Similarity constraint style transfer mapping for emotion recognition. (February 2023)

• Record Type: Journal Article
• Authors: Chen, Lei; She, Qingshan; Meng, Ming; Zhang, Qizhong; Zhang, Jianhai
• Is Part Of: Biomedical signal processing and control. Volume 80(2023)Part 1
• Publication Date: 2023-02
• Subjects: Brain-computer interfaces -- Transfer learning -- Emotion recognition -- Style transfer mapping; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2022.104314
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: The proposed algorithm maintains the local structure of the target domain in the transfer process. It can achieve good performance in the case of less labeled target data. Experimental results show that SCSTM can achieve better performance than most other transfer algorithms in SEED, SEED-IV and DEAP. Abstract: Transfer learning plays a vital role in emotion recognition based on electroencephalogram (EEG). In practical application, only little labeled data from the target subject can be obtained, so there is still a problem of how to solve the situation of no large amount of unlabeled data from the target subject. Therefore, this paper proposes a novel method of similarity constraint style transfer mapping (SCSTM) and domain selection strategy with geodesic flow kernel (DSSWGFK). When calculating the mapping matrix, SCSTM maintains the local structure of the target domain by constraining the similarity of the distance among the samples of the target subject before and after mapping, which further makes use of the existing data to reduce the demand for the quantity of data from target subject. DSSWGFK obtains the weights of source domain classifiers in the ensemble classifier based on the similarity between the target subject and each source domain, which makes full use of the source domain data and reduces the demand for the quantity of data from the target subject. Experimental results show that our SCSTM method can achieve better average classification accuracy, 1.14%, 6.84% and 8.77% higher than that of supervised STM on SEED, SEED-IV and DEAP, respectively. Furthermore, DSSWGFK is capable in improving the performance of SCSTM. Finally, it can be concluded that the proposed method has achieved superior performance for emotion recognition.