN can be used in the emotional recognition study and has good research significance.