WO2022178978A1 - 一种基于最大化比率和线性判别分析的数据降维方法 - Google Patents

一种基于最大化比率和线性判别分析的数据降维方法 Download PDF

Info

Publication number
WO2022178978A1
WO2022178978A1 PCT/CN2021/090835 CN2021090835W WO2022178978A1 WO 2022178978 A1 WO2022178978 A1 WO 2022178978A1 CN 2021090835 W CN2021090835 W CN 2021090835W WO 2022178978 A1 WO2022178978 A1 WO 2022178978A1
Authority
WO
WIPO (PCT)
Prior art keywords
matrix
data
label
discriminant analysis
linear discriminant
Prior art date
Application number
PCT/CN2021/090835
Other languages
English (en)
French (fr)
Inventor
王靖宇
王红梅
聂飞平
李学龙
Original Assignee
西北工业大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 西北工业大学 filed Critical 西北工业大学
Priority to US17/997,684 priority Critical patent/US20240029431A1/en
Publication of WO2022178978A1 publication Critical patent/WO2022178978A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/7715—Feature extraction, e.g. by transforming the feature space, e.g. multi-dimensional scaling [MDS]; Mappings, e.g. subspace methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G06V20/13—Satellite images
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
    • G06F18/24133—Distances to prototypes
    • G06F18/24143—Distances to neighbourhood prototypes, e.g. restricted Coulomb energy networks [RCEN]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing
    • G06V10/34—Smoothing or thinning of the pattern; Morphological operations; Skeletonisation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16—Human faces, e.g. facial parts, sketches or expressions
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G06V20/194—Terrestrial scenes using hyperspectral data, i.e. more or other wavelengths than RGB

