WO2021003951A1 - 一种基于标签约束弹性网图模型的高光谱图像分类方法 - Google Patents
一种基于标签约束弹性网图模型的高光谱图像分类方法 Download PDFInfo
- Publication number
- WO2021003951A1 WO2021003951A1 PCT/CN2019/120491 CN2019120491W WO2021003951A1 WO 2021003951 A1 WO2021003951 A1 WO 2021003951A1 CN 2019120491 W CN2019120491 W CN 2019120491W WO 2021003951 A1 WO2021003951 A1 WO 2021003951A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- matrix
- label
- hyperspectral image
- constrained
- image classification
- Prior art date
Links
Images
Classifications
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/28—Determining representative reference patterns, e.g. by averaging or distorting; Generating dictionaries
Definitions
- the invention belongs to the technical field of image information processing, and particularly relates to a hyperspectral image classification method.
- hyperspectral remote sensing technology began to rise and develop rapidly, and the ability of humans to observe and understand things on the surface has undergone a qualitative leap.
- the hyperspectral remote sensing technology can capture the corresponding spectral information while acquiring the space image of the observed ground object. Therefore, the hyperspectral image is presented as a three-dimensional cube data, which realizes the first true map-spectrum imaging.
- the multiple spectral segments of each pixel form a spectral curve, which contains rich information about the composition of things on the surface, and can be used to identify different types of ground features.
- hyperspectral image classification has become a hot research in the field of hyperspectral remote sensing. Due to its own characteristics of hyperspectral images, its classification also faces certain challenges.
- the learning of hyperspectral image classifiers includes supervised models, unsupervised models and semi-supervised classification models.
- Unsupervised methods mainly use clustering algorithms to obtain the distribution characteristics of hyperspectral pixels, but they cannot accurately distinguish their categories. In comparison, the effect of the supervised model is better.
- the semi-supervised classification model was born. It makes full use of the limited labeled samples and a large number of unlabeled samples of hyperspectral images for learning classification. Studies have shown that the semi-supervised classification model requires far less training samples than the supervised samples. Its appearance greatly alleviates the problem of insufficient training samples and effectively improves the classification performance.
- Cheng and Yang proposed a l 1 graph, which uses the sparse representation of the data to construct the graph and cluster the hyperspectral images.
- Song proposed sparse representations (SR+EMAPs) based on morphological attribute contours for hyperspectral classification.
- Zhou and Huang classify hyperspectral images by constructing hypergraphs. Hyperedges can connect multiple vertices at the same time, which is beneficial to express high-order complex relationships.
- Constructing an effective graph representation model is the key to the semi-supervised classification algorithm of hyperspectral graphs.
- the current algorithm does not make full use of the label information to compose the graph.
- How to construct the graph model and use the effective calibration information is the key to the hyperspectral graph semi-supervised classification algorithm.
- the present invention proposes a hyperspectral image classification method based on a label-constrained elastic net graph model.
- the technical solution of the present invention is:
- a hyperspectral image classification method based on label-constrained elastic network graph model including the following steps:
- V represents the space spectrum joint characteristic matrix
- X represents the spectral characteristic matrix
- EMP represents the EMP characteristic matrix
- m is the number of main components
- n is the number of circular structural elements with different radii
- d is the number of spectral bands
- N Is the number of image pixels.
- step (2) the model of label constraint transfer is as follows:
- U is the global constraint matrix
- Z is the initial constraint matrix
- superscript T represents matrix transpose
- tr represents matrix trace
- ⁇ is regularization parameter
- L is regularized Laplacian matrix
- L ID W -1/2 WD W -1 /2
- W is the adjacency matrix obtained by composing the graph based on the space spectrum joint characteristic matrix V
- D W is the degree matrix of the graph
- I is the identity matrix
- each element in the global constraint matrix solved in step (2) corresponds to the confidence between two pixels, and for each pixel x i , the first K confidences are selected
- step (4) the elastic net representation of all pixels is solved by solving the following constraint optimization problem:
- c i is the expression coefficient obtained by x i based on the dictionary D i
- C [c 1 ,c 2 ,...,c N ] is the expression coefficient matrix
- E is the error matrix
- ⁇ and ⁇ are the regularization parameters
- 1 means l 1 norm
- 2,1 means l 2,1 norm;
- a label-constrained elastic net graph model is constructed.
