基于贝叶斯优化WGAN-GP的fNIRS数据增强与情绪识别OA
fNIRS Data Enhancement and Emotion Recognition Based on Bayesian Optimization WGAN-GP
因收集大量功能性近红外光谱(functional near-infrared spectroscopy,fNIRS)情绪数据的过程漫长且烦琐,导致数据量有限,从而影响深度学习分类模型的训练和准确性,提出一种基于贝叶斯优化梯度惩罚的 Wasserstein 生成对抗网络(Bayesian optimization with gradient penalty for Wasserstein generative adversarial network,BO-WGAN-GP)进行数据增强的方法.在不同分类模型上,对原始数据与生成数据混合后的数据进行大量情绪分类实验,并与其他生成对抗网络模型进行比较.实验结果表明,BO-WGAN-GP 模型生成的数据在 fNIRS 情绪识别方面表现最佳,氧合血红蛋白(oxyhemoglobin,HbO2)和脱氧血红蛋白(deoxyhemoglobin,HbR)指标的平均分类准确率分别达到了97.92%和 99.31%.
Collecting large amounts of functional near-infrared spectroscopy(fNIRS)emotion data is a lengthy and tedious process.Limited data can affect training and accuracy of deep learning classification models.To address is issue of a method using Bayesian optimization with gradient penalty for Wasserstein generative adversarial networks(BO-WGAN-GP)was proposed for data augmentation.Extensive emotion classification experiments were conducted on data mixed from original and generated data using different classification models,and comparisons were made with other generative adversarial networks.The experi-mental results showed that data generated by the BO-WGAN-GP model performed best in fNIRS emotion recognition.The average classification accuracies for oxyhemoglobin(HbO2)and deoxyhemoglobin(HbR)reached 97.92%and 99.31%respectively.
李修军;葛雄心;杨菁菁
长春理工大学 计算机科学技术学院 吉林 长春 130022长春理工大学 计算机科学技术学院 吉林 长春 130022||长春理工大学中山研究院 广东 中山 528437长春理工大学 计算机科学技术学院 吉林 长春 130022
信息技术与安全科学
生成对抗网络数据增强贝叶斯优化功能性近红外光谱技术情绪识别
generative adversarial networkdata augmentationBayesian optimizationfunctional near-infrared spectroscopy technologyemotion recognition
《郑州大学学报（理学版）》 2026 (3)
17-24,8
吉林省教育厅科学技术研究项目(JJKH20220780KJ,JJKH20230847KJ)
评论