基于生成对抗网络的情绪识别数据增强方法

Data Enhancement Method of Emotion Recognition Based on GAN

使用深度学习方法构建高准确率的情绪识别模型需要大量的情绪脑电数据。生成对抗网络(GAN)最近在生成逼真的数据方面取得了巨大成功,但一直没有客观的评价指标衡量生成的数据质量,无法保证生成的样本总是有助于分类。针对此问题,提出了一种将带条件和梯度惩罚的生成对抗网络(Conditional Wasserstein GAN-Gradient Penalty,CWGAN-GP)与序列后向选择(Sequential Backword Selection,SBS)相结合的数据增强方法。利用SBS自动从CWGAN-GP生成的人工样本中选择高质量的人工样本加入到训练集中,在DEAP数据集中评估提出的CWGAN-GP-SBS方法。实验结果表明,使用CWGAN-GP-SBS方法得到样本的测试分类准确率相比传统SBS方法平均高出5.86%,说明CWGAN-GP-SBS生成的人工样本可以显著提高情绪识别模型的准确性。

Using deep learning method to construct emotion recognition model with high recognition accuracy requires a lot of emotional EEG data.Recently,generative adversarial network(GAN)has achieved great success in generating realistic data,but there has been no objective evaluation index to measure the quality of generated data to ensure that the generated samples always contribute to classification.Aiming at solving this problem,a data augmentation method is proposed which combines CWGAN-Gradient Penalty(CWGAN-GP)with sequential backword selection(SBS).High-quality artificial samples are automatically selected from the artificial samples generated by CWGAN-GP using SBS and added to the training set.The proposed CWGAN-GP-SBS method is evaluated on the DEAP data set.The experimental results show that the accuracy of the samples obtained by the CWGAN-GP-SBS method is on average 5.86%higher compared to those without SBS,indicating that the accuracy of emotion recognition model can be significantly improved by using the artificial samples generated by CWGAN-GP-SBS.