Definitions

  • the invention belongs to the field of image classification and pattern recognition, and in particular relates to a data dimension reduction method based on maximizing ratio and linear discriminant analysis.
  • Data dimensionality reduction technology is an important research topic in the field of image classification and pattern recognition. In the context of big data, the amount of raw data directly obtained in practical application scenarios is huge. Handle the requirements of the hardware platform.
  • Data dimensionality reduction is the dimensionality reduction of the original high-dimensional data. While ensuring that the data after dimensionality reduction still retains most of the information contained in the original data, the dimensionality of the data is reduced as much as possible to improve the efficiency of data storage and processing and reduce the need for Requirements for hardware and subsequent data processing algorithms. Since data dimensionality reduction can reduce the data dimension and required storage space, save model training and calculation time, and improve the accuracy of subsequent applied algorithms, data dimensionality reduction technology has been widely used in face recognition, hyperspectral image classification, Chinese herbal medicine classification, etc. In the early data processing of practical application scenarios.
  • the current data dimensionality reduction methods include feature selection and feature extraction.
  • the feature selection method is to directly select key features from the original high-dimensional data, and the feature extraction is to project the existing features into a new space to form new features.
  • the former is beneficial to preserve the physical meaning of the data, while the new features obtained by the latter are difficult to explain, but the effect of feature selection is slightly inferior to that of feature extraction.
  • Linear discriminant analysis is a common method for feature extraction, which can well preserve the discriminant information of the data, and is often used in the preprocessing step of data classification. However, the traditional linear discriminant analysis can not be higher than or equal to the total number of categories of the data after dimensionality reduction, and it is easy to be unsolvable due to the non-singularity of the matrix during the solution process.
  • the original data features of image classification are often high-dimensional, and too high a dimension may reduce the classification accuracy, and the original data contains redundant information. Using it directly for classification will lead to problems such as slow data processing and low classification accuracy. .
  • hyperspectral imaging has been widely used in the classification of ground objects. How to reduce the dimensionality of high-dimensional hyperspectral data, thereby reducing the cost of data storage and processing, and extracting key features and category information of the data is of great importance. practical meaning.
  • the traditional linear discriminant analysis method tends to select features with small variance, which are difficult to effectively distinguish between categories, and the traditional linear discriminant analysis needs to invert the intra-class covariance matrix in the solution process, but in many cases the matrix is singular , at this time, the method fails, and the data dimension reduction of the highlight image cannot be performed.
  • the traditional linear discriminant analysis must ensure the singularity of the intra-class covariance matrix through preprocessing, in order to achieve dimensionality reduction of high-dimensional data, which leads to complex data processing flow and interaction between the preprocessing algorithm and the data dimensionality reduction algorithm. question.
  • the present invention proposes a method based on maximizing ratio and linear discriminant
  • the dimensionality reduction method of the analyzed data due to the problems of low efficiency and low accuracy in the image classification method due to the imperfection of the dimensionality reduction method, the present invention proposes a ground object classification method for hyperspectral images.
  • a data dimensionality reduction method based on maximizing ratio and linear discriminant analysis characterized in that the steps are as follows:
  • Step 1 construct a data matrix, a label vector and a label matrix according to the image;
  • the image is a hyperspectral image, a Chinese herbal medicine image or a face image;
  • Step 2 Calculate the intra-class covariance matrix and the inter-class covariance matrix
  • Step 3 Construct an optimization problem based on linear discriminant analysis that maximizes the sum of ratios
  • Step 4 Solve the projection matrix that maximizes the objective function.
  • Step 2 is as follows:
  • X and G are the sample matrix and label matrix obtained in step 1, respectively, is a unit matrix of order n, is an n-dimensional all 1-column vector.
  • step 4 an alternate iterative optimization algorithm is used to solve the projection matrix
  • a method for classifying features of hyperspectral images by adopting the above-mentioned dimensionality reduction method wherein the samples in step 1 are hyperspectral images, and the eigenvalues are grayscale values of a single waveband; n is a single waveband The total number of pixels, c is the total number of pixel ground object categories; perform steps 1-4 in turn to obtain the projection matrix; use the projection matrix to project the data matrix formed by the grayscales of the corresponding pixels in all bands in the acquired hyperspectral image of the unknown label, to obtain Projected sample matrix Z; take each column of Z as all feature sequences of pixels corresponding to the new unknown label, and classify the projected new pixel samples using the K-nearest neighbor classifier that has been trained with the training samples , and finally get the category label of the pixel corresponding to the unknown label feature.
  • k 3 of the K-nearest neighbor classifier.
  • a data dimensionality reduction method based on maximizing ratio and linear discriminant analysis proposed by the present invention establishes the objective function of the linear discriminant analysis method based on maximizing ratio sum, avoiding the tendency of traditional linear discriminant analysis to select small variance and discriminating ability. For the problem of weak features, it is possible to select features that are more conducive to classification.
  • Alternate optimization iterative algorithm is used to solve the optimization problem of linear discriminant analysis of maximizing the ratio sum.
  • This algorithm does not rely on the calculation of the inverse matrix of the intra-class covariance matrix, and does not require data preprocessing, which improves the data dimensionality reduction method for the original data. adaptability of traits.
  • a method for classifying ground objects of hyperspectral images proposed by the present invention maximizes the sum of the ratios of the inter-class distances and the intra-class distances of all feature dimensions in the projection subspace, which can avoid selecting sample features with small variance during feature extraction , which is beneficial to improve the classification accuracy.
  • the alternate iterative optimization method is used to solve the problem of maximizing the sum of ratios.
  • the solution process does not involve the matrix inversion step, which avoids the problem that the classification method based on linear discriminant analysis cannot be solved due to the singularity of the intra-class covariance matrix. Therefore, the present invention can better realize dimensionality reduction of high-dimensional data and extract more effective features, thereby reducing the difficulty of storing hyperspectral data, improving data processing speed, extracting more effective features of data, and finally improving the classification accuracy of ground object classification .
  • Fig. 1 is a flow chart of the dimensionality reduction method of the present invention.
  • Figure 2 is a grayscale image of the actual scene.
  • Figure 3 is a graph of the results of the classification accuracy of ground objects.
  • the present invention proposes a ground object classification method based on maximizing ratio and linear discriminant analysis of hyperspectral images, comprising the following steps:
  • Step 1 Obtain a set of hyperspectral images with a feature dimension of d (that is, the total number of hyperspectral bands is d), and the feature dimension d in the actual ground object dataset used is 103.
  • the value of the feature is the gray value of the corresponding pixel in each band.
  • the total number of pixels in a single band is n.
  • the number of training samples is 2074, and a total of 10 categories of ground object category labels are obtained for all pixels, and then the data matrix, label vector, label matrix, Within-class covariance matrix and between-class covariance matrix. It is mainly divided into the following two processes:
  • X and G are the sample matrix and label matrix calculated according to step (1), respectively, is a unit matrix of order n, is an n-dimensional all 1-column vector.
  • Step 2 Establish an optimization problem and solve the optimal projection matrix, which is mainly divided into the following two processes:
  • m is the final feature dimension of the subspace to be projected
  • ⁇ is an adaptive parameter, which needs to be a sufficiently large number to ensure that the algorithm converges, and the value here is Tr( Sw ) ⁇ 10 10 .
  • ⁇ 2 is the convergence accuracy, which can be artificially given according to the actual application situation, and is set to 10 -6 here.
  • ⁇ 1 is the convergence accuracy, which can be artificially given according to the actual application situation, and is set to 10 -6 here.
  • Step 3 Use the same hyperspectral camera to take a hyperspectral image of the area that needs to be classified, and obtain a hyperspectral image whose feature dimension is still d.
  • the feature dimension of the image used this time is 103, and the feature value is the gray of a single band.
  • the gray value after metricization, the total number of pixels in a single band is n', and the total number of test samples is 8296. The acquisition of the original features of these samples is exactly the same as that of the training data set.
  • step 2 Use the projection matrix obtained in step 2 to project the data matrix formed by the gray levels of corresponding pixels in all bands in the acquired hyperspectral image of the unknown label, and obtain the projected sample matrix where each column represents the value of a new set of features for a hyperspectral image pixel of unknown label, the total number of new features is m, that is, the subspace dimension is,
  • Baseline is the classification result of directly using the trained K-nearest neighbor classifier by using the original training data
  • RSLDA is the classification result of the present invention using the trained K-nearest neighbor classifier after dimensionality reduction of the original data.
  • the dimension of the subspace that is, the number of new features is specified from 1 to 60
  • the data dimensionality reduction method of the present invention can obtain higher classification accuracy by combining the classifier for classification, and it will not be affected by intra-class coordination during the calculation process.
  • the singularity of the variance matrix makes data dimensionality reduction algorithms unavailable.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Multimedia (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Data Mining & Analysis (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Software Systems (AREA)
  • Medical Informatics (AREA)
  • Databases & Information Systems (AREA)
  • Computing Systems (AREA)
  • Remote Sensing (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Spectroscopy & Molecular Physics (AREA)
  • Oral & Maxillofacial Surgery (AREA)
  • Human Computer Interaction (AREA)
  • Astronomy & Astrophysics (AREA)
  • Image Analysis (AREA)

