Abstract:

This paper proposes a new algorithm on MEG (magnetoencephalography) signals to classify MEG signals generated when human brain confronts the stimulation of different visual objects. At first, filtered MEG signals are projected to a new feature space and they will generate new features. The covariance features of new features, on behalf of the MEG signals features, will be used to presort samples in a vector space named tangent space. Then, the second time labeling results can be derived after adjusting prelabeled samples using Bhattacharyya distance. Finally, the ultimate classification results can be got by applying the Riemann distance method to adjust covariance matrices in manifold. The extensive experiments show that the combination of supervised and unsupervised algorithms can remarkably improve the classification accuracy of MEG signals.

Key words: magnetoencephalography (MEG), classification algorithm, covariance matrices, Riemann distance, Bhattacharyya distance

摘要：

针对人脑对不同视觉目标刺激产生的脑磁图（magnetoencephalography，MEG）信号，提出了一种新型的脑磁图信号分类算法。该算法首先将滤波后的脑磁图信号投影到新的特征空间，然后将脑磁图信号投影后新特征的协方差特征投影到切线空间中，用协方差特征作为信号的特征，进而对样本进行预分类；接着将预分类的样本通过巴氏距离的调整，得到二次标记结果；最后采用黎曼距离对协方差特征矩阵在流形上进行调整，得到最终的分类结果。实验结果表明，该有监督与无监督相结合的算法有助于提高脑磁图信号分类的准确率。

关键词: 脑磁图（MEG）, 分类算法, 协方差矩阵, 黎曼距离, 巴氏距离