WO2022041678A1 - 张量协作图判别分析遥感图像特征提取方法 - Google Patents
张量协作图判别分析遥感图像特征提取方法 Download PDFInfo
- Publication number
- WO2022041678A1 WO2022041678A1 PCT/CN2021/079598 CN2021079598W WO2022041678A1 WO 2022041678 A1 WO2022041678 A1 WO 2022041678A1 CN 2021079598 W CN2021079598 W CN 2021079598W WO 2022041678 A1 WO2022041678 A1 WO 2022041678A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- tensor
- matrix
- data
- training
- feature extraction
- Prior art date
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
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/42—Global feature extraction by analysis of the whole pattern, e.g. using frequency domain transformations or autocorrelation
- G06V10/422—Global feature extraction by analysis of the whole pattern, e.g. using frequency domain transformations or autocorrelation for representing the structure of the pattern or shape of an object therefor
-
- 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
- G06V10/771—Feature selection, e.g. selecting representative features from a multi-dimensional feature space
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- 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
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/29—Graphical models, e.g. Bayesian networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/42—Global feature extraction by analysis of the whole pattern, e.g. using frequency domain transformations or autocorrelation
- G06V10/422—Global feature extraction by analysis of the whole pattern, e.g. using frequency domain transformations or autocorrelation for representing the structure of the pattern or shape of an object therefor
- G06V10/426—Graphical representations
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/58—Extraction of image or video features relating to hyperspectral data
-
- 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
- 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 relates to image feature extraction in the field of image processing, in particular to a graph discriminant analysis feature extraction technology of remote sensing images, in particular to a tensor collaboration graph discriminant analysis remote sensing image feature extraction method.
- grayscale images are second-order tensors
- color images are third-order tensors.
- data are often assembled into tensor patterns artificially, such as environment
- the data in monitoring can be regarded as a third-order tensor with patterns of time, location, and type. Patterns in the form of tensors are used in network graph mining, network debate, and face recognition.
- data is generally represented in vector mode, that is, no matter whether the original data is a one-dimensional vector, two-dimensional matrix or high-order tensor, it is almost always converted into the corresponding vector mode for processing.
- Image features are an important basis for image analysis, and the operation of obtaining image feature information is called feature extraction. It is important as a basis for pattern recognition, image understanding or information volume compression.
- the extraction and selection of image features is a very important part in the image processing process, which has an important impact on the subsequent image classification, and has the characteristics of few samples and high dimensionality for image data.
- To extract useful information from images it is necessary to Image features are subjected to dimensionality reduction processing.
- Feature extraction and feature selection are the most effective dimensionality reduction methods. The purpose is to obtain a feature subspace that reflects the essential structure of the data and has a higher recognition rate.
- the spectral cluster groups between different images and images of different time periods cannot maintain their continuity, which makes it difficult to compare different images.
- the traditional way of manually interpreting remote sensing images has been difficult to apply. Instead, the computer automatically extracts remote sensing image information.
- the corresponding data processing algorithms generally have the shortcomings of insufficient adaptive ability.
- the original sampling data must be transformed to obtain the features that can best reflect the essence, which is the process of feature extraction and selection.
- the so-called hyperspectral image feature extraction refers to reducing the spectral dimension on the basis of removing redundancy and retaining effective information to reduce the complexity of the data.
- Hyperspectral image classification refers to the use of different features with different spectral feature information to distinguish the categories of different features in the image.
- Hyperspectral remote sensing earth observation technology provides refined image data for the detection of ground objects.
- Hyperspectral images are multi-spectral images, containing dozens or even hundreds of continuous bands with rich spectral characteristics. These data not only contain rich ground objects Spectral information also contains spatial structure information with higher and higher resolution.
- Spectral information also contains spatial structure information with higher and higher resolution.
- the increase of frequency bands will inevitably lead to the redundancy of information and the increase of the complexity of data processing.
- There is a strong correlation between these bands of hyperspectral images which not only brings great information redundancy, but also increases the computational burden of hyperspectral data classification.
- the "Hughes phenomenon” also known as the curse of dimensionality
- feature extraction becomes a key preprocessing step for hyperspectral image analysis.
- Principal Component Analysis is one of the most classic unsupervised feature extraction methods. Its purpose is to find a linear transformation matrix that maximizes the variance of the data, so as to keep the important information contained in the data at a low level obtained by projection. dimension feature. Due to the prior label information of unused samples, the performance of unsupervised methods is usually difficult to meet the needs of practical applications. In order to use the prior information of the data to further improve the data processing performance, researchers have done a lot of research work on supervised feature extraction.
- LDA Linear Discriminant Analysis
- spectral vector is usually used as the basic research unit in hyperspectral image analysis.
- studies have shown that spatial information plays a crucial role in hyperspectral image processing, and making full use of its spatial structure information can improve the feature extraction and classification performance of hyperspectral images. Therefore, combining spatial information to carry out hyperspectral image feature extraction research has become a research hotspot.