Abstract

本发明涉及一种基于最大化比率和线性判别分析的数据降维方法，属于图像分类与模式识别领域。包括构建数据矩阵、标签向量和标签矩阵；计算类内协方差矩阵和类间协方差矩阵；构建基于最大化比率和的线性判别分析的优化问题；采用交替优化迭代算法求解能够最大化目标函数的投影矩阵。本发明建立了基于最大化比率和的线性判别分析方法的目标函数，避免了传统线性判别分析倾向于选择小方差、鉴别能力弱的特征的问题，能够选择更有利于分类的特征。不依赖类内协方差矩阵的逆矩阵的计算，不需要进行数据预处理，提高了数据降维方法对原始数据特征的适应性。

Description

一种基于最大化比率和线性判别分析的数据降维方法 技术领域
本发明属于图像分类与模式识别领域，具体涉及一种基于最大化比率和线性判别分析的数据降维方法。
背景技术
数据降维技术是图像分类和模式识别领域的一个重要的研究课题。在大数据背景下，在实际应用场景中直接获取的原始数据量是巨大的，这些数据的高维度和高冗余对数据存储和数据处理造成了极大的困难，并且提高了对数据存储和处理的硬件平台的要求。数据降维是对原始高维数据进行降维处理，在保证降维后的数据仍然保留原始数据包含的大部分信息的同时，尽可能降低数据的维度，以提高数据存储和处理效率，降低对硬件和后续数据处理算法的要求。由于数据降维能减少数据维度和需要的存储空间，节约模型训练计算时间，提高后面应用算法的准确度，数据降维技术已经被广泛应用于人脸识别、高光谱图像分类、中草药分类的等实际应用场景的前期数据处理中。
目前的数据降维方法包括特征选择和特征提取两类，特征选择方法是从原始高维数据中直接选择关键特征，特征提取是将现有特征投影到新空间形成新特征。前者有利于保留数据的物理意义，而后者获得的新特征难以解释，但是特征选择的效果相对于特征提取略逊一筹。线性判别分析时特征提取的常用方法，它能够很好地保留数据的判别信息，常用于数据分类的预处理步骤。但是传统的线性判别分析的降维后数据维度不能高于或等于数据的总类别数，并且在求解过程中容易由于矩阵的非奇异性导致不可求解。
图像分类的原始数据特征往往是高维度的，过高的维度反而可能导致分类精度降低，并且原始数据含有冗余信息，将其直接用于分类工作会导致数据处理速度慢、分 类精度低等问题。随着高光谱技术的发展，高光谱成像已经广泛应用于地物分类，如何对高维度的高光谱数据进行降维，进而降低数据存储处理成本、提取数据的关键特征和类别信息，具有重大的实际意义。
杨明莉，范玉刚，李宝芸(《基于LDA和ELM的高光谱图像降维与分类方法研究》,电子测量与仪器学报,2020,34(05):190-196.)采用线性判别分析对高光谱图像数据进行降维，然后采用极限学习机进行分类。线性判别分析通过学习一个投影矩阵将原始数据从高维空间投影到低维空间，使子空间内的同类样本的相似性更高，不同类别的样本的区别更大。因此线性判别分析能够较好地降低数据维度，并且强化类别特征，进而使最终的高光谱图像地物分类的速度提高、分类精度提高。
但是采用传统的线性判别分析方法趋向于选择方差小的特征，这些特征难以有效区分类别，并且传统线性判别分析在求解过程中需要对类内协方差矩阵求逆，但是很多时候该矩阵是奇异的，此时该方法失效，不能进行对高光图像进行数据降维。通常情况下，传统的线性判别分析必须通过预处理保证类内协方差矩阵的奇异性，才能实现高维数据的降维，从而导致数据处理流程复杂以及预处理算法和数据降维算法相互影响的问题。
发明内容
要解决的技术问题
针对已经提出的传统线性判别分析趋向于选择方差小且判别力低的特征，并且优化问题的求解依赖于可逆的类内协方差矩阵的问题，本发明提出了一种基于最大化比率和线性判别分析的数据降维方法；由于降维方法的不完善导致在图像分类方法中的效率低和准确率不高的问题，本发明提出了一种高光谱图像的地物分类方法。
技术方案
一种基于最大化比率和线性判别分析的数据降维方法，其特征在于步骤如下：
步骤1：根据图像构建数据矩阵、标签向量和标签矩阵；所述的图像为高光谱图像、中草药图像或人脸图像；
步骤2：计算类内协方差矩阵和类间协方差矩阵；
步骤3：构建基于最大化比率和的线性判别分析的优化问题
设计最优化问题的目标函数为
Figure PCTCN2021090835-appb-000001
其中优化变量为投影矩阵
Figure PCTCN2021090835-appb-000002
约束为W TW＝I；构建F(W)的等价目标函数
Figure PCTCN2021090835-appb-000003
其中优化变量为投影矩阵
Figure PCTCN2021090835-appb-000004
以及向量
Figure PCTCN2021090835-appb-000005
约束为W TW＝I；其中，m是最终要投影的子空间特征维度，σ是一个自适应参数；S w为类内协方差矩阵和S b为类间协方差矩阵；
步骤4：求解能够最大化目标函数的投影矩阵。
本发明技术方案更进一步的说：步骤1具体如下：获得一组特征维度为d的n个样本，
Figure PCTCN2021090835-appb-000006
x i的每一个元素值为数据对应的样本的特征的值，n个样本的类别标签向量为
Figure PCTCN2021090835-appb-000007
