Scholarlink

<p>Emotion recognition is receiving more and more attention as the core technology for understanding and responding to the user’s inner state in human-computer interaction system. Recently, deep learning has been widely used in EEG emotion recognition. However, there are still great challenges in accurately modeling their spatial, temporal and frequency characteristics. To solve this problem, we propose a emotion recognition model CBSAtt that fuses channel-independent convolution, BiLSTM and MHSA. Firstly, short time fourier transform (STFT) is used to convert the original EEG into a time-frequency matrix to effectively reveal the dynamic changes, and then an independent convolution module is constructed for each channel, so as to extract the local spatial features of the channel. Subsequently, BiLSTM is used to model the time series features to capture dynamic information of emotion evolution over time. Finally, MHSA was introduced to enhance the perception of key features of CBSAtt and further improve the discrimination ability. Experimental results show that CBSAtt achieves superior performance on the EEG emotion recognition dataset, showing strong feature learning ability and good generalization performance.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

CBSAtt: a CNN-BiLSTM network with multi-head self-attention for EEG emotion recognition

  • 期刊论文

  • 发表日期：

    2025年9月22日

查看全文
  • Yi Zhou,
  • Ruiwen Jiang,
  • Zilong Zhou,
  • Yue Yu,
  • Jingxiang Zhang

摘要

Emotion recognition is receiving more and more attention as the core technology for understanding and responding to the user’s inner state in human-computer interaction system. Recently, deep learning has been widely used in EEG emotion recognition. However, there are still great challenges in accurately modeling their spatial, temporal and frequency characteristics. To solve this problem, we propose a emotion recognition model CBSAtt that fuses channel-independent convolution, BiLSTM and MHSA. Firstly, short time fourier transform (STFT) is used to convert the original EEG into a time-frequency matrix to effectively reveal the dynamic changes, and then an independent convolution module is constructed for each channel, so as to extract the local spatial features of the channel. Subsequently, BiLSTM is used to model the time series features to capture dynamic information of emotion evolution over time. Finally, MHSA was introduced to enhance the perception of key features of CBSAtt and further improve the discrimination ability. Experimental results show that CBSAtt achieves superior performance on the EEG emotion recognition dataset, showing strong feature learning ability and good generalization performance.