- the coefficients between samples are directly used as the weight of the edge, and the connection of the edges is established between the hyperspectral pixels.
- the adjacency matrix is directly represented by the elastic net
- the coefficient matrix C represents.
- step (5) a semi-supervised learning model is established:
- F is the classification result label matrix
- Y is the initial label matrix
- ⁇ is the regularization parameter
- L 1 is the Laplacian matrix of the label-constrained elastic net graph
- the superscript T represents the matrix transposition
- tr represents the trace of the matrix
- the present invention uses tag pair constraint information to establish a hyperspectral image classification model based on tag-constrained elastic net graphs.
- the model combines spatial features and spectral features to form space-spectrum joint features; and uses tag constraints
- the information selects K samples that are most relevant to it to construct a dictionary, and then decomposes the elastic net and constructs an elastic net graph, which reduces the computational complexity and makes the composition more accurate; semi-supervised classification learning based on the constructed label constraint graph model reaches a high level The purpose of accurate classification of spectral images.
- Figure 1 is an overall flow chart of the present invention.
- the present invention designs a hyperspectral image classification method based on a label-constrained elastic network graph model, as shown in Figure 1, and the steps are as follows:
- S101 Perform EMP feature extraction on the hyperspectral image to be processed, and construct a space-spectrum joint feature
- S102 Perform label constraint transfer according to the space spectrum joint feature to obtain a global constraint matrix
- S105 Perform semi-supervised classification based on the elastic network graph model to obtain a label matrix to realize hyperspectral image classification.
- step S101 is implemented using the following preferred solution:
- V represents the space spectrum joint characteristic matrix
- X represents the spectral characteristic matrix
- EMP represents the EMP characteristic matrix
- m is the number of main components
- n is the number of circular structural elements with different radii
- d is the number of spectral bands
- N Is the number of image pixels.
- step S102 is implemented using the following preferred solution:
- U is the global constraint matrix
- Z is the initial constraint matrix
- superscript T represents matrix transpose
- tr represents matrix trace
- ⁇ is regularization parameter
- L is regularized Laplacian matrix
- L ID W -1/2 WD W -1 /2
- W is the adjacency matrix obtained by composing the graph based on the space spectrum joint characteristic matrix V
- D W is the degree matrix of the graph
- I is the identity matrix.
- step S103 is implemented using the following preferred solution:
- Each element in the global constraint matrix solved in step S102 corresponds to the confidence between two pixels, that is, the element in the i-th row and j-th column in the global constraint matrix represents the confidence between the pixel points x i and x j .
- step S104 is implemented using the following preferred solution:
- c i is the expression coefficient obtained by x i based on the dictionary D i
- C [c 1 ,c 2 ,...,c N ] is the expression coefficient matrix
- E is the error matrix
- ⁇ and ⁇ are the regularization parameters
- 1 means l 1 norm
- 2,1 means l 2,1 norm.
- a label-constrained elastic net graph model is constructed.
- the coefficients between samples are directly used as the weight of the edge, and the connection of the edges is established between the hyperspectral pixels.
- the adjacency matrix is directly represented by the elastic net
- the coefficient matrix C represents.
- step S105 is implemented using the following preferred solution:
- F is the classification result label matrix
- Y is the initial label matrix
- ⁇ is the regularization parameter
- L 1 is the Laplacian matrix of the label-constrained elastic net graph
- the superscript T represents the matrix transposition
- tr represents the trace of the matrix
- the test sequence specifications of Indian Pines (IP) and Salinas Scene (SS) are 145 ⁇ 145 and 512 ⁇ 217, respectively.
- the evaluation of the experiment uses quantitative analysis methods.
- OA is the overall accuracy of all sample classifications (Overall Accuracy)
- AA is the average accuracy of various sample classifications (Average Accuracy)
- k is the Kappa coefficient. The calculations are as follows:
- c is the number of sample categories
- m ii represents the number of samples in the i-th category that are classified into the i-th category in the classification process
- N is the total number of samples
- p i represents the correct rate of classification of each type of sample
- N i represents the The total number of samples of type i.
- C1 ⁇ C16 represent the first to 16 types of hyperspectral images
- LCE+SSL represents the hyperspectral image classification method based on the label-constrained elastic network model proposed by the present invention.