其中y i＝1,2,...,c表示第i个样本的类别，c为样本的类别总数；采用n个样本构建训练样本矩阵
Figure PCTCN2021090835-appb-000008
利用标签向量构造标签矩阵
Figure PCTCN2021090835-appb-000009
其中标签矩阵的每一个元素的值定义为：
Figure PCTCN2021090835-appb-000010
其中i＝1,2,...,n，j＝1,2,...,c。
本发明技术方案更进一步的说：步骤2具体如下：
S w＝X((I n-G(G TG) -1G T)/n)X T
Figure PCTCN2021090835-appb-000011
其中X和G分别是步骤1得到的样本矩阵和标签矩阵，
Figure PCTCN2021090835-appb-000012
是n阶单位矩阵，
Figure PCTCN2021090835-appb-000013
是一个n维全1列向量。
本发明技术方案更进一步的说：步骤4中采用交替迭代优化算法求解投影矩阵
Figure PCTCN2021090835-appb-000014
一种采用上述的降维方法实现的高光谱图像的地物分类方法，其中步骤1中的样本为高光谱图像，特征值取单一波段的灰度化之后的灰度值；n为单一波段的像素总数，c为像素地物类别总数；依次进行步骤1-4得到投影矩阵；利用投影矩阵对获取的未知标签的高光谱图像内对应像素在所有波段的灰度构成的数据矩阵进行投影，得到投影后的样本矩阵Z；将Z的每一列作为新的未知标签的地物对应的像素的所有特征序列，将投影后的新的像素样本采用已经用训练样本训练好的K近邻分类器进行分类，最后得到未知标签地物对应的像素的类别标签。
优选地：所述的K近邻分类器的k＝3。
有益效果
本发明提出的一种基于最大化比率和线性判别分析的数据降维方法，建立了基于最大化比率和的线性判别分析方法的目标函数，避免了传统线性判别分析倾向于选择小方差、鉴别能力弱的特征的问题，能够选择更有利于分类的特征。
采用交替优化迭代算法求解最大化比率和的线性判别分析的最优化问题，该算法不依赖类内协方差矩阵的逆矩阵的计算，不需要进行数据预处理，提高了数据降维方法对原始数据特征的适应性。
本发明提出的一种高光谱图像的地物分类方法，最大化投影子空间内所有特征维度的类间距离和类内距离之比的和，这能够避免在特征提取时选择小方差的样本特征，有利于提高分类精度。采用交替迭代优化方法对最大化比率和的问题进行求解，求解过程不涉及矩阵求逆步骤，避免了类内协方差矩阵奇异导致的基于线性判别分析的分 类方法不可求解的问题。因此，本发明能够更好地实现高维数据降维并且提取更加有效的特征，从而降低高光谱数据存储难度、提高数据处理速度、提取数据更加有效的特征，最终提高地物分类的分类精确度。
附图说明
图1是本发明降维方法流程图。
图2是实际场景的灰度图像。
图3是地物分类精度结果图。
具体实施方式
现结合实施例、附图对本发明作进一步描述：
本发明最大化比率和线性判别分析的数据降维方法基本流程如图1所示，下面结合对实际场景的高光谱图像的地物分类实例说明本发明的具体实施方式，但本发明的技术内容不限于所述的范围。
本发明提出一种基于最大化比率和线性判别分析的高光谱图像的地物分类方法，包括以下步骤：
步骤一、获取一组特征维度均为d的高光谱图像(即高光谱波段总数为d)，采用的实际地物数据集中特征维度d为103。特征的值为每个波段对应像素的灰度值。其单一波段的像素总数为n个，对于使用的真实数据集等距采样，训练样本为2074个，并且获得所有像素的地物类别标签共10类，进而构造数据矩阵、标签向量、标签矩阵、类内协方差矩阵和类间协方差矩阵。主要分为如下两个过程：
(1)一组特征维度均为d的高光谱图像(特征值取单一波段的灰度化之后的灰度值)，其单一波段的像素总数为n个，第i个像素的所有特征表示为
Figure PCTCN2021090835-appb-000015
其中i＝1,2,...,n，x i的第j个元素代表第i个像素的第j个特征的值，
Figure PCTCN2021090835-appb-000016
表示所有训练数据集的标签向量，其中 y i＝1,2,...,c，c为像素地物类别总数。
(2)利用标签向量
Figure PCTCN2021090835-appb-000017
构造标签矩阵
Figure PCTCN2021090835-appb-000018
其中标签矩阵的每一个元素的值定义为：
Figure PCTCN2021090835-appb-000019
其中i＝1,2,...,n，j＝1,2,...,c，即每一个像素在其对应的类的位置值为1，其他位置为零。计算类内协方差矩阵S w和类间协方差矩阵S b：
S w＝X((I n-G(G TG) -1G T)/n)X T
Figure PCTCN2021090835-appb-000020
其中X和G分别是根据步骤(1)计算的样本矩阵和标签矩阵，
Figure PCTCN2021090835-appb-000021
是n阶单位矩阵，
Figure PCTCN2021090835-appb-000022
是一个n维全1列向量。
步骤二、建立最优化问题，并求解最优投影矩阵，主要分为如下两个过程：