- Early dimensionality reduction methods based on spatial spectral features consider both spatial information and spectral information to improve the performance to a certain extent, but these methods need to convert the spatial spectral features into vector form for analysis, which usually results in local pixels. The spatial connection is lost.
- At least some embodiments of the present invention provide a supervised feature extraction method with low complexity and good feature extraction performance, so as to at least partially solve the problems in the related art for hyperspectral data with high spectral dimensions, large information redundancy and complex existing methods high degree and insufficient spatial information mining.
- a method for extracting features of remote sensing images by discriminant analysis of tensor collaboration graphs including the following steps:
- the data cooperative representation model calculates the representation coefficient of the current training pixel under each category of training data, and constructs the graph weight matrix and the tensor local preservation projection model; projection matrix; finally, use the low-dimensional projection matrix to obtain the training set and test set represented by the three-dimensional low-dimensional representation, and expand them into the form of column vectors according to the feature dimension, and input the extracted low-dimensional features into the support vector machine classifier for classification.
- the categories of the test set to evaluate the performance of feature extraction in terms of classification effect.
- the embodiment of the present invention constructs a tensor collaboration graph discriminant analysis feature extraction model from two aspects of algorithm complexity and spatial information mining.
- the technology focuses on L2 norm sparse constraint, weight constraint matrix, tensor representation, etc. Cutting-edge mathematical theory, and gives the optimal solution of the model.
- the embodiment of the present invention uses the L2 norm to construct a constrained cooperative representation model to solve the representation coefficient of each pixel in the training set.
- the collaborative representation model based on the L2 norm can obtain closed-form solutions through model derivation, which avoids the high complexity of the orthogonal matching pursuit method of the L1 norm in the sparse graph model.
- the embodiment of the present invention designs a weight constraint matrix, which can constrain the model to select training data similar to the current pixel as much as possible for representation, thereby improving the quality of the representation coefficients.
- the embodiment of the present invention uses the mathematical theory of tensor analysis as a tool to mine the spatial structure information of hyperspectral data by using a tensor representation method for some existing problems of feature extraction and classification algorithms based on tensor data.
- Hyperspectral data which is three-dimensional stereo data consisting of two spatial dimensions and one spectral dimension, fits well with third-order tensors. Therefore, using tensor data blocks to perform cooperative representation operations can better preserve the spatial neighborhood information of the data and improve the accuracy of the representation coefficients.
- the core of the embodiment of the present invention is to construct a tensor cooperative representation model with weight constraints, realize the effective mining of spectral information and spatial information of hyperspectral data, and improve the discrimination ability of low-dimensional features.
- the present invention is effective as long as it relates to image feature extraction or dimensionality reduction. Simulation experiments show that the embodiment of the present invention is obviously superior to the sparse graph discriminant analysis method, the collaboration graph discriminant analysis method and other spatial spectral feature extraction methods in the performance of hyperspectral image feature extraction.
- the embodiments of the present invention are suitable for feature extraction of hyperspectral images.
- FIG. 1 is a schematic diagram of feature extraction of remote sensing image by discriminant analysis of tensor collaboration graph according to one embodiment of the present invention
- FIG. 2 is a flowchart of feature extraction of a Tensor Collaboration Graph discriminant analysis image according to one embodiment of the present invention
- FIG. 3 is a schematic diagram of a modulo-3 expansion of a third-order tensor according to one embodiment of the present invention.
- the vector machine classifier performs classification, determines the category of the test set, and evaluates the performance of feature extraction based on the classification effect.
- Step 1 in an optional embodiment, convert the input raw hyperspectral data According to the set sliding window size, it is cut into third-order tensor blocks, and the tensor data block is divided into training set and test set according to a certain ratio, where A and B respectively represent the two spatial dimensions of hyperspectral data, and D represents hyperspectral data.
- the sliding window size is set to w ⁇ w, then a third-order tensor data block can be expressed as
- the training set obtained by proportional division consists of N samples containing C categories, which are expressed as
- the test set consists of M samples, denoted as in, Indicates the jth test data block, 1 ⁇ j ⁇ M.
- Step 2 In the construction of the weight constraint matrix, the data blocks in the training set are divided into C sub-data sets according to categories, and the l-th sub-data set is There are a total of Nl samples, and the lth sub-data set is the ith sample in Expand into vector form modulo 3
- the Euclidean distance from the jth sample in the lth subdataset is Finally, (N l -1) Euclidean distances are obtained, where 1 ⁇ j ⁇ N l , j ⁇ i, and
- This embodiment of the present invention adopts the method of intra-class representation, therefore, when calculating the Euclidean distance does not include Euclidean distance from itself. Taking (N l -1) Euclidean distances as the diagonal elements of the symmetric matrix, construct the weight constraint matrix of the lth class as shown below
- Step 3 In the construction of the collaborative representation model with weight constraints, the L2 norm is used to realize the training samples Represents the sparse constraints of the coefficients, reduces the complexity of the model, and at the same time improves the representation ability of the coefficients with the weight constraint matrix.
