基于ECG生物识别并行化的研究与实现

发布时间：2018-04-26 22:30

本文选题：大数据 + 基于基准点特征　；参考：《华中科技大学》2016年硕士论文

【摘要】：在过去十几年中,生物识别技术已经相当成熟了,它是一门利用统计学方法和人体生理活动数据来验证个人身份的技术。心电信号ECG(Electrocardiograph)本身因人而异的,并且在每个人当中不可复制,目前大量的ECG信号用于生物识别技术,最新研究指出了ECG生物识别技术的一个待解决问题:在应用场景多样化和人群数量庞大的情况下,如何充分的利用ECG各类特征来保持识别鲁棒性的问题。针对上述问题,本研究首先利用传统的特征提取方法,将基于特征点特征(Fiducial based features)和基于非特征点特征(Non-Fiducial based features)进行结合,提取一种结合Fiducial和Non-fidicuial的混合特征提取方法,以完成ECG信号多维特征的建模。其中,Fiducial特征包括ECG信号的波幅特征、ECG信号的时序特征和ECG信号的频谱特征;Non-Fiducial特征包括ECG心电图的波形形状。进而将两类特征混合并进行统一建模,经验证,在多样化场景中本研究提出的ECG混合特征比传统的ECG单维度特征拥有更高的识别率。第二步,针对人群数据庞大时,数据训练的时间开销大的问题,本研究基于上述ECG混合特征,提出新的LDA的算法(LDA Based On Multiple Features,LOMF),LOMF算法包含了ECG信号的预处理、子块划分和分块训练。并利用MapReduce分布式计算框架进行算法并行化,提出一种基于多维特征空间的二级检索方式,在保证计算效率提高的同时,将识别率提升到一个更高的等级。文中实验部分将ECG混合特征分别与Fiducial,Non-Fiducial两种单维特征方法进行对比,发现在同一种识别算法中,ECG混合特征有更高的识别率。并且本文提出的基于多维特征空间二次检索的LOMF算法比传统的LDA,SVM等算法精度有7%-8%的提升,且LOMF最大的优势在于很好的契合于MapReduce并行框架,更适于互联网这种数据集增长速度快的应用场景。
[Abstract]:In the past decade, biometric technology has been developed, it is a use of statistical methods and human physiological activity data to verify the identity of individuals. ECG electrocardiography (ECG) itself varies from person to person and cannot be duplicated in everyone. A large number of ECG signals are currently used in biometrics. The latest research has pointed out an unsolved problem of ECG biometrics: how to make full use of the ECG features to maintain the robustness of recognition under the circumstances of diverse application scenarios and large number of people. In order to solve the above problems, this paper firstly uses the traditional feature extraction method to combine the feature point based feature based (feature) and non-Fiducial based feature (Non-Fiducial based feature) to extract a hybrid feature extraction method which combines Fiducial and Non-fidicuial. In order to complete the modeling of multidimensional features of ECG signal. Fiducial features include the amplitude characteristics of ECG signals and the frequency spectrum features of ECG signals, including the waveform shapes of ECG electrocardiograms. Then the two kinds of features are mixed and unified modeling. It is proved that the ECG hybrid feature proposed in this paper has a higher recognition rate than the traditional ECG one-dimensional feature in the diverse scenarios. The second step is to solve the problem that the time cost of data training is large when the crowd data is large. Based on the mixed characteristics of ECG mentioned above, a new LDA algorithm named LDA Based On Multiple Features-LOMF algorithm is proposed, which includes the preprocessing of ECG signal, subblock partition and block training. The algorithm is parallelized by using MapReduce distributed computing framework, and a two-level retrieval method based on multidimensional feature space is proposed, which can improve the efficiency of computation and raise the recognition rate to a higher level. In the experiment part, we compare the mixed features of ECG with those of Fiducial Non-Fiducial and find that the mixed feature of ECG has higher recognition rate in the same recognition algorithm. Moreover, the proposed LOMF algorithm based on multidimensional feature space quadratic retrieval has 7- 8% higher accuracy than the traditional LDA-SVM algorithm, and the biggest advantage of LOMF is that it fits well with the MapReduce parallel framework. More suitable for the Internet, such as the rapid growth of data sets application scenarios.
【学位授予单位】：华中科技大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【参考文献】