(1)建立最优化问题，得到最优投影矩阵需要求解最大化子空间每一个特征维度内样本类间距离和类内距离的比率和，即求解最优化问题：
Figure PCTCN2021090835-appb-000023
并建立其等价问题：
Figure PCTCN2021090835-appb-000024
其中，m是最终要投影的子空间特征维度，σ是一个自适应参数，需要取一个充分大的数保证算法收敛，这里取值为Tr(S w)×10 10。
(2)采用交替迭代优化算法按如下步骤求解最优化问题得到投影矩阵
Figure PCTCN2021090835-appb-000025
①初始化投影矩阵：任意给定初值W 0＝[w 0,1,w 0,2,...,w 0,m]，使其满足W 0 TW 0＝I，t 1＝0。
②计算
Figure PCTCN2021090835-appb-000026
其中k＝1,2,...,m。
③按如下步骤更新W，计算
Figure PCTCN2021090835-appb-000027
i.首先进行初始化，t 2＝0，初始化
Figure PCTCN2021090835-appb-000028
Figure PCTCN2021090835-appb-000029
ii.计算矩阵
Figure PCTCN2021090835-appb-000030
其中k＝1,2,...,m：
Figure PCTCN2021090835-appb-000031
Figure PCTCN2021090835-appb-000032
iii.计算
Figure PCTCN2021090835-appb-000033
且计算矩阵
Figure PCTCN2021090835-appb-000034
iv.对
Figure PCTCN2021090835-appb-000035
进行完整的奇异值分解：
Figure PCTCN2021090835-appb-000036
其中
Figure PCTCN2021090835-appb-000037
为酉矩阵，
Figure PCTCN2021090835-appb-000038
为对角阵，且该阵的对角线奇异值是由大到小排列，
Figure PCTCN2021090835-appb-000039
为酉矩阵。
v.计算
Figure PCTCN2021090835-appb-000040
vi.判断是否满足收敛条件
Figure PCTCN2021090835-appb-000041
不满足收敛条件，则令t 2＝t 2+1，并回到步骤ii.继续计算；满足收敛条件，则令
Figure PCTCN2021090835-appb-000042
停止计算进入步骤④。其中ε 2为收敛精度可以根据实际应用情况人为给定，此处设置为10 -6。
④判断目标函数值
Figure PCTCN2021090835-appb-000043
是否满足收敛条件
Figure PCTCN2021090835-appb-000044
若不满足收敛条件，令t 1＝t 1+1，返回步骤②；若满足收敛条件，则停止计算，需要求解的投影矩阵
Figure PCTCN2021090835-appb-000045
至此，经过学习得到的投影矩阵就得到了。其中ε 1为收敛精度可以根据实际应用情况人为给定，此处设置为10 -6。
步骤三、用同一个高光谱相机拍摄需要进行地物分类的区域的高光谱图像，获得特征维度仍为d的高光谱图像，本次使用的图像特征维度为103，特征值取单一波段的 灰度化之后的灰度值，其单一波段像素总数为n'，测试样本的总数为8296，这些样本的原始特征的获取与训练数据集的获取方式完全相同。
第i个像素的所有特征表示为
Figure PCTCN2021090835-appb-000046
x i'的第j个元素代表图像的第j个特征的值。然后对未知标签的高光谱图像像素进行分类，主要分为以下两个过程：
(1)利用步骤二得到的投影矩阵对获取的未知标签的高光谱图像内对应像素在所有波段的灰度构成的数据矩阵进行投影，得到投影后的样本矩阵
Figure PCTCN2021090835-appb-000047
其中每一列代表一个未知标签的高光谱图像像素的新的一组特征的值，新特征总数为m，即子空间维度为，即
Figure PCTCN2021090835-appb-000048
(2)将Z的每一列作为新的未知标签的地物对应的像素的所有特征序列，将投影后的新的像素样本采用已经用训练样本训练好的K近邻分类器(k＝3)进行分类，最后得到未知标签地物对应的像素的类别标签。
从分类结果可以看出，Baseline是采用原始训练数据直接用训练好的K近邻分类器的分类结果，RSLDA是本发明对原始数据降维后采用训练好的K近邻分类器的分类结果。当子空间维度即新的特征数规定分别从1取到60时，本发明的数据降维方法结合分类器进行分类可以得到较高的分类精确度，并且在计算过程中不会因为类内协方差矩阵的奇异性导致数据降维算法不可用。
以上所述，仅为本发明的具体实施方式，但本发明的保护范围并不局限于此，任何熟悉本技术领域的技术人员在本发明公开的技术范围内，可轻易想到各种等效的修改或替换，这些修改或替换都应涵盖在本发明的保护范围之内。