- This embodiment adopts the method of intra-class representation, that is, the training sample Only the samples belonging to the same class l are used for representation learning, and the collaborative representation model with weight constraints is constructed as follows:
- arg min represents the minimum value of the objective function
- ⁇ is a regularization parameter.
- Step 4 the collaborative representation model with weight constraints is solved.
- the collaborative representation model is based on the L2 norm, and the representation coefficient can be obtained by derivation. the optimal solution of where T represents the transpose of the matrix and ( ) -1 represents the inverse of the matrix.
- Step 5 in the construction of the graph weight matrix, according to the representation coefficient
- the graph weight coefficients of the lth class are obtained as
- the final graph weight matrix constructed from the training samples is
- Step 6 in the solution of the projection matrix, this embodiment adopts the tensor local preservation projection algorithm to solve the projection of the three dimensions in the hyperspectral data block, as shown in the following expression:
- min represents the minimum value of the objective function
- ⁇ represents the summation operation
- ⁇ n represents the n-th module multiplication
- U n represents the projection matrix on the n-th module
- Wi ,j represents that the row number of the graph weight matrix is i
- the column number is j.
- Tr( ) represents the trace of the matrix, represents the n-modulo expansion of the ith data block.
- Step 7 in calculating the low-dimensional features of the training set and the test set, according to the projection matrices U 1 , U 2 , U 3 in the three dimensions obtained in step 6, calculate the low-dimensional features of the training set and the test set:
- Step 8 use the support vector machine classifier to calculate the category of the test set samples after feature extraction, and use the low-dimensional features of the training set Train a support vector machine classifier and then perform low-dimensional features on the test set The classification is performed to evaluate the performance of the feature extraction algorithm with the accuracy of the classification of the test set sample categories.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Data Mining & Analysis (AREA)
- Computing Systems (AREA)
- Databases & Information Systems (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Medical Informatics (AREA)
- Software Systems (AREA)
- Bioinformatics & Computational Biology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- General Engineering & Computer Science (AREA)
- Spectroscopy & Molecular Physics (AREA)
