| The prevalence of mental illness has increased significantly worldwide as a result of the accelerating growth of the global economy and the growing strain of social rivalry,and a younger age trend has been observed in an increasing number of cases and suicides.It is important to pay attention to and research the prevention and treatment of mental disease.The rapid advancement of brain science research and artificial intelligence has resulted in excellent mental disease diagnostic and treatment approaches.Emotion,as a wind vane of human mental state,is inextricably linked to mental health.As a result,electroencephalography(EEG)emotion recognition has emerged as a research priority in the fields of brain science and artificial intelligence.There are numerous study findings in the area of EEG signal emotion detection at the moment,however there are still many obstacles and hurdles in emotional feature extraction,such as inadequate feature information,single-modal data constraints,and feature redundancy.Several studies employ a high number of leads without screening,resulting in data duplication and hardware complexity,which not only impairs recognition accuracy but also makes it difficult to translate results into portable solutions.Given the aforementioned issues,this research investigates an EEG emotion identification approach based on Differential Entropy(DE)characteristics and Convolutional Neural Network(CNN).This research offers an EEG emotion recognition model of SE-DE-CNN based on CNN and SENet(Squeeze and Excitation Network)attention mechanisms to improve the accuracy of emotion recognition.Firstly,the raw data is preprocessed by sliding window segmentation sample and baseline calibration.Then,the differential entropy DE features of each frequency band are extracted.The one-dimensional signal features are mapped to a two-dimensional plane according to the arrangement of electrode positions.The two-dimensional feature matrix of each frequency band is stacked with reference to the RGB organization of the image,and the strong feature extraction ability of CNN is fully utilized to mine the correlation between channels.Moreover,the SENet attention method is implemented after the convolutional layer to improve attention to the goal information while suppressing extraneous feature information.Lastly,the model’s retrieved features are utilized to recognize emotions using the Softmax classification method.Experiments are run on the DEAP dataset to evaluate the proposed model’s performance.For the Valence-Arousal dimension,the average classification accuracy is 91.08% and 91.93%,respectively.The experimental findings suggest that the SE-DE-CNN model developed in this research is capable of accurately identifying emotions.This paper proposes a spatial lead selection algorithm based on Fisher Score to solve the problem of EEG lead selection,in which the value of each lead for judging emotion was measured by the method of weight addition to get the value of the common lead independent of the subject,so as to realize the screening of the lead with strong correlation of emotion.On the DEAP data set,the classification performance of the whole lead set and the filtered lead set is also tested.Finally,the SE-DE-CNN model maintained classification accuracy of 89.82% and 91.10% in the Valence-Arousal dimension,which were only 1.26% and 0.83% lower than the full-lead set,respectively;however,the number of leads was reduced to less than half of the total leads,greatly reducing the complexity and calculation time.The experimental results show that the Fisher Score-based spatial electrode lead selection algorithm can achieve higher emotion recognition accuracy while reducing the number of leads significantly,and the distribution of the selected leads is consistent with the cortical area dealing with general emotional tasks.It serves as a model for decreasing the complexity of wearable device construction and non-invasive emotional state measurement systems in daily life or clinical care. |
PDF Full Text Request