Claims (6)

  1. 一种基于最大化比率和线性判别分析的数据降维方法，其特征在于步骤如下：
    步骤1：构建数据矩阵、标签向量和标签矩阵；所述的数据为高光谱图像、中草药图像或人脸图像；
    步骤2：计算类内协方差矩阵和类间协方差矩阵；
    步骤3：构建基于最大化比率和的线性判别分析的优化问题
    设计最优化问题的目标函数为
    Figure PCTCN2021090835-appb-100001
    其中优化变量为投影矩阵
    Figure PCTCN2021090835-appb-100002
    约束为W TW＝I；S w、S b分别为类内协方差矩阵和为类间协方差矩阵；构建F(W)的等价目标函数
    Figure PCTCN2021090835-appb-100003
    其中优化变量为投影矩阵
    Figure PCTCN2021090835-appb-100004
    以及向量
    Figure PCTCN2021090835-appb-100005
    约束为W TW＝I；其中，m是最终要投影的子空间特征维度，σ是一个自适应参数；
    步骤4：求解能够最大化目标函数的投影矩阵。
  2. 根据权利要求1所述的一种基于最大化比率和线性判别分析的数据降维方法，其特征在于步骤1具体如下：获得一组特征维度为d的n个样本，
    Figure PCTCN2021090835-appb-100006
    x i的每一个元素值为数据对应的样本的特征的值，n个样本的类别标签向量为
    Figure PCTCN2021090835-appb-100007
    其中y i＝1,2,...,c表示第i个样本的类别，c为样本的类别总数；采用n个样本构建训练样本矩阵
    Figure PCTCN2021090835-appb-100008
    利用标签向量构造标签矩阵
    Figure PCTCN2021090835-appb-100009
    其中标签矩阵的每一个元素的值定义为：
    Figure PCTCN2021090835-appb-100010
    其中i＝1,2,...,n，j＝1,2,...,c。
  3. 根据权利要求2所述的一种基于最大化比率和线性判别分析的数据降维方法，其 特征在于步骤2具体如下：
    S w＝X((I n-G(G TG) -1G T)/n)X T
    Figure PCTCN2021090835-appb-100011
    其中X和G分别是步骤1得到的样本矩阵和标签矩阵，
    Figure PCTCN2021090835-appb-100012
    是n阶单位矩阵，
    Figure PCTCN2021090835-appb-100013
    是一个n维全1列向量。
  4. 根据权利要求1所述的一种基于最大化比率和线性判别分析的数据降维方法，其特征在于步骤4中采用交替迭代优化算法求解投影矩阵
    Figure PCTCN2021090835-appb-100014
  5. 一种采用权利要求2所述的降维方法的高光谱图像的地物分类方法，其特征在于其中步骤1中的样本为高光谱图像，特征值取单一波段的灰度化之后的灰度值；n为单一波段的像素总数，c为像素地物类别总数；依次进行步骤1-4得到投影矩阵；利用投影矩阵对获取的未知标签的高光谱图像内对应像素在所有波段的灰度构成的数据矩阵进行投影，得到投影后的样本矩阵Z；将Z的每一列作为新的未知标签的地物对应的像素的所有特征序列，将投影后的新的像素样本采用已经用训练样本训练好的K近邻分类器进行分类，最后得到未知标签地物对应的像素的类别标签。
  6. 根据权利要求5所述的一种高光谱图像的地物分类方法，其特征在于所述的K近邻分类器的k＝3。
PCT/CN2021/090835 2021-02-26 2021-04-29 一种基于最大化比率和线性判别分析的数据降维方法 WO2022178978A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US17/997,684 US20240029431A1 (en) 2021-02-26 2021-04-29 A data dimension reduction method based on maximizing ratio sum for linear discriminant analysis