- Remote Sensing (AREA)
- Image Analysis (AREA)
Abstract
一种张量协作图判别分析遥感图像特征提取方法:以每个像素为中心截取三维的张量数据块;按比例将实验数据划分成训练集和测试集;计算当前训练像素与每个类别训练数据的欧式距离,构建对角权重约束矩阵;然后,设计带约束的L2范数协作表示模型,构建图权重矩阵和张量局部保持投影模型;求取对应张量数据块每一个维度的投影矩阵;最后,利用低维投影矩阵得到三维低维表示的训练集和测试集,并按特征维展开成列向量的形式,将提取到的低维特征输入支持向量机分类器进行分类,判定测试集的类别,以分类效果评估特征提取的性能。
Description
本发明涉及图像处理领域的图像特征提取,具体涉及遥感图像的图判别分析特征提取技术,尤其是张量协作图判别分析遥感图像特征提取方法。
在许多应用领域,特别是在云计算、移动互联网、大数据应用方面,会产生大量的高维高阶的数据,采用张量的数学形式能够恰当地表示这些具有多维结构的数据。这些数据往往含有大量的冗余信息,需要对其进行有效地降维。在模式识别中,特征提取(降维)和分类是两个关键步骤。大多经典的特征提取和分类的算法都是基于向量数据的,处理张量数据时需要将其向量化。张量数据向量化的过程会破坏数据的内部结构,维数也会显著增加,使算法的计算量和复杂度也显著增加。在模式识别中经常遇到张量形式的模式.如,灰度图像是二阶张量,彩色图像是三阶张量.为了处理的需要,数据常被人为组装成张量模式,如,环境监控中的数据可视为以时间、位置和类型为模式的三阶张量,网络图挖掘、网络辩论及人脸识别中都用到张量形式的模式。然而在传统统计模式识别中,数据一般采用向量模式表示,即无论原始数据是一维向量、两维矩阵还是高阶张量,几乎总是转换成对应的向量模式来处理。为了便于有效地分析和研究,往往需要对给定的遥感图像用更为简单明确的数值、符号或图形来表征,它们反映了图像中基本的重要的信息,称之为图像的特征。图像特征是图像分析的重要依据,获取图像特征信息的操作称为特征提取。它作为模式识别、图像理解或信息量压缩的基础是很重要的。图像特征的提取和选择是图像处理过程中很重要的环节,对后续图像分类有着重要的影响,并且对于图像数据具有样本少,维数高的特点,要从图像中提取有用的信息,必须对图像特征进行降维处理,特征提取与特征选择就是最有效的降维方法,其目的是得到一个反映数据本质结构、识别率更高的特征子空间。
随着遥感技术的发展,获得遥感图像的波段数不断增多,为我们了解地物提供了极其丰富的遥感信息,这有助于完成更加细致的遥感地物分类和目标识别,然而波段的增多也必然导致信息的冗余和数据处理复杂度的增加。虽然每一种图像数据都可能包含了一些用于自动分类的信息,但就某些指定的地物分类而言,并不是所获得的全部波段图像数据都可用。由于图像中同一类别的光谱差异,造成训练样本并没有很好的代表性。训练样本的选取和评 估需花费较多的人力、时间。如果不加区别地将大量原始图像直接用来分类,不仅数据量太大,计算复杂,而且分类的效果也不一定好。由于图像中各类别的光谱特征会随时间、地形等变化,不同图像以及不同时段的图像之间的光谱集群组无法保持其连续性,从而使其不同图像之间的对比变得困难。传统的人工解译遥感影像的方式已经很难应用,取而代之的是计算机全自动提取遥感影像信息的方法。但是相应的数据处理算法普遍存在自适应能力不足的缺点。为了有效地实现分类识别,必须对原始采样数据进行变换,得到最能反映本质的特征,这就是特征提取和选择的过程。所谓的高光谱图像特征提取是指在去除冗余、保留有效信息的基础上对光谱维进行维数约减,以降低数据的复杂度。高光谱图像分类是指利用不同地物具有不同的光谱特征信息,来区分图像中不同地物的类别。
高光谱遥感对地观测技术为地物探测提供精细化的影像数据,高光谱图像为多光谱图像,包含几十甚至上百个具有丰富光谱特征的连续波段,这些数据不但包含了丰富的地物光谱信息,还包含了分辨率越来越高的空间结构信息。然而波段的增多也必然导致信息的冗余和数据处理复杂度的增加。高光谱图像的这些波段之间具有很强的相关性,不仅带来了极大的信息冗余,也增加了高光谱数据分类的计算负担。此外,样本维数高、数量少所导致的“Hughes现象”(也称为维数的诅咒),也使得高光谱数据分类更具挑战性。因此,特征提取成为了高光谱图像分析的关键预处理步骤。
通常,根据是否使用样本先验信息,将特征提取方法大致地分为无监督和有监督两种类型。主成分分析(Principal Component Analysis,PCA)是一种最经典的无监督特征提取方法,其目的是寻找一个使得数据方差最大化的线性变换矩阵,以将数据蕴含的重要信息保留在投影得到的低维特征中。由于未使用样本的的先验标签信息,无监督方法的性能通常难以满足实际应用需求。为了利用数据的先验信息进一步提高数据处理性能,学者在有监督特征提取方面做了大量的研究工作。线性判别分析(Linear Discriminant Analysis,LDA)是最经典有监督特征提取方法,其目标是寻找一个投影变换,使得投影得到的子空间中作为瑞利商的Fisher比值最大,以增强低维特征的可分性。然而,在小样本情况下(small-sample-size,SSS),LDA的性能通常欠佳。在高光谱遥感影像分类问题中,由于训练样本数量常常远小于光谱特征维数,因此,直接使用常规的线性判别分析算法必然会遭遇到上述小样本问题。为了解决这一问题,研究学者基于LDA提出了大量的判别分析方法。随着稀疏表示(Sparse Representation,SR)在人脸识别方向的成功应用,大量研究学者将稀疏表示引入高光谱图像特征提取与分类领域,提出了稀疏图嵌入、稀疏图判别分析等方法,在特征提取的性能上取得较大突破。后 来,基于低秩表示理论提出了低秩图嵌入的方法。
实际上,前面介绍的特征提取方法都是在向量空间的基础上发展而来的,在高光谱图像分析中通常是以光谱向量作为基本研究单位的。然而,研究表明空间信息在高光谱图像处理中起着至关重要的作用,充分利用其空间结构信息能够提高高光谱图像的特征提取和分类性能。因而,结合空间信息开展高光谱图像特征提取研究成为研究热点。早期基于空谱特征的降维方法同时考虑空间信息和光谱信息虽然在一定程度上带来性能的提升,但这些方法需将空谱特征向转化为向量形式进行分析,通常会造成局部像素之间的空间联系丢失。