- PX+SVM Use only support vector machines with spectral features for pixel classification
- EMP+SVM Use support vector machine to classify pixels with extended morphological features and spectral features cascade
- Graph+SSL semi-supervised learning based on graph
- SR+EMAPs use the sparse representation of the morphological attribute profile for classification
- GGraph+SSL Semi-supervised classification learning using only Gaussian kernel composition for space spectrum joint features
- LGG+SSL Semi-supervised classification learning of label-constrained Gaussian kernel composition of space spectrum joint features.
- the label information cannot be fully utilized, and the classification result has a large pixel misclassification problem.
- the present invention uses label pair constraint information to establish a hyperspectral image classification model of label-constrained elastic net graphs, selects a plurality of highly correlated sample points to characterize the target sample, suppresses sample points with high probability of belonging to different classes, and improves The accuracy of the composition is improved.
- the algorithm in this paper has a greater inhibitory effect on pixel misclassification, and can more effectively classify hyperspectral pixels. It shows certain advantages in terms of classification accuracy and visual effects.
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
一种基于标签约束弹性网图模型的高光谱图像分类方法。步骤:对待高光谱图像进行EMP特征提取,构建空谱联合特征(S101);根据空谱联合特征进行标签约束传递,获得全局约束矩阵(S102);根据全局约束矩阵,针对每个像素点构建字典(S103);根据字典求解弹性网表示,构建标签约束的弹性网图模型(S104);基于弹性网图模型进行半监督分类,得到标签矩阵,实现高光谱图像分类(S105)。该方法可以降低计算的复杂度,提高构图的准确性,提升算法分类性能。
Description
本发明属于图像信息处理技术领域,特别涉及了一种高光谱图像分类方法。
上世纪80年代,高光谱遥感技术开始崛起并且迅速发展,人类观测与认识地表事物的能力发生了质的飞跃。高光谱遥感技术在获取观测地物的空间图像的同时,还可以捕捉到其对应的光谱信息,因此高光谱图像呈现为一个三维立方体数据,实现了第一次真正意义上的图谱合一成像。每个像素的多个谱段形成了光谱曲线,包含了地表事物成分的丰富信息,可以用于识别不同的地物类型。目前,高光谱图像分类已经成为当前高光谱遥感领域的热门研究。由于高光谱图像的自身特征,其分类也面临着一定的挑战。
在高光谱分类的特征表示方面,研究者们进行了很多尝试。最初的方法通常直接使用光谱信息,较难得到稳健的分类结果。在认识到空间信息在高光谱分类中的重要性后,Pesaresi和Benediktsson两人采用形态学转换的方法去构建形态学分布特征,用来提取空间结构信息。考虑到高光谱图像空谱合一的特性,Fauvel和Chanussot将可拓展的形态学空间特征EMP与光谱特征进行联合提取到一种空谱联合的特征表示,然后采用支持向量机模型进行分类,实现了算法分类的性能提升。