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202110216054.1 2021-02-26
CN202110216054.1A CN112836671B (zh) 2021-02-26 2021-02-26 一种基于最大化比率和线性判别分析的数据降维方法

Publications (1)

Publication Number Publication Date
WO2022178978A1 true WO2022178978A1 (zh) 2022-09-01

Family

ID=75933718

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/090835 WO2022178978A1 (zh) 2021-02-26 2021-04-29 一种基于最大化比率和线性判别分析的数据降维方法

Country Status (3)

Country Link
US (1) US20240029431A1 (zh)
CN (1) CN112836671B (zh)
WO (1) WO2022178978A1 (zh)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116126931A (zh) * 2022-12-08 2023-05-16 湖北华中电力科技开发有限责任公司 一种配电台区用电数据挖掘方法、装置、系统及存储介质
CN116347104A (zh) * 2023-05-22 2023-06-27 宁波康达凯能医疗科技有限公司 基于高效判别分析的帧内图像编码方法、装置及存储介质
CN116341396A (zh) * 2023-05-30 2023-06-27 青岛理工大学 一种基于多源数据融合的复杂装备数字孪生建模方法
CN116542956A (zh) * 2023-05-25 2023-08-04 广州机智云物联网科技有限公司 一种织物组分自动检测方法、系统及可读存储介质
CN117493858A (zh) * 2023-12-29 2024-02-02 湖北神龙工程测试技术有限公司 基于人工智能的基桩完整性识别方法

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104008394A (zh) * 2014-05-20 2014-08-27 西安电子科技大学 基于近邻边界最大的半监督高光谱数据降维方法
CN108520279A (zh) * 2018-04-12 2018-09-11 上海海洋大学 一种局部稀疏嵌入的高光谱图像半监督降维方法
CN108805061A (zh) * 2018-05-30 2018-11-13 西北工业大学 基于局部自适应判别分析的高光谱图像分类方法
WO2019045147A1 (ko) * 2017-08-29 2019-03-07 한밭대학교 산학협력단 딥러닝을 pc에 적용하기 위한 메모리 최적화 방법
CN111553417A (zh) * 2020-04-28 2020-08-18 厦门大学 基于判别正则化局部保留投影的图像数据降维方法及系统

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6609093B1 (en) * 2000-06-01 2003-08-19 International Business Machines Corporation Methods and apparatus for performing heteroscedastic discriminant analysis in pattern recognition systems
CN104809475B (zh) * 2015-05-06 2018-05-04 西安电子科技大学 基于增量线性判别分析的多类标场景分类方法
CN107944482B (zh) * 2017-11-17 2021-10-19 上海海洋大学 一种基于半监督学习的高光谱图像的降维方法
CN108845974A (zh) * 2018-04-24 2018-11-20 清华大学 采用最小最大概率机的分离概率的有监督线性降维方法
CN111191700B (zh) * 2019-12-20 2023-04-18 长安大学 基于自适应协同图判别分析的高光谱图像降维方法及装置
CN112116017B (zh) * 2020-09-25 2024-02-13 西安电子科技大学 基于核保持的图像数据降维方法

