基于多尺度卷积和身份子空间的脑纹识别  

Brainprint recognition based on multi-scale convolution and subspace

作　　者：刘国文 唐佳佳 金宣妤 孔万增[1] LIU Guowen;TANG Jiajia;JIN xuanyu;KONG Wanzeng(School of Computer,Hangzhou Dianzi University,Hangzhou Zhejiang 310018,China)

机构地区：[1]杭州电子科技大学计算机学院,浙江杭州310018

出　　处：《杭州电子科技大学学报(自然科学版)》2025年第1期82-88,共7页Journal of Hangzhou Dianzi University:Natural Sciences

基　　金：国家自然科学基金项目(U20B2074);浙江省重点研究与发展项目(2023C03026,2021C03001,2021C03003);浙江省脑机协同智能重点实验室项目(2020E10010)。

摘　　要：脑电图信号具有难窃取性、抗伪造性和活体性等独特优势,可以为身份识别提供更安全的生物识别方法。传统的脑纹识别仅针对单一任务状态下的脑电图信号,身份识别的可靠性取决于特定刺激和诱发的脑电特征。在现实环境中,人类接收到的刺激是多种多样且瞬息万变的,诱发的脑电信号是复杂多变的,因此极大降低了这些研究结果在现实环境中的可靠性。卷积神经网络可以学习特征区域内的变化,且不同尺度的卷积核可以学习不同粒度的信息,因此针对任务无关的脑纹识别提出了一种基于多尺度卷积的方法。首先,对数据进行降采样和带通滤波,提取出具备时域、频域和空间特性的功率谱密度特征。然后,在各个通道的时间、频率维度使用多尺度卷积融合,再对具有时频变化的深度特征间使用多尺度卷积学习不同的脑域特征。最后,将高维空间的统计特征映射到子空间,去除冗余信息并保留个体独有特征,用映射到低维空间的本征脑电均值构成身份本征,通过计算样本与身份本征信号的余弦距离来判断受试者的身份信息。在具有多个任务的数据集上进行了实验,以评估所提出方法在不同任务中的泛化能力。提出的方法在印度理工任务无关数据集上,识别准确率和等错误率分别为87.8%和4.9%。Electroencephalography(EEG)signals have the unique advantages of being difficult to steal,resistant to forgery and in vivo,which can provide a more secure biometric method for identity recognition.Traditional brainprint recognition only targets EEG signals in a single task state,and the reliability of identification depends on the specific stimuli and evoked EEG features.In real-world environments,the stimuli received by humans are diverse and rapidly changing,and the evoked EEG signals are complex and variable,thus greatly reducing the reliability of these findings in real-world environments.A multi-scale convolution-based approach is proposed for task-independent brainprint recognition.Firstly,the data is downsampled and band-pass filtered to extract Power Spectral Density(PSD)features with time-domain,frequency-domain,and spatial characteristics.Then,multi-scale convolution is used to fuse both the temporal and the frequency dimensions of each channel,and then multi-scale convolution is applied to the time-frequency variation.Finally,the statistical features in the high-dimensional space are mapped to the subspace to remove the redundant information and retain the unique features of individuals,and the eigen EEG mapped to the low-dimensional space is used to form the identity eigen.Experiments are conducted on a dataset with multiple tasks to evaluate the generalization ability of the proposed method across different tasks.The proposed method achieves 87.8%recognition accuracy and 4.9%equal error rate on the IIT task-independent dataset,respectively.

关 键 词：脑纹识别 任务无关 多尺度卷积 子空间 

分 类 号：R318[医药卫生—生物医学工程]