高光谱图像分类器学习方面包括监督模型、非监督模型与半监督分类模型,区别在于在模型训练阶段是否使用到了带有标签的数据样本。非监督方法主要使用聚类算法来获得高光谱像素的分布特性,但无法准确判别其类别。相比较而言,监督模型的效果更好,想要训练出好的分类模型,训练样本的数量必须是样本维 度的十倍,但是想要获得大量的训练样本需要付出巨大的成本代价。半监督分类模型因此而诞生,它充分利用了高光谱图像有限的有标记样本和大量的未标记样本进行学习分类。研究表明,半监督分类模型对于训练样本的数量要求远低于监督样本的要求,它的出现大大缓解了训练样本不足的问题,并且有效的提高了分类性能。
诸多文献研究了高光谱图像的半监督分类算法,而基于支持向量机分类器则是目前主流的半监督方法之一。Leng和Xu提出了主动学习与半监督分类相结合用于SVM分类器,改善仅使用SVM进行分类的性能。另一种主流的半监督分类方法则是基于图的半监督学习方法,它通过边将每个样本的标签信息传播给近邻,直到整个数据集达成全局稳定,该方法的优势体现在计算速度快、准确率高、便于解决大规模问题。例如,Camps-Valls等人利用k近邻法(k-nearest neighbors,KNN)构造了一个半监督图,并利用复合核计算每条边的权值。但是KNN的方法泛化能力往往不是很好,很有可能发生过拟合的情况。为了进一步整合空间上下文信息,Cheng和Yang提出一种l
1的图,利用数据的稀疏表示去构建图,进行高光谱图像的聚类。Song提出了基于形态属性轮廓的稀疏表示(SR+EMAPs)用于高光谱分类。Zhou和Huang通过构建超图来进行高光谱图像的分类,超边可以同时连接多个顶点,有利于表示高阶的复杂关系。
构建有效的图表示模型是高光谱图半监督分类算法的关键。然而当前的算法并未能充分利用标签信息进行构图,如何构造图模型以及利用有效的标定信息是高光谱的图半监督分类算法的关键。
发明内容
为了解决上述背景技术提到的技术问题,本发明提出了一种基于标签约束弹性网图模型的高光谱图像分类方法。
为了实现上述技术目的,本发明的技术方案为:
一种基于标签约束弹性网图模型的高光谱图像分类方法,包括以下步骤:
(1)对待高光谱图像进行EMP特征提取,构建空谱联合特征;
(2)根据空谱联合特征进行标签约束传递,获得全局约束矩阵;
(3)根据全局约束矩阵,针对每个像素点构建字典;
(4)根据字典求解弹性网表示,构建标签约束的弹性网图模型;
(5)基于弹性网图模型进行半监督分类,得到标签矩阵,实现高光谱图像分类。
进一步地,在步骤(2)中,标签约束传递的模型如下:
其中,U为全局约束矩阵,Z为初始约束矩阵,
表示F范数的平方,上标T表示矩阵转置,tr表示矩阵的迹,δ为正则化参数,L为正则化的拉普拉斯矩阵,L=I-D
W
-1/2WD
W
-1/2,W是根据空谱联合特征矩阵V构图得到的邻接矩阵,D
W是图的度矩阵,I是单位阵;
求解上述模型,得到全局约束矩阵的最优解U
*。
进一步地,在步骤(3)中,步骤(2)求解的全局约束矩阵中的每个元素为对应两个像素点之间的置信度,针对每个像素点x
i,挑选出前K个置信度最高的像素点x
j构建字典D
i=VB
i,其中,B
i是一个对角矩阵,若x
j被选出构建字 典D
i,则(B
i)
jj=1,否则(B
i)
jj=0,(B
i)
jj表示矩阵B
i中第j行第j列的元素,i=1,2,…,N,j∈{1,2,...,N}。
进一步地,在步骤(4)中,通过求解以下约束优化问题求解所有像素点的弹性网表示:
s.t.x
i=D
ic
i,1≤i≤N
其中,c
i是x
i基于字典D
i获得的表示系数,C=[c
1,c
2,...,c
N]是表示系数矩阵,E是误差矩阵,λ和γ是正则化参数,||C||
1表示l
1范数,||E||
2,1表示l
2,1范数;
根据每个样本点的弹性网表示系数,构建标签约束的弹性网图模型,样本之间的系数直接作为边的权重,在高光谱像素点之间建立边的连接,邻接矩阵直接用弹性网表示系数矩阵C表示。
进一步地,在步骤(5)中,建立半监督学习模型:
求解半监督学习模型,得到分类结果标签矩阵的最优解F
*,实现高光谱图像分类。
采用上述技术方案带来的有益效果:
本发明利用标签成对约束信息,建立一种基于标签约束弹性网图的高光谱图像分类模型,模型在特征表示上,将空间特征与光谱特征相联合构成空谱联合特征;并且利用了标签约束信息选择K个与其最相关的样本构建字典,进而进行 弹性网分解并构建弹性网图,降低了计算复杂度,使得构图更加准确;基于构建的标签约束图模型进行半监督分类学习,达到了高光谱图像精准分类的目的。
图1是本发明的整体流程图。
以下将结合附图,对本发明的技术方案进行详细说明。
本发明设计了基于标签约束弹性网图模型的高光谱图像分类方法,如图1所示,步骤如下:
S101、对待高光谱图像进行EMP特征提取,构建空谱联合特征;
S102、根据空谱联合特征进行标签约束传递,获得全局约束矩阵;
S103、根据全局约束矩阵,针对每个像素点构建字典;
S104、根据字典求解弹性网表示,构建标签约束的弹性网图模型;
S105、基于弹性网图模型进行半监督分类,得到标签矩阵,实现高光谱图像分类。
在本实施例中,步骤S101采用如下优选方案实现:
其中,V表示空谱联合特征矩阵,X表示光谱特征矩阵,EMP表示EMP特征矩阵,m为主成分个数,n为半径不同的圆形结构元素的个数,d为光谱波段的数量,N为图像像素点数目。
