Research On EEG Emotion Recognition Based On Deep Learning

Emotion is a complex physiological and psychological activity,it is closely related to brain response and plays an essential role in social life.Therefore,it is of great practical value to use existing techniques to study emotion recognition.Emotion recognition based on EEG is also a challenging topic in pattern recognition research.It is difficult to find a method that can effectively represent the temporal and spatial features of EEG signals.Therefore,this thesis proposes the following two emotion recognition methods based on deep learning.Firstly,a parallel spatial-temporal projection convolutional neural network is proposed,which mainly includes temporal stream sub-network,spatial stream subnetwork and fusion classification module.Temporal stream extracts temporal continuity features through sequence-projection layer,spatial stream captures spatial correlation features by channel-projection layer,and the fusion classification module combines the extracted spatial and temporal features into a joint spatial-temporal feature vector for emotion prediction.Both sequence-projection and channel-projection adopt lengthsynchronized convolutional kernel to decode whole time and space information.The size of length-synchronized convolutional kernel is equal to the length of transmitted EEG sequence.In addition,considering that data preprocessing contributes to model training,a baseline removal method based on sorting filtering is designed to correct the baseline offset of emotional signals.In order to increase the size of the training set,a data augmentation strategy of randomly exchanging corresponding channels among similar samples is proposed,and the best exchange intervals are explored through experiments.Experimental results on DEAP and DREAMER emotion database show that the proposed method performs superior performance in emotion recognition task and surpasses other advanced methods.On this basis,this thesis further proposes an global spatial-temporal fusion memory analysis emotion recognition model.The model is composed of following three functional modules: global spatial-temporal feature stack(GSFS),multiscale spatial-temporal fusion(MSF)and deep Bi-LSTM memory analysis(DBMA)modules.The GSFS module consists of two symmetrical sub-networks,which are responsible for extracting the local and global spatial-temporal information from the input signals,and dynamically stacking local and global information into the global spatial-temporal features combined with the original signals.The MSF module performs multi-scale fusion and deep excavating on the extracted global spatial-temporal features from the three dimensions of spatial,temporal and spatial-temporal,as well as the global and local perspectives.The DBMA module is composed of two-layer bidirectional long short term memory networks.The deep emotional features after spatial-temporal fusion are memorized and recursively analyzed.Finally,the emotional features are classified through the full connection layer.Experimental results show that the performance of the proposed method is better than the current mainstream emotion recognition methods.