尽管人们提出了很多特征提取算法,但现有的特征提取算法基本上还处于实验阶段,其准确性、实用性、通用性等方面离大规模实际应用的要求还有很大差距。综合来看,现有的高光谱图像特征提取算法尚存在两个问题:(1)特征提取算法模型的复杂度过高,以L1范数为基础的稀疏图嵌入和以核范数为基础的低秩图嵌入在求解图权重矩阵的过程中涉及复杂的求解过程;(2)高光谱图像空间信息利用不充分,部分方法通过局部正则化的方式保持像素的局部信息,空间信息的利用存在局限性。
发明内容
本发明至少部分实施例提供了一种复杂度低、特征提取性能好的有监督特征提取方法,以至少部分地解决相关技术中针对高光谱数据光谱维度高、信息冗余大及现有方法复杂度高、空间信息挖掘不充分等问题。
在本发明其中一实施例中,提供了一种张量协作图判别分析遥感图像特征提取方法,包括如下步骤:
首先,设定正方形的滑动窗口尺寸,将高光谱数据的第一个像素作为起点,以每个像素为中心截取得到三维的张量数据块;根据得到的数据块按比例将实验数据划分成训练集和测试集,并将每个数据块按光谱维展开成一个列向量;计算当前训练像素与每个类别训练数据的欧式距离,进而构建对角权重约束矩阵;然后,设计带约束的L2范数协作表示模型,计算当前训练像素在每个类别训练数据下的表示系数,构建图权重矩阵和张量局部保持投影模型;通过张量局部保持投影模型求得对应张量数据块每一个维度的投影矩阵;最后,利用低维投影矩阵得到三维低维表示的训练集和测试集,并按特征维展开成列向量的形式,将提取到的低维特征输入支持向量机分类器进行分类,判定测试集的类别,以分类效果评估特征提取的性能。
本发明实施例相比于现有技术的技术效果在于:
(1)本发明实施例从算法复杂度和空间信息挖掘两个方面,构建了一种张量协作图判别分析 特征提取模型,技术着眼于L2范数稀疏约束、权重约束矩阵、张量表示等前沿的数学理论,并给出了模型的优化求解。
(2)本发明实施例利用L2范数构建带约束的协作表示模型来求解训练集中每个像素的表示系数。与稀疏图判别分析模型相比,基于L2范数的协作表示模型通过模型求导可以得到闭式解,避免了稀疏图模型中L1范数的正交匹配追踪方法求解的高复杂度;与协作图判别分析模型相比,本发明实施例设计了权重约束矩阵,可以约束模型尽可能选择与当前像素相似的训练数据进行表示,改善了表示系数的质量。
(3)本发明实施例以张量分析的数学理论为工具,针对基于张量数据的特征提取和分类算法现有的一些问题,采用张量表示的方法挖掘高光谱数据的空间结构信息。高光谱数据,是由两个空间维和一个光谱维组成的三维立体数据与三阶张量非常契合。因此采用张量数据块的方式进行协作表示运算,能够更好地保留数据的空间邻域信息,提升表示系数的准确性。
本发明实施例的核心是构建带权重约束的张量协作表示模型,实现高光谱数据的光谱信息和空间信息的有效挖掘,提升了低维特征的判别能力。只要是有关图像特征提取或降维,本发明都是有效的。仿真实验表明,本发明实施例在高光谱图像特征提取的性能上明显优于稀疏图判别分析方法、协作图判别分析方法以及其它的空谱特征提取方法。
本发明实施例适用于高光谱图像特征提取。
图1是根据本发明其中一实施例的张量协作图判别分析遥感图像特征提取示意图;
图2是根据本发明其中一实施例的张量协作图判别分析图像特征提取流程图;
图3是根据本发明其中一实施例的三阶张量的模3展开示意图;
为使本发明的目的、技术方案和优点更加清楚明白,以下结合具体实施例,并参照附图,对本发明进一步详细说明。
参阅图1-图3。根据本发明其中一实施例,首先,设定正方形的滑动窗口尺寸,将高光谱数据的第一个像素作为起点,以每个像素为中心截取得到三维的张量数据块;根据得到的数据块按比例将实验数据划分成训练集和测试集,并将每个数据块按光谱维展开成一个列向量;计算当前训练像素与每个类别训练数据的欧式距离,进而构建对角权重约束矩阵;然后,设计带约束的L2范数协作表示模型,计算当前训练像素在每个类别训练数据下的表示系 数,构建图权重矩阵和张量局部保持投影模型;通过张量局部保持投影模型求得对应张量数据块每一个维度的投影矩阵;最后,利用低维投影矩阵得到三维低维表示的训练集和测试集,并按特征维展开成列向量的形式,将提取到的低维特征输入支持向量机分类器进行分类,判定测试集的类别,以分类效果评估特征提取的性能。
参阅图2。本发明实施例具体包括以下步骤:
步骤1,在可选实施例中,将输入的原始高光谱数据
按设置的滑动窗口尺寸切割成三阶张量块,根据一定的比例将张量数据块划分为训练集和测试集,其中A和B分别表示高光谱数据的两个空间维,D表示高光谱数据的光谱维,R表示实数空间。
滑动窗口尺寸设置为w×w,则切割的一个三阶张量数据块可以表示为
根据比例划分得到的训练集由包含了C个类别的N个样本组成,表示为
第l类的样本表示为
其中l=1,2,…,C,
其中,
表示训练集中的第i个数据块,1≤i≤N,N
l表示第l类的训练样本个数,
表示第l类的训练中的第i个数据块。
参见图3。步骤2,在权重约束矩阵构建中,将训练集中的数据块按类别划分成C个子数据集,第l个子数据集为
共有Nl个样本,将第l个子数据集
中的第i个样本
按模3展开成向量形式
与第l类子数据集中第j个样本的欧式距离为
最终获得(N
l-1)个欧式距离,其中,1≤j≤N
l,j≠i,||·||
2表示L2范数。本发明实施例采用类内表示的方法,因此,在计算欧式距离
时不包含
与其自身的欧式距离。将(N
l-1)个欧式距离作为对称矩阵的对角线元素,构建如下所示第l类的权重约束矩阵
步骤3,在带权重约束的协作表示模型构建中,采用L2范数实现训练样本
表示系数的稀疏约束,降低模型的复杂度,同时以权重约束矩阵提升表示系数的表示能力。本实施例采用类内表示的方法,即训练样本
仅用同属于第l类的样本进行表示学习,带权重约束的协作表示模型构建如下:
其中,arg min表示目标函数的最小值,
表示字典,其中的元素包含去掉
的(N
l-1)个样本,样本的维度为Dw
2,
表示矩阵L2范数的平方,
表示
以X
l′为字典时的表示系数,λ表示正则化参数。
其中,W
i表示第i类的类内权重矩阵,i=1,2,…,C,C表示高光谱数据中的类别总数。
步骤6,在投影矩阵求解中,本实施例采用张量局部保持投影算法求解高光谱数据块中三个维度的投影,如下表达式所示,
其中,min表示目标函数最小值,∑表示求和运算,
表示第i个数据块按第n模运算,×
n表示第n模相乘,U
n表示第n模上的投影矩阵,W
i,j表示图权重矩阵的行号为i、列号位j的元素,Tr(·)表示矩阵的迹,