在本实施例中,步骤S102采用如下优选方案实现:
标签约束传递的模型如下:
其中,U为全局约束矩阵,Z为初始约束矩阵,
表示F范数的平方,上标T表示矩阵转置,tr表示矩阵的迹,δ为正则化参数,L为正则化的拉普拉斯矩阵,L=I-D
W
-1/2WD
W
-1/2,W是根据空谱联合特征矩阵V构图得到的邻接矩阵,D
W是图的度矩阵,I是单位阵。
在本实施例中,步骤S103采用如下优选方案实现:
步骤S102求解的全局约束矩阵中的每个元素为对应两个像素点之间的置信度,即全局约束矩阵中第i行第j列元素代表像素点x
i与x
j之间的置信度。针对每个像素点x
i,挑选出前K个置信度最高的像素点x
j构建字典D
i=VB
i,其中,B
i是一个对角矩阵,若x
j被选出构建字典D
i,则(B
i)
jj=1,否则(B
i)
jj=0,(B
i)
jj表示矩阵B
i中第j行第j列的元素,i=1,2,…,N,j∈{1,2,...,N}。
在本实施例中,步骤S104采用如下优选方案实现:
通过求解以下约束优化问题求解所有像素点的弹性网表示:
s.t.x
i=D
ic
i,1≤i≤N
其中,c
i是x
i基于字典D
i获得的表示系数,C=[c
1,c
2,...,c
N]是表示系数矩阵,E是误差矩阵,λ和γ是正则化参数,||C||
1表示l
1范数,||E||
2,1表示l
2,1范数。
根据每个样本点的弹性网表示系数,构建标签约束的弹性网图模型,样本之间的系数直接作为边的权重,在高光谱像素点之间建立边的连接,邻接矩阵直接用弹性网表示系数矩阵C表示。
在本实施例中,步骤S105采用如下优选方案实现:
建立半监督学习模型:
求解半监督学习模型,得到分类结果标签矩阵的最优解F
*=(I-βL
1)
-1Y,β是正则化参数,实现高光谱图像分类。
为验证本发明的效果,对本发明进行仿真实验,Indian Pines(IP)和Salinas Scene(SS)测试序列规格分别是145×145和512×217,两种测试序列设置正则化参数α和β:α
IP=0.3,β
IP=0.1,α
SS=0.1,β
SS=0.5。
实验的评价使用定量的分析方法。
关于定量比较分析,使用OA、AA和k来评判。其中,OA为所有样本分类的总体准确率(Overall Accuracy),AA为各类样本分类的平均准确度(Average Accuracy),k为Kappa系数,其计算分别如下:
其中,c为样本类别数量,m
ii表示分类过程中第i类的样本被分到第i类的样本 数量,N为样本总量,p
i表示每类样本分类的正确率,N
i表示第i类样本的总数。
在作定量比较时,我们在两个高光谱图像数据集上,每类样本选取10%的样本给予标签,进行分类,将每个数据集的分类结果与Groundtruth比对,计算出对应的OA、AA和k值。表2和表3给出了本发明与其他方法分别在数据集Indian Pines(IP)和Salinas Scene(SS)OA、AA和k值,可见本发明在高光谱图像的分类中起到了很大的效果,使用标签约束信息,使得构图更加准确,分类精度有所提升。
表1
Class | PX+SVM | EMP+SVM | Graph+SSL | SR+EMAPs | GGraph+SSL | LCG+SSL | LCE+SSL |
C1 | 38.05 | 73.91 | 20.87 | 91.95 | 84.78 | 89.13 | 86.96 |
C2 | 78.72 | 87.82 | 71.05 | 82.51 | 89.78 | 90.83 | 95.17 |
C3 | 68.70 | 86.51 | 51.52 | 89.06 | 84.94 | 79.76 | 88.80 |
C4 | 52.96 | 83.97 | 48.95 | 63.24 | 80.59 | 84.39 | 80.59 |
C5 | 90.11 | 90.64 | 91.39 | 93.62 | 96.89 | 93.58 | 93.37 |
C6 | 96.44 | 98.44 | 93.51 | 99.27 | 98.36 | 99.18 | 98.49 |
C7 | 66.40 | 90.00 | 89.29 | 97.20 | 32.14 | 89.29 | 96.43 |
C8 | 96.51 | 98.95 | 100 | 100 | 98.74 | 99.16 | 99.58 |
C9 | 38.89 | 53.00 | 18.00 | 27.78 | 10.00 | 95.00 | 60.00 |
C10 | 71.20 | 85.49 | 77.53 | 84.49 | 93.83 | 95.78 | 94.75 |
C11 | 78.62 | 87.98 | 83.60 | 94.83 | 93.73 | 94.01 | 94.42 |
C12 | 70.79 | 84.25 | 39.87 | 89.27 | 78.08 | 82.80 | 88.03 |
C13 | 97.28 | 98.63 | 99.41 | 98.21 | 98.54 | 99.51 | 98.05 |
