| Emotion classification can be applied in many fields,such as cognitive intervention and auxiliary medical treatment,which has significant research value.The EEG signal as the physiological carrier of emotion signal has the advantages of non-invasive acquisition,high temporal resolution,and portable and inexpensive required equipment compared with other emotion signals.Processing and analyzing the EEG signal has become one of the most promising directions of affective computing.However,there are still several challenges in EEG-based emotion recognition.First,the acquisition cost of EEG signals is high,which requires professional experts,equipment,and a strict acquisition process.Second,due to the rich emotional categories and the complex labeling system,the learning cost required for labeling data is relatively high.Third,labeling emotion data is highly subjective,and the obtained label can be used only when different annotators return the same.The above challenges all lead to insufficient high-quality labeled data in EEG-based emotion recognition.To address the above mentioned challenges,this thesis proposes an "unlabeled EEG representation-EEG data augmentation-EEG emotion classification" framework for EEGbased emotion representation and recognition in weakly-supervised scenarios.(1)This thesis proposes a semi-supervised EEG-based emotion feature representation method based on contrastive learning,which fully leverages unlabeled EEG data to improve the feature representation ability.Specifically,based on the temporal instability of biological signals,this method builds an auxiliary task to assign pseudo labels for unlabeled EEG data,and then uses a recurrent training strategy to reduce the noise interference caused by unlabeled data.At the same time,to alleviate the problem of over-fitting and improve the model generalization,this thesis introduces a label smoothing regularization technique to revise the labeled data learning.(2)This thesis proposes an EEG-based emotion classification method based on the generative adversarial network,which applies the data augmentation for labeled samples to improve the performance of classification.Specifically,the method uses the learned discriminative EEG features as actual samples to learn synthetic EEG samples,and then introduces an emotion classifier to maximize the inter-class differences between EEG fake samples.Additionally,to avoid the problem of information degradation,this method applies an emotion regressor to reconstruct the semantic representation of EEG fake samples to achieve semantic consistency across modalities.(3)To further verify the practicability and generalization of the method in real-world scenarios,this thesis constructs a dataset by recruiting 20 volunteers to watch the pictures from "Peace and War" and conducts the experimental evaluation of the method.The experimental results show that the proposed method has generalizability and can effectively identify human emotions in real scenes.The research in this thesis proposes an effective method to address the lack of labeled samples in EEG-based emotion analysis in practical applications.The proposed method reduces the requirements for labeled data and improves the performance of emotion recognition.Furthermore,this thesis expands the practical application scenarios of EEG-based emotion analysis and provides a solid theoretical analysis,which has great significance and value in this field. |
