基于半监督的脑磁图视觉目标分类算法

发布时间：2018-01-25 09:32

本文关键词： 脑磁图去噪降维分类巴氏距离黎曼距离　出处：《天津大学》2016年硕士论文　论文类型：学位论文

【摘要】：人脑活动会产生电磁信号,可以通过设备对产生的电磁信号进行捕捉,捕捉的信号为脑磁图。研究脑磁信号,对人类攻克各种疾病和对人类各种高级行为比如语言、动作、思维的研究都具有深远的意义。脑磁信号在采集与传输的过程中,不可避免会受到大量的噪声干扰。同时脑磁图多通道多采样时间点构成的高维数据特性,给脑磁图的研究带来了难题。由于大量的噪声干扰,根据脑磁图的特性,需要设计合理的滤波器对脑磁图进行去噪。脑磁图的高维数据需要采用有效的降维方式进行维度的降维。同时维度降低后的特征数据,需要采用有效的方式对数据进行分类。本文结合脑磁图的特性,分析了各种滤波方法特性的基础上,采用组合滤波器对脑磁图信号去噪。在对比了线性降维和非线性降维方式的基础上,采用协方差式的降维对脑磁图信号进行降维并提取特征。在降维的基础上,本文采用SVM和Lasso模型对数据特征进行分类并做了对比,提出了组合分类器来提高模型的泛化特性。在有监督的组合分类器基础上,本文加入无监督的巴氏距离和黎曼距离对有监督的分类结果进行无监督的聚类,提升分类的精度。本文的创新性在于将有监督的算法与无监督的算法相结合,采用半监督的方式对数据进行分类。通过实验对设计的算法进行验证,在两个实验数据集上都取得了成功,两个实验数据集交叉验证的平均正确率分别为80.11%和99.2%,高于原先两个数据集比赛中最好的成绩73.6%的正确率和97.1%的正确率。
[Abstract]:Human brain activity produces electromagnetic signals, which can be captured by devices, and the signals captured are magnetograms. It is of great significance to study various diseases and advanced behaviors such as language, action and thinking. Magnetic signals are collected and transmitted in the process of brain magnetic signal acquisition and transmission. At the same time, the high dimensional data characteristics of multichannel and multi-sampling time points of the brain magnetic map bring difficulties to the study of the brain magnetic map, due to a large number of noise interference. According to the characteristics of magnetoencephalogram, we need to design a reasonable filter to de-noising the magnetoencephalogram. The high-dimensional data of the magnetoencephalogram need to be dimensionally reduced in an effective way. At the same time, the dimensionally reduced feature data should be used. It is necessary to classify the data in an effective way. In this paper, the characteristics of various filtering methods are analyzed based on the characteristics of the brain magnetic map. Based on the comparison of linear and nonlinear dimensionality reduction methods, covariance dimension reduction is used to reduce the dimension of the signal and extract the features. On the basis of dimensionality reduction, the signal is de-noised by a combined filter. In this paper, SVM and Lasso models are used to classify and compare the data features, and a combined classifier is proposed to improve the generalization of the model. This paper adds unsupervised Pasteur distance and Riemann distance to unsupervised clustering of supervised classification results to improve the accuracy of classification. The innovation of this paper is to combine supervised algorithm with unsupervised algorithm. The semi-supervised method is used to classify the data, and the designed algorithm is verified by experiments, which is successful in both experimental data sets. The average accuracy of cross-validation of the two experimental data sets is 80.11% and 99.2, respectively, which is higher than that of 73.6% and 97.1% of the best scores in the original two datasets.
【学位授予单位】：天津大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：R318;TP391.41

【相似文献】