C14 | 90.97 | 96.27 | 98.91 | 99.70 | 99.37 | 98.34 | 99.84 |
C15 | 58.27 | 90.26 | 49.64 | 97.61 | 66.84 | 77.46 | 90.93 |
C16 | 87.38 | 99.78 | 88.17 | 96.02 | 95.70 | 98.92 | 93.55 |
OA | 79.65 | 89.98 | 77.76 | 91.76 | 91.40 | 92.33 | 94.40 |
AA | 73.83 | 87.78 | 70.11 | 87.80 | 81.39 | 91.70 | 91.19 |
k | 76.76 | 88.56 | 74.51 | 90.88 | 90.20 | 91.26 | 93.61 |
表2
Class | PX+SVM | EMP+SVM | Graph+SSL | SR+EMAPs | GGraph+SSL | LCG+SSL | LCE+SSL |
C1 | 99.37 | 99.93 | 99.89 | 100 | 99.45 | 99.95 | 98.95 |
C2 | 99.92 | 99.83 | 96.34 | 99.94 | 99.97 | 99.95 | 99.97 |
C3 | 99.66 | 99.78 | 94.10 | 99.93 | 100 | 99.75 | 100 |
C4 | 99.27 | 99.00 | 99.17 | 99.44 | 98.92 | 99.64 | 99.57 |
C5 | 98.83 | 99.21 | 98.43 | 98.51 | 98.28 | 98.58 | 97.91 |
C6 | 99.90 | 99.94 | 99.93 | 99.86 | 99.80 | 99.97 | 99.77 |
C7 | 99.71 | 99.75 | 99.96 | 99.80 | 100 | 99.89 | 99.41 |
C8 | 87.54 | 94.36 | 96.48 | 99.04 | 98.45 | 96.42 | 99.02 |
C9 | 99.60 | 99.95 | 99.57 | 99.93 | 99.97 | 100 | 100 |
C10 | 96.71 | 99.19 | 98.77 | 98.98 | 98.38 | 98.32 | 99.51 |
C11 | 97.96 | 99.57 | 96.42 | 98.82 | 93.63 | 96.16 | 98.78 |
C12 | 99.79 | 99.98 | 95.23 | 100 | 99.79 | 100 | 99.64 |
C13 | 98.96 | 99.48 | 94.67 | 99.23 | 98.58 | 98.80 | 98.47 |
C14 | 96.68 | 99.66 | 94.15 | 97.89 | 96.45 | 97.38 | 97.85 |
C15 | 76.51 | 88.31 | 92.61 | 86.97 | 95.12 | 95.01 | 97.04 |
C16 | 99.00 | 99.61 | 99.81 | 100 | 100 | 100 | 100 |
OA | 93.71 | 97.05 | 97.12 | 97.79 | 98.54 | 98.23 | 99.05 |
AA | 96.86 | 98.60 | 97.22 | 98.65 | 98.55 | 98.74 | 99.12 |
k | 92.99 | 96.72 | 96.79 | 97.12 | 98.38 | 98.03 | 98.94 |
在上表中,C1~C16表示高光谱图像中的第1~16类,LCE+SSL代表本发明提出的基于标签约束弹性网图模型的高光谱图像分类方法,其他方法解释如下:PX+SVM:只使用光谱特征的支持向量机进行像素分类;
EMP+SVM:采用支持向量机对扩展形态学特征和光谱特征级联的像素进行分类;
Graph+SSL:基于图的半监督学习;
SR+EMAPs:利用形态学属性剖面的稀疏表示进行分类;
GGraph+SSL:对空谱联合特征仅使用高斯核构图的半监督分类学习;
LGG+SSL:空谱联合特征的标签约束高斯核构图的半监督分类学习。
综上可知,相对于传统高光谱分类算法并不能够充分利用标签信息,分类的结果存在很大的像素错分问题。本发明利用标签成对约束信息,建立一种标签约束弹性网图的高光谱图像分类模型,选取多个高度相关的样本点来表征目标样本,抑制了大概率属于不同类的样本点,进而提高了构图的精准性。本文的算法对于像素错分有了较大的抑制作用,可以更有效的将高光谱像素进行分类,无论是从分类的准确率还是从视觉效果上来看,都呈现出一定的优势。
实施例仅为说明本发明的技术思想,不能以此限定本发明的保护范围,凡是按照本发明提出的技术思想,在技术方案基础上所做的任何改动,均落入本发明保护范围之内。
Claims (6)