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104008394A (zh) * 2014-05-20 2014-08-27 西安电子科技大学 基于近邻边界最大的半监督高光谱数据降维方法
WO2019045147A1 (ko) * 2017-08-29 2019-03-07 한밭대학교 산학협력단 딥러닝을 pc에 적용하기 위한 메모리 최적화 방법
CN108520279A (zh) * 2018-04-12 2018-09-11 上海海洋大学 一种局部稀疏嵌入的高光谱图像半监督降维方法
CN108805061A (zh) * 2018-05-30 2018-11-13 西北工业大学 基于局部自适应判别分析的高光谱图像分类方法
CN111553417A (zh) * 2020-04-28 2020-08-18 厦门大学 基于判别正则化局部保留投影的图像数据降维方法及系统

Cited By (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116126931A (zh) * 2022-12-08 2023-05-16 湖北华中电力科技开发有限责任公司 一种配电台区用电数据挖掘方法、装置、系统及存储介质
CN116126931B (zh) * 2022-12-08 2024-02-13 湖北华中电力科技开发有限责任公司 一种配电台区用电数据挖掘方法、装置、系统及存储介质
CN116347104A (zh) * 2023-05-22 2023-06-27 宁波康达凯能医疗科技有限公司 基于高效判别分析的帧内图像编码方法、装置及存储介质
CN116347104B (zh) * 2023-05-22 2023-10-17 宁波康达凯能医疗科技有限公司 基于高效判别分析的帧内图像编码方法、装置及存储介质
CN116542956A (zh) * 2023-05-25 2023-08-04 广州机智云物联网科技有限公司 一种织物组分自动检测方法、系统及可读存储介质
CN116542956B (zh) * 2023-05-25 2023-11-17 广州机智云物联网科技有限公司 一种织物组分自动检测方法、系统及可读存储介质
CN116341396A (zh) * 2023-05-30 2023-06-27 青岛理工大学 一种基于多源数据融合的复杂装备数字孪生建模方法
CN116341396B (zh) * 2023-05-30 2023-08-11 青岛理工大学 一种基于多源数据融合的复杂装备数字孪生建模方法
CN117493858A (zh) * 2023-12-29 2024-02-02 湖北神龙工程测试技术有限公司 基于人工智能的基桩完整性识别方法
CN117493858B (zh) * 2023-12-29 2024-03-26 湖北神龙工程测试技术有限公司 基于人工智能的基桩完整性识别方法

Also Published As

Publication number Publication date
CN112836671A (zh) 2021-05-25
US20240029431A1 (en) 2024-01-25
CN112836671B (zh) 2024-03-08

Similar Documents

Publication Publication Date Title
WO2022178978A1 (zh) 一种基于最大化比率和线性判别分析的数据降维方法
Martin et al. Learning to detect natural image boundaries using brightness and texture
Jiang Asymmetric principal component and discriminant analyses for pattern classification
CN111126482B (zh) 一种基于多分类器级联模型的遥感影像自动分类方法
Li et al. Overview of principal component analysis algorithm
Bekhouche et al. Pyramid multi-level features for facial demographic estimation
CN108647690B (zh) 基于判别稀疏保持投影的非约束人脸图像降维方法
CN111126240B (zh) 一种三通道特征融合人脸识别方法
CN104318219A (zh) 基于局部特征及全局特征结合的人脸识别方法
CN113033398B (zh) 一种手势识别方法、装置、计算机设备及存储介质
CN109241813B (zh) 基于判别稀疏保持嵌入的非约束人脸图像降维方法
Jiang Feature extraction for image recognition and computer vision
Gao et al. A robust geometric mean-based subspace discriminant analysis feature extraction approach for image set classification
CN110188646B (zh) 基于梯度方向直方图与局部二值模式融合的人耳识别方法
CN111325275A (zh) 基于低秩二维局部鉴别图嵌入的鲁棒图像分类方法及装置
Huang et al. Asymmetric 3D/2D face recognition based on LBP facial representation and canonical correlation analysis
CN103942572A (zh) 一种基于双向压缩数据空间维度缩减的面部表情特征提取方法和装置
CN110287973B (zh) 一种基于低秩鲁棒线性鉴别分析的图像特征提取方法
Yuan et al. Holistic learning-based high-order feature descriptor for smoke recognition
Kejun et al. Automatic nipple detection using cascaded adaboost classifier
Yin et al. Video text localization based on Adaboost
Huang et al. Skew correction of handwritten Chinese character based on ResNet
Liu et al. Gabor feature representation method based on block statistics and its application to facial expression recognition
Li et al. Shadow determination and compensation for face recognition
Wang et al. Feature extraction method of face image texture spectrum based on a deep learning algorithm

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21927411

Country of ref document: EP

Kind code of ref document: A1

WWE Wipo information: entry into national phase

Ref document number: 17997684

Country of ref document: US

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 21927411

Country of ref document: EP

Kind code of ref document: A1

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 19.02.2024)