摘要
在神经美学研究中已经证明,中文字体审美偏好的情绪刺激可以通过观察3种偏好(喜欢、不喜欢和中性)之间的事件相关电位(event related potential,ERP)波动获得.本文通过引入一种核化张量奇异值分解的多视角聚类方法分别构建了基于脑电图(electroencephalogram,EEG)和ERP的审美偏好识别模型,通过这些模型首次确认了该结论.本文方法将来自不同频段的数据视为描述中文字体审美偏好的不同视角,通过张量多秩最小化的约束探索所有视角特征的一致性和关联性,并通过之后的聚类获取审美偏好的识别结果.采用多视角无监督聚类方法得到的识别精度达到97.1%.此外,通过输入–扰动关联方法将电极的振幅与不同种类的审美偏好相关联,可视化关键频段组合以及电极之间的关系,分别取出与喜欢、不喜欢、中性最相关的3个电极,包含次相关的6个电极,包含第三相关的9个电极,包含第四相关的12个电极,分别形成4种不同组合的脑电特征.通过比较实验,验证了相对于62个电极信号,上述4种组合方式在字体美学分类上更具有优势,并且最相关的3个电极的组合特征对审美偏好最具判别性.实验结果表明,基于多视角聚类的方法能够解决神经信号与审美偏好的相关分析,并能挖掘出与字体审美偏好最相关的电极.
Previous neuroesthetic studies have proved that Chinese typefaces can be viewed as an esthetic preference stimulus by observing differences in event related potential(ERP)waves among three preferences,namely,like,dislike,and neutral.We first reconfirm this conclusion by introducing a multiview clustering method of kernelized tensor singular value decomposition(KT-SVD)to construct an esthetic preference recognition model based on electroencephalograms(EEGs).Our approach regards data from different frequency bands as different views describing the esthetic preferences of Chinese fonts,explore the relevance of all view features through the constraint of tensor multi-rank minimization,and obtains the esthetic preferences using the clustering results.Additionally,the input-perturbation correlation method is used to correlate the amplitude of the electrodes with different types of esthetic preferences and describe the relationship between the key frequency-band combinations and electrodes,and take out the electrodes most relevant to likes,dislikes,and neutrality,including 3 electrodes of Top-1,6 electrodes of Top-2,9 electrodes of Top-3,and 12 electrodes of Top-4,forming four different combinations of EEG features for esthetic preference recognition experiments.Experimental results show that the method based on multiview clustering can solve the correlation analysis of neural signals and esthetic preferences and mine the electrodes most relevant to the esthetic preferences of fonts.
作者
张艳
谢源
洪辰
曲延云
李睿
张俊松
李翠华
Yan ZHANG;Yuan XIE;Chen HONG;Yanyun QU;Rui LI;Junsong ZHANG;Cuihua LI(Computer Seience Department,School of Informatics,Xiamen Universitg,Xiamen 361005,China;School of Computer Science and Technology,East China Normal University,Shanghai 200062,China;National Engimeering Laboratory for Technology of Big Data Applications in Education,Central China NormalUniversitg,Wuhan 430079,China;Mind,Art&Computation Group,Department of Artificial Intelligence,School of Informatics,Xiamen Universityg,Xiamen 361005,China;School of Mathematics Science,Guizhou Normal University,Guigang 550001,China)
出处
《中国科学:信息科学》
CSCD
北大核心
2021年第3期383-398,共16页
Scientia Sinica(Informationis)
基金
国家自然科学基金面上项目(批准号:61876161,61772524,61671397,U1065252,61772440)资助。
关键词
中文字体
审美评价
计算美学
事件相关电位
核化张量奇异值分解
数据挖掘
Chinese typeface
esthetic evaluation
computational esthetics
event-related potentials
kernelized tensor-SVD
data mining