表示第i个数据块的n模展开。
步骤7,在计算训练集和测试集的低维特征中,根据步骤6中求得的三个维度上的投影矩阵U
1、U
2、U
3,计算训练集和测试集的低维特征:
以上所述的具体实施例,对本发明的目的、技术方案和有益效果进行了进一步详细说明,所应理解的是,以上所述仅为本发明的具体实施例而已,并不用于限制本发明,凡在本发明的精神和原则之内,所做的任何修改、等同替换、改进等,均应包含在本发明的保护范围之内。
Claims (10)
- 一种张量协作图判别分析遥感图像特征提取方法,包括:设定正方形的滑动窗口尺寸,将输入的原始高光谱数据的第一个像素作为起点,以每个像素为中心截取得到三维的张量数据块;根据得到的数据块按比例将实验数据划分成训练集和测试集,并将每个数据块按光谱维展开成一个列向量;计算当前训练像素与每个类别训练数据的欧式距离,进而构建对角权重约束矩阵;设计带约束的L2范数协作表示模型,计算当前训练像素在每个类别训练数据下的表示系数,构建图权重矩阵和张量局部保持投影模型;通过张量局部保持投影模型求得对应张量数据块每一个维度的投影矩阵;利用低维投影矩阵得到三维低维表示的训练集和测试集,并按特征维展开成列向量的形式,将提取到的低维特征输入支持向量机分类器进行分类,判定测试集的类别,以分类效果评估特征提取的性能。
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US17/913,855 US20230186606A1 (en) | 2020-08-30 | 2021-03-08 | Tensor Collaborative Graph Discriminant Analysis Method for Feature Extraction of Remote Sensing Images |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010891063.6A CN112101381B (zh) | 2020-08-30 | 2020-08-30 | 张量协作图判别分析遥感图像特征提取方法 |
CN202010891063.6 | 2020-08-30 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2022041678A1 true WO2022041678A1 (zh) | 2022-03-03 |
Family
ID=73756630
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/CN2021/079598 WO2022041678A1 (zh) | 2020-08-30 | 2021-03-08 | 张量协作图判别分析遥感图像特征提取方法 |
Country Status (3)
Country | Link |
---|---|
US (1) | US20230186606A1 (zh) |
CN (1) | CN112101381B (zh) |
WO (1) | WO2022041678A1 (zh) |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114299398A (zh) * | 2022-03-10 | 2022-04-08 | 湖北大学 | 一种基于自监督对比学习的小样本遥感图像分类方法 |
CN114781576A (zh) * | 2022-04-19 | 2022-07-22 | 广东海洋大学 | 一种基于随机森林算法的声速剖面估计方法及装置 |
CN114897818A (zh) * | 2022-05-09 | 2022-08-12 | 北京理工大学 | 基于时空距离矩阵分析的遥感时间序列图像变化检测方法 |
CN115049944A (zh) * | 2022-06-02 | 2022-09-13 | 北京航空航天大学 | 一种基于多任务优化的小样本遥感图像目标检测方法 |
CN115082770A (zh) * | 2022-07-04 | 2022-09-20 | 青岛科技大学 | 一种基于机器学习的图像中心线结构提取方法 |
CN116188995A (zh) * | 2023-04-13 | 2023-05-30 | 国家基础地理信息中心 | 一种遥感图像特征提取模型训练方法、检索方法及装置 |
CN116610927A (zh) * | 2023-07-21 | 2023-08-18 | 傲拓科技股份有限公司 | 基于fpga的风机齿轮箱轴承故障诊断方法及诊断模块 |
CN117173102A (zh) * | 2023-08-04 | 2023-12-05 | 航天恒星科技有限公司 | 基于多源数据融合的遥感数据稳定性分析方法、电子设备及存储介质 |
CN117671673A (zh) * | 2023-11-21 | 2024-03-08 | 江南大学 | 一种基于自适应张量子空间的小样本宫颈细胞分类方法 |
Families Citing this family (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112751792B (zh) * | 2019-10-31 | 2022-06-10 | 华为技术有限公司 | 一种信道估计方法及装置 |
CN112101381B (zh) * | 2020-08-30 | 2022-10-28 | 西南电子技术研究所(中国电子科技集团公司第十研究所) | 张量协作图判别分析遥感图像特征提取方法 |
CN113378942B (zh) * | 2021-06-16 | 2022-07-01 | 中国石油大学(华东) | 基于多头特征协作的小样本图像分类方法 |
CN116863327B (zh) * | 2023-06-05 | 2023-12-15 | 中国石油大学(华东) | 一种基于双域分类器协同对抗的跨域小样本分类方法 |
CN116539167B (zh) * | 2023-07-04 | 2023-09-08 | 陕西威思曼高压电源股份有限公司 | 一种高压电源工作温度分布数据分析方法 |
