希尔伯特黄变换方法及其在特征提取中的应用研究

发布时间：2018-01-07 16:22

本文关键词：希尔伯特黄变换方法及其在特征提取中的应用研究　出处：《北京科技大学》2017年博士论文　论文类型：学位论文

【摘要】：希尔伯特黄变换(Hilbert Huang transform简记HHT)具有数据驱动性,它是依据信号自身特性设定特征时间尺度,可将信号分解成由高频到低频的内膜函数(intrinsic mode functions,简记IMFs),更能反映非平稳信号的局部特征,从而更能准确的提取非平稳信号的特征。中药色谱图谱识别问题是中药质量控制、中药真假鉴别、指导中药材栽培等的重要依据。同一种中药因产地不同、野生和栽培不同、栽培方法不同等,其色谱图存在一定差异,但差异性微小,相似性极高,对识别分类带来了困难,这是中药质量控制、中药真假鉴别、指导中药材栽培中的难点问题。这一问题的解决,关键是对中药色谱图谱特征提取的研究,而中药色谱图谱是一系列不同频率的高斯函数构成的一维信号,其特征提取问题应是一类非平稳信号特征提取问题。光照人脸图像识别问题是人脸识别难点问题,研究其特征提取方法仍是现在人脸识别研究的热点,光照人脸图像是二维信号,不同频率成分在图像上表现不均匀,增加了识别的难度。这两类模式识别问题虽然研究的模式背景不同,模式表达形式不同,但是它们都具有非平稳信号特点。希尔伯特黄变换恰恰针对非线性非平稳信号处理,具有独特的优势,本论文尝试应用希尔伯特黄变换方法解决中药色谱图谱和光照人脸图像特征提取问题,并在中药甘草色谱图谱识别和光照人脸图像识别中加以验证。本文主要研究工作与创新点如下:1)研究应用希尔伯特黄变换(HHT),提取在不同栽培条件下的中药甘草色谱图谱的特征。针对在不同栽培条件下的中药甘草色谱图谱,提出了希尔伯特黄变换的经验模式分解(EMD)与分形维数相结合的方法,应用于中药甘草色谱图谱识别,提取不同栽培条件下的中药甘草色谱图谱的特征,即EMD分形特征,并与小波分形特征相比较,识别结果表明EMD分形特征分辨能力好于小波分形特征。在此基础上,为进一步提高分类识别效果,有效提取甘草中药特征,设计了一种分割窗EMD分形特征提取算法,并应用于不同栽培条件下的甘草色谱图谱的分类。从实验验证结果可以明显看出,分割窗EMD分形特征好于单纯使用EMD和EMD分形特征,而且随着训练样本集中样本数和样本分解层数的增加,分类识别率表现非常稳定。2)研究利用希尔伯特黄变换(HHT),提取光照人脸的高频特征。针对同一个人在不同光照条件下的人脸识别问题,本文提出了基于EMD高频IMF特征提取方法和高频IMF的人脸融合特征提取方法。根据EMD的自适应性和数据驱动特性,可以将信号分解成由高频到低频的若干个IMFs,依据高频成分对光照表现较稳定的特点,提出用第一个IMF作为光照人脸识别特征,在PIE人脸库实验结果显示用此特征识别效果好于用db4小波变换的高频特征识别;在此基础上,进一步提出了高频IMF的人脸融合特征提取方法,光照人脸经过融合,基本消除了不同方向光源在人脸图像上的影响,在PIE人脸库上实验结果显示,融合后识别率比融合前识别率提高了近30%。3)研究用数学分析方法改进希尔伯特黄变换的数值计算方法,提取光照人脸的相位特征。由于希尔伯特黄变换的EMD分解过程,上下包络的获取是采用了三次样条插值拟合法,缺乏数学原理分析,本文提出了一种改进的二维移动平均滤波方法,替代三次样条插值拟合得到的上下包络均值,经过筛分过程(BEMD),获得二维IMFs(BIMF),并将其应用于光照人脸相位特征提取,根据相位一致性原理,对每一个BIMF经Riesz变换,获得每个IMF的单演信号,计算相位一致函数值(Phase Congruence 简记PC),从而得到光照人脸图像的相位特征,经 PIE(Pose,Illumination,and Expression(PIE)face database ofthe Carnegie Mellon University)人脸库识别验证,效果好于改进前的传统EMD方法。本文研究成果不仅对中药色谱图谱和光照人脸识别特征提取方法研究具有一定指导意义,而且对一类非平稳信号特征提取问题以及希尔伯特黄变换方法本身理论和应用研究都具有一定的参考价值。
[Abstract]:Hilbert Huang transform (Hilbert Huang transform abbreviated HHT) is a data driven, it is to set up the characteristic time scale according to the signal characteristics, the signals can be decomposed into low-frequency function into the intima by high frequency (intrinsic mode functions, abbreviated IMFs), to better reflect the local characteristics of non stationary signals, which can more accurately extract features non stationary signal. Chromatographic pattern recognition problem is the quality control of traditional Chinese medicine, Chinese medicine identification, an important basis for the cultivation of medicinal materials in the guide. The same kind of traditional Chinese medicine from different habitats, wild and cultivated different, different cultivation methods, there are some differences between the chromatograms, but the difference is small, high similarity, brought it is difficult to identify the classification and quality control of traditional Chinese medicine, Chinese medicine identification, and difficult problems in cultivation of medicinal materials in the guidance. To solve this problem, the key is to extract the feature of chromatographic From the study, and the traditional Chinese medicine chromatographic fingerprinting is a series of one-dimensional signals of different frequencies of the Gauss function, the feature extraction problem is the extraction problem for a class of nonstationary signal feature. Illumination face image recognition is the face recognition problem, study the method of feature extraction is still a research hotspot now face recognition, illumination face the image is a two-dimensional signal of different frequency components is not uniform in the image, increase the recognition difficulty. These two kinds of pattern recognition problems while different mode of background research, expression patterns of different forms, but they all have non-stationary characteristics. Hilbert Huang transform just to deal with nonlinear and nonstationary signal, has unique advantages. This thesis attempts to apply Hilbert Huang transform method of chromatographic and illumination face image feature extraction problem, and in the chromatographic fingerprints of licorice