摘要
基于脑电信号的智能情绪识别系统具有便携性、高时间分辨率、实时性等特点,能够在健康、娱乐、教育等多个领域实现情绪监控与调节的应用。但由于脑电信号的非平稳性和个体差异性,传统分类器难以深入提取脑电信号中潜在的与情绪语义相关的特征。为了有效地提取脑电特征,提高脑电-情绪识别的准确性,提出一种新型的基于深浅特征融合的深度卷积残差网络情绪识别模型,主要包括浅层-深层特征提取两个模块和分类模块。首先,通过设计多层不同卷积核的卷积层,以实现浅层时-空特征提取;其次,将所提取的浅层时-空特征输入到双向GRU网络和注意力机制网络,进一步提取得到浅层-深层融合特征;最后,将浅层-深层融合特征输入到全连接层进行分类。使用DEAP数据集中76 800个脑电样本进行基于被试独立的留一交叉验证,在效价和唤醒度的维度上,跨个体、跨试次、跨时间的二分类准确率分别为96.95%和97.22%,比现有同类模型的最优识别性能分别提升3.53%和4.25%。另外,模型的性能也在MAHNOB-HCI和SEED数据集上得到验证。结果表明,提出的模型能有效地提取与情绪语义相关的脑电特征。
Electroencephalography( EEG) has advantages of portability,high temporal resolution and real-time operation,therefore has been used to recognize, monitor, and track human’ s emotion in the fields of healthcare,entertainment,education and so on. However,due to the non-stationarity and individual differences in EEG signals,it is difficult to effectively and efficiently extract informative and useful emotion related characteristics using traditional methods. To obtain representative features in an efficient manner and improve emotion classification accuracy,we proposed a feature fusion based deep residual networks using deep and shallow learning for EEG-based emotion recognition. The proposed model consisted of three modules: a shallow feature extraction module,a deep feature extraction module,and a classification module. First,the shallow feature extraction module was designed with multiple convolution layers to extract shallow tempo-spatial features.Second,the deep feature extraction module employed a Bi-GRU layer and the attention mechanism to extract deep tempo-spatial features from the extracted shallow features. Third,the classification module was designed with a fully connected layer for binary classification. The proposed model used a subject-based leave-one-out cross-validation on the DEAP database with 76 800 samples,and achieved a good performance in emotion recognition with the binary classification accuracy of 96. 95% for valence and 97. 22% for arousal. Comparing to the existing methods,our proposed model increased the accuracies by 3. 53% and 4. 25% for valence and arousal,respectively. Further,the good performance of the proposed model in binary emotion classification was also validated in MAHNOB-HCI and SEED databases as well.
作者
周如双
赵慧琳
林玮玥
胡婉柔
张力
黄淦
李琳玲
张治国
梁臻
Zhou Rushuang;Zhao Huilin;Lin Weiyue;Hu Wanrou;Zhang Li;Huang Gan;Li Linling;Zhang Zhiguo;Liang Zhen(School of Biomedical Engineering,Health Science Center,Shenzhen University,Shenzhen 518071,Guangdong,China;National Regional Key Technology Engineering Laboratory for Medical Ultrasound,Shenzhen 518071,Guangdong,China;Guangdong Provincial Key Laboratory of Biomedical Measurements and Ultrasound Imaging,Shenzhen 518071,Guangdong,China)
出处
《中国生物医学工程学报》
CAS
CSCD
北大核心
2021年第6期641-652,共12页
Chinese Journal of Biomedical Engineering
基金
国家自然科学基金(61906122,91859122)
腾讯“犀牛鸟”-深圳大学青年教师科研基金项目。
关键词
脑电信号
情绪识别
深度卷积残差网络
深浅特征融合
双向门控循环单元
electroencephalography(EEG)
emotion recognition
deep residual networks
feature extraction
bidirectional gated recurrent unit