CN117115669B (zh) * | 2023-10-25 | 2024-03-15 | 中交第二公路勘察设计研究院有限公司 | 双条件质量约束的对象级地物样本自适应生成方法及系统 |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110619263A (zh) * | 2019-06-12 | 2019-12-27 | 河海大学 | 低秩联合协同表示的高光谱遥感影像异常探测方法 |
US20200026953A1 (en) * | 2018-07-23 | 2020-01-23 | Wuhan University | Method and system of extraction of impervious surface of remote sensing image |
CN111191700A (zh) * | 2019-12-20 | 2020-05-22 | 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 |
CN111368691A (zh) * | 2020-02-28 | 2020-07-03 | 西南电子技术研究所(中国电子科技集团公司第十研究所) | 无监督高光谱遥感图像空谱特征提取方法 |
CN111369457A (zh) * | 2020-02-28 | 2020-07-03 | 西南电子技术研究所(中国电子科技集团公司第十研究所) | 稀疏判别张量鲁棒性pca的遥感图像去噪方法 |
CN112101381A (zh) * | 2020-08-30 | 2020-12-18 | 西南电子技术研究所(中国电子科技集团公司第十研究所) | 张量协作图判别分析遥感图像特征提取方法 |
Family Cites Families (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101520896B (zh) * | 2009-03-30 | 2012-05-30 | 中国电子科技集团公司第十研究所 | 利用光学遥感影像自动检测云层干扰舰船目标的方法 |
CN104778482B (zh) * | 2015-05-05 | 2018-03-13 | 西安电子科技大学 | 基于张量半监督标度切维数约减的高光谱图像分类方法 |
CN105740799B (zh) * | 2016-01-27 | 2018-02-16 | 深圳大学 | 基于三维Gabor特征选择的高光谱遥感图像分类方法及系统 |
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 |
CN108896499A (zh) * | 2018-05-09 | 2018-11-27 | 西安建筑科技大学 | 结合主成分分析与正则化多项式的光谱反射率重建方法 |
CN110232317B (zh) * | 2019-05-05 | 2023-01-03 | 五邑大学 | 基于超像素分割及两阶段分类策略的超光谱图像分类方法、装置及介质 |
CN110334618B (zh) * | 2019-06-21 | 2022-08-05 | 河海大学 | 基于稀疏张量局部Fisher判别分析算法的人体行为识别方法 |
CN111191637A (zh) * | 2020-02-26 | 2020-05-22 | 电子科技大学中山学院 | 一种基于无人机视频采集的人群聚集度检测及呈现方法 |
CN111539316B (zh) * | 2020-04-22 | 2023-05-05 | 中南大学 | 基于双注意力孪生网络的高分遥感影像变化检测方法 |
-
2020
- 2020-08-30 CN CN202010891063.6A patent/CN112101381B/zh active Active
-
2021
- 2021-03-08 US US17/913,855 patent/US20230186606A1/en active Pending
- 2021-03-08 WO PCT/CN2021/079598 patent/WO2022041678A1/zh active Application Filing
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20200026953A1 (en) * | 2018-07-23 | 2020-01-23 | Wuhan University | Method and system of extraction of impervious surface of remote sensing image |
CN110619263A (zh) * | 2019-06-12 | 2019-12-27 | 河海大学 | 低秩联合协同表示的高光谱遥感影像异常探测方法 |
CN111191700A (zh) * | 2019-12-20 | 2020-05-22 | 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 |
CN111368691A (zh) * | 2020-02-28 | 2020-07-03 | 西南电子技术研究所(中国电子科技集团公司第十研究所) | 无监督高光谱遥感图像空谱特征提取方法 |
CN111369457A (zh) * | 2020-02-28 | 2020-07-03 | 西南电子技术研究所(中国电子科技集团公司第十研究所) | 稀疏判别张量鲁棒性pca的遥感图像去噪方法 |
CN112101381A (zh) * | 2020-08-30 | 2020-12-18 | 西南电子技术研究所(中国电子科技集团公司第十研究所) | 张量协作图判别分析遥感图像特征提取方法 |
Cited By (15)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114299398B (zh) * | 2022-03-10 | 2022-05-17 | 湖北大学 | 一种基于自监督对比学习的小样本遥感图像分类方法 |
CN114299398A (zh) * | 2022-03-10 | 2022-04-08 | 湖北大学 | 一种基于自监督对比学习的小样本遥感图像分类方法 |