The identification and illumination face image recognition is verified. The main research work and innovations are as follows: 1) study on the application of Hilbert Huang transform (HHT) feature extraction, chromatographic fingerprints of licorice in different cultivation conditions. The chromatographic fingerprints of licorice in different cultivation conditions, put forward the empirical mode decomposition of Hilbert Huang transform (EMD) method combined with fractal dimension, used in traditional Chinese medicine licorice chromatographic identification, feature extraction of licorice Chromatognun under different cultivation condition, namely EMD fractal characteristics, and the fractal features and wavelet phase comparison, identification results show that the EMD fractal feature resolution better than wavelet based fractal characteristics., in order to further improve the classification effect of traditional Chinese medicine, licorice extract features effectively, design a EMD algorithm to extract the fractal feature segmentation window, and applied in different cultivation conditions The classification of licorice chromatogram. From the experimental results it is clear that the split window EMD fractal feature is better than the simple use of EMD and EMD fractal characteristics, and with the increase of the training sample set the sample number and sample decomposition, classification rate performance is very stable.2) study using Hilbert Huang transform (HHT), extraction of light high frequency facial features. Aiming at the problem of face recognition in the same person under different illumination conditions, this paper put forward a method to extract facial feature extraction method EMD high frequency IMF and high frequency IMF fusion based on the characteristics of driving adaptability and. According to the characteristic of EMD data, it can decompose the signal into several IMFs from high frequency to low frequency. According to the characteristics of high frequency light performance is stable, the first IMF as a light face recognition feature in PIE face database, experimental results show that the recognition effect is better than DB High frequency feature recognition 4 wavelet transform; on this basis, further puts forward the extraction method of high frequency IMF face feature fusion, illumination face after fusion, the effects of different light source in the direction of face image basically eliminated, show the experimental results on PIE face database, the fusion recognition rate before the fusion recognition rate increased 30%.3) numerical calculation method of mathematical analysis method improved Hilbert Huang transform, extracting illumination phase facial features. Because of the decomposition process of Hilbert Huang transform EMD, obtain the envelope is the use of the three spline interpolation fitting, lack of mathematical analysis principle, this paper proposes an improved two-dimensional moving average filtering method, substitute three times by the spline interpolation on the mean envelope, through the screening process (BEMD), IMFs (BIMF), obtained and applied to illumination face feature extraction phase According to the principle, the consistency of phase, for each BIMF by Riesz transform, obtain the monogenic signal of each IMF, calculate the phase congruency function value (Phase Congruence or PC), so as to obtain the light phase features of face image, by PIE (Pose, Illumination, and Expression (PIE) face database ofthe Carnegie Mellon University) face database identification, the effect is better than the traditional EMD method before improvement. This paper research not only on chromatographic profiles and illumination face recognition feature extraction method has a certain guiding significance, but also has a certain reference value and the extraction problem of Hilbert Huang transform method theory and Application Research of a class of nonstationary signal feature.

【学位授予单位】：北京科技大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】