- 一种基于标签约束弹性网图模型的高光谱图像分类方法,其特征在于,包括以下步骤:(1)对待高光谱图像进行EMP特征提取,构建空谱联合特征;(2)根据空谱联合特征进行标签约束传递,获得全局约束矩阵;(3)根据全局约束矩阵,针对每个像素点构建字典;(4)根据字典求解弹性网表示,构建标签约束的弹性网图模型;(5)基于弹性网图模型进行半监督分类,得到标签矩阵,实现高光谱图像分类。
- 根据权利要求2所述基于标签约束弹性网图模型的高光谱图像分类方法, 其特征在于,在步骤(3)中,步骤(2)求解的全局约束矩阵中的每个元素为对应两个像素点之间的置信度,针对每个像素点x i,挑选出前K个置信度最高的像素点x j构建字典D i=VB i,其中,B i是一个对角矩阵,若x j被选出构建字典D i,则(B i) jj=1,否则(B i) jj=0,(B i) jj表示矩阵B i中第j行第j列的元素,i=1,2,…,N,j∈{1,2,...,N}。
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910609358.7A CN110399909B (zh) | 2019-07-08 | 2019-07-08 | 一种基于标签约束弹性网图模型的高光谱图像分类方法 |
CN201910609358.7 | 2019-07-08 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2021003951A1 true WO2021003951A1 (zh) | 2021-01-14 |
Family
ID=68322833
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/CN2019/120491 WO2021003951A1 (zh) | 2019-07-08 | 2019-11-25 | 一种基于标签约束弹性网图模型的高光谱图像分类方法 |
Country Status (2)
Country | Link |
---|---|
CN (1) | CN110399909B (zh) |
WO (1) | WO2021003951A1 (zh) |
Cited By (16)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112801028A (zh) * | 2021-02-09 | 2021-05-14 | 北京工业大学 | 基于归纳表示学习网络的光谱和空间图高光谱图像分类方法 |
CN113111834A (zh) * | 2021-04-23 | 2021-07-13 | 大连海事大学 | 一种基于注意力的全局局部光谱权重的波段选择方法 |
CN113139513A (zh) * | 2021-05-14 | 2021-07-20 | 辽宁工程技术大学 | 基于超像素轮廓和改进pso-elm的空谱主动学习高光谱分类方法 |
CN113191287A (zh) * | 2021-05-08 | 2021-07-30 | 大连海事大学 | 一种面向分类的高光谱图像波段选择方法 |
CN113688697A (zh) * | 2021-08-06 | 2021-11-23 | 南京审计大学 | 一种基于局部相似性保持特征表示的掌纹识别方法 |
CN113807438A (zh) * | 2021-09-17 | 2021-12-17 | 中国石油大学(华东) | 基于语义保留的对抗混合半监督高光谱图像分类方法 |
CN113836996A (zh) * | 2021-08-10 | 2021-12-24 | 中国地质大学(武汉) | 高光谱遥感图像部分迁移方法及*** |
CN113837258A (zh) * | 2021-09-17 | 2021-12-24 | 华中师范大学 | 基于局部相关熵矩阵的高光谱图像分类方法和*** |
CN114332986A (zh) * | 2021-12-07 | 2022-04-12 | 北京理工大学 | 一种基于子空间学习的小样本人脸识别方法 |
CN114863291A (zh) * | 2022-04-20 | 2022-08-05 | 重庆市地理信息和遥感应用中心 | 基于mcl和光谱差异度量的高光谱影像波段选择方法 |
CN115019368A (zh) * | 2022-06-09 | 2022-09-06 | 南京审计大学 | 一种基于2desdlpp的审计调查中人脸识别特征提取方法 |
CN116310459A (zh) * | 2023-03-28 | 2023-06-23 | 中国地质大学(武汉) | 基于多视图空谱联合的高光谱图像子空间聚类方法 |
CN116611001A (zh) * | 2023-07-19 | 2023-08-18 | 中国海洋大学 | 基于多维自适应增量图的近红外光谱数据分类方法 |
CN116755397A (zh) * | 2023-05-26 | 2023-09-15 | 北京航空航天大学 | 一种基于图卷积策略梯度的多机协同任务调度方法 |