CN114781576A (zh) * | 2022-04-19 | 2022-07-22 | 广东海洋大学 | 一种基于随机森林算法的声速剖面估计方法及装置 |
CN114897818A (zh) * | 2022-05-09 | 2022-08-12 | 北京理工大学 | 基于时空距离矩阵分析的遥感时间序列图像变化检测方法 |
CN115049944A (zh) * | 2022-06-02 | 2022-09-13 | 北京航空航天大学 | 一种基于多任务优化的小样本遥感图像目标检测方法 |
CN115049944B (zh) * | 2022-06-02 | 2024-05-28 | 北京航空航天大学 | 一种基于多任务优化的小样本遥感图像目标检测方法 |
CN115082770B (zh) * | 2022-07-04 | 2024-02-23 | 青岛科技大学 | 一种基于机器学习的图像中心线结构提取方法 |
CN115082770A (zh) * | 2022-07-04 | 2022-09-20 | 青岛科技大学 | 一种基于机器学习的图像中心线结构提取方法 |
CN116188995A (zh) * | 2023-04-13 | 2023-05-30 | 国家基础地理信息中心 | 一种遥感图像特征提取模型训练方法、检索方法及装置 |
CN116188995B (zh) * | 2023-04-13 | 2023-08-15 | 国家基础地理信息中心 | 一种遥感图像特征提取模型训练方法、检索方法及装置 |
CN116610927B (zh) * | 2023-07-21 | 2023-10-13 | 傲拓科技股份有限公司 | 基于fpga的风机齿轮箱轴承故障诊断方法及诊断模块 |
CN116610927A (zh) * | 2023-07-21 | 2023-08-18 | 傲拓科技股份有限公司 | 基于fpga的风机齿轮箱轴承故障诊断方法及诊断模块 |
CN117173102A (zh) * | 2023-08-04 | 2023-12-05 | 航天恒星科技有限公司 | 基于多源数据融合的遥感数据稳定性分析方法、电子设备及存储介质 |
CN117671673A (zh) * | 2023-11-21 | 2024-03-08 | 江南大学 | 一种基于自适应张量子空间的小样本宫颈细胞分类方法 |
CN117671673B (zh) * | 2023-11-21 | 2024-05-28 | 江南大学 | 一种基于自适应张量子空间的小样本宫颈细胞分类方法 |
Also Published As
Publication number | Publication date |
---|---|
CN112101381A (zh) | 2020-12-18 |
US20230186606A1 (en) | 2023-06-15 |
CN112101381B (zh) | 2022-10-28 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2022041678A1 (zh) | 张量协作图判别分析遥感图像特征提取方法 | |
CN110399909B (zh) | 一种基于标签约束弹性网图模型的高光谱图像分类方法 | |
Krisshna et al. | Face recognition using transform domain feature extraction and PSO-based feature selection | |
CN111860612A (zh) | 无监督高光谱图像隐低秩投影学习特征提取方法 | |
CN108304357B (zh) | 一种基于字体流形的中文字库自动生成方法 | |
CN110659665B (zh) | 一种异维特征的模型构建方法及图像识别方法、装置 | |
CN108647690B (zh) | 基于判别稀疏保持投影的非约束人脸图像降维方法 | |
CN109766858A (zh) | 结合双边滤波的三维卷积神经网络高光谱影像分类方法 | |
CN110458192B (zh) | 基于视觉显著性的高光谱遥感图像分类方法及系统 | |
CN109726725B (zh) | 一种基于大间隔类间互异性多核学习的油画作者识别方法 | |
CN102880875B (zh) | 基于lrr图的半监督学习人脸识别方法 | |
CN109241813B (zh) | 基于判别稀疏保持嵌入的非约束人脸图像降维方法 | |
CN107341505B (zh) | 一种基于图像显著性与Object Bank的场景分类方法 | |
CN109376787B (zh) | 流形学习网络及基于其的计算机视觉图像集分类方法 | |
CN103020265A (zh) | 图像检索的方法和系统 | |
CN104715266B (zh) | 基于src‑dp与lda相结合的图像特征提取方法 | |
CN106778714B (zh) | 基于非线性特征和模型合并的lda人脸识别方法 | |
CN109657567A (zh) | 一种基于3d指纹图像的弱监督的特征分析方法及系统 | |
CN109190511A (zh) | 基于局部与结构约束低秩表示的高光谱分类方法 | |
CN106803105B (zh) | 一种基于稀疏表示字典学习的图像分类方法 | |
CN110414516B (zh) | 一种基于深度学习的单个汉字识别方法 | |
Han et al. | Deep low-rank graph convolutional subspace clustering for hyperspectral image | |
CN114155572A (zh) | 一种面部表情识别方法和系统 | |
Wang et al. | Feature extraction method of face image texture spectrum based on a deep learning algorithm | |
CN105975996A (zh) | 一种基于k均值和奈斯特龙逼近的图像分割方法 |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 21859527 Country of ref document: EP Kind code of ref document: A1 |