文章摘要

文章信息/Info

Title:: Research on emotion recognition based on image features of ECG signal and convolutional neural network

Author(s):: LI Yongkang; FANG Ancheng; CHEN Yanang; XIE Ziqi; PAN Fan; HE Peiyu; （College of Electronics and Information Engineering, Sichuan University, Chengdu 610065,China）

Keywords:: Emotion recognition; Electrocardiography; Feature extraction; Dual input; Convolutional neural network

摘要:: 为提高情绪识别的准确率，本研究利用卷积神经网络和迁移学习，提出了一种基于心电（electrocardiography,ECG)信号图像特征的情绪识别方法。首先对ECG信号进行预处理，去除噪声；然后提取ECG信号的时域波形图和时频图；最后，利用迁移学习和双输入EfficientNetV2网络学习图像的时域和频域特征并进行分类，得到对应的情绪类别。在公开数据集Amigos上进行验证，结果显示，本研究在唤醒度、效价和优势度的识别准确率分别为91.63%，95.27%和92.32%。相较于其它情绪识别方法，本研究方法具有更高的准确率。

Abstract:: In order to improve the accuracy of emotion recognition, we used convolutional neural network and transfer learning method to propose an emotion recognition method based on electrocardiography(ECG) signal image features. First, the ECG signal was preprocessed to remove noise, and then the time-domain waveform and time-frequency graph of the ECG signal were extracted. Finally, transfer learning and the time-domain and frequency-domain features contained in the dual input EfficientNetV2 network learning images were used and classified to obtain the corresponding emotion categories. The results of validation on the public dataset Amigos showed that the recognition accuracy of arousal, titer and dominance were 91.63%, 95.27% and 92.32%，respectively. Compared to other emotion recognition methods, this method has higher accuracy.