高级检索

    宋振振, 陈兰岚, 娄晓光. 基于时序卷积网络的情感识别算法[J]. 华东理工大学学报(自然科学版), 2020, 46(4): 564-572. DOI: 10.14135/j.cnki.1006-3080.20190508001
    引用本文: 宋振振, 陈兰岚, 娄晓光. 基于时序卷积网络的情感识别算法[J]. 华东理工大学学报(自然科学版), 2020, 46(4): 564-572. DOI: 10.14135/j.cnki.1006-3080.20190508001
    SONG Zhenzhen, CHEN Lanlan, LOU Xiaoguang. Emotion Recognition Algorithm Based on Temporal Convolution Network[J]. Journal of East China University of Science and Technology, 2020, 46(4): 564-572. DOI: 10.14135/j.cnki.1006-3080.20190508001
    Citation: SONG Zhenzhen, CHEN Lanlan, LOU Xiaoguang. Emotion Recognition Algorithm Based on Temporal Convolution Network[J]. Journal of East China University of Science and Technology, 2020, 46(4): 564-572. DOI: 10.14135/j.cnki.1006-3080.20190508001

    基于时序卷积网络的情感识别算法

    Emotion Recognition Algorithm Based on Temporal Convolution Network

    • 摘要: 采用脑电数据集DEAP进行情感识别。由于脑电信号具有时序性,采用深度学习中的时序卷积网络(TCN)对数据进行训练识别。首先使用小波包分解提取各子带小波系数能量值作为特征;然后通过TCN对特征进行训练,在训练过程中加入了Snapshot寻优思想保存多个模型;最后采用投票集成策略建立集成模型,以提高识别精度,并增强结果稳健性。实验结果表明,本文方法将情感分为二类和四类的平均识别精度分别能够达到95%和93%,相对于同类研究有较大的提高。

       

      Abstract: This paper uses EEG dataset DEAP to undergo the emotion recognition. By considering the time sequence characteristics of EEG signals, we adopt temporal convolution network (TCN) as a deep learning approach to train and identify EEG data. TCN model can not only grasp the characteristics of time series, but also retain the parallel computing property of CNN. Hence, it can improve the computational efficiency, while maintaining high identification accuracy. Firstly, wavelet packet transform is employed to decompose the raw signals into sub-frequency bands, and then, wavelet coefficient energy values are computed as the features. Wavelet packet analysis is an extension of wavelet analysis, whose basic idea is to concentrate the information energy, find the order in the details, and filter out the rules. Hence, it can provide a more sophisticated analysis method for the signal. Besides, the extracted features are trained through the TCN and Snapshot optimization idea is adopted for saving multiple models during the process of training. The principle of Snapshot optimization is to avoid local pole values by constantly resetting the learning rate. Finally, the ensemble learning with the voting integration strategy is implemented to combine multiple machine learning models into an integrated model such that the recognition accuracy can be improved and the robustness of the results can be enhanced. Finally, it is shown via the experimental results that the proposed method can realize the two classification and four classification identification tasks of emotions with the average recognition accuracy, 95% and 93%, respectively. Compared with similar methods, this proposed method attains significant improvement.

       

    /

    返回文章
    返回