CN116824282A (zh) * | 2023-08-31 | 2023-09-29 | 中国石油大学(华东) | 融合多跳图卷积和注意力的半监督高光谱图像分类方法 |
CN117173350A (zh) * | 2023-08-09 | 2023-12-05 | 中国科学技术大学 | 基于主动领域适应学习的地质建模方法、***及介质 |
Families Citing this family (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110399909B (zh) * | 2019-07-08 | 2021-07-27 | 南京信息工程大学 | 一种基于标签约束弹性网图模型的高光谱图像分类方法 |
CN112329818B (zh) * | 2020-10-20 | 2023-07-07 | 南京信息工程大学 | 基于图卷积网络嵌入表征的高光谱图像非监督分类方法 |
CN112381144B (zh) * | 2020-11-13 | 2022-09-13 | 南京理工大学 | 非欧与欧氏域空谱特征学习的异构深度网络方法 |
CN112417188B (zh) * | 2020-12-10 | 2022-05-24 | 桂林电子科技大学 | 一种基于图模型的高光谱图像分类方法 |
CN113111969B (zh) * | 2021-05-03 | 2022-05-06 | 齐齐哈尔大学 | 一种基于混合度量的高光谱图像分类方法 |
CN113821664A (zh) * | 2021-08-30 | 2021-12-21 | 湖南军芃科技股份有限公司 | 一种基于依据直方统计频率的图像分类方法、***、终端及可读存储介质 |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20140204092A1 (en) * | 2012-04-09 | 2014-07-24 | The Regents Of The University Of California | Classification of high dimensional data |
CN105787516A (zh) * | 2016-03-09 | 2016-07-20 | 南京信息工程大学 | 一种基于空谱局部性低秩超图学习的高光谱图像分类方法 |
CN110399909A (zh) * | 2019-07-08 | 2019-11-01 | 南京信息工程大学 | 一种基于标签约束弹性网图模型的高光谱图像分类方法 |
Family Cites Families (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US8194938B2 (en) * | 2009-06-02 | 2012-06-05 | George Mason Intellectual Properties, Inc. | Face authentication using recognition-by-parts, boosting, and transduction |
CN104866871B (zh) * | 2015-06-02 | 2018-06-05 | 西安电子科技大学 | 基于投影结构稀疏编码的高光谱图像分类方法 |
CN109886887A (zh) * | 2019-01-28 | 2019-06-14 | 深圳大学 | 一种壁画图像修复处理方法、***、装置及存储介质 |
-
2019
- 2019-07-08 CN CN201910609358.7A patent/CN110399909B/zh active Active
- 2019-11-25 WO PCT/CN2019/120491 patent/WO2021003951A1/zh active Application Filing
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20140204092A1 (en) * | 2012-04-09 | 2014-07-24 | The Regents Of The University Of California | Classification of high dimensional data |
CN105787516A (zh) * | 2016-03-09 | 2016-07-20 | 南京信息工程大学 | 一种基于空谱局部性低秩超图学习的高光谱图像分类方法 |
CN110399909A (zh) * | 2019-07-08 | 2019-11-01 | 南京信息工程大学 | 一种基于标签约束弹性网图模型的高光谱图像分类方法 |
Non-Patent Citations (2)
Title |
---|
LIU, QINGSHAN ET AL.: "Elastic Net Hypergraph Learning for Image Clustering and Semi-Supervised Classification", IEEE TRANSACTIONS ON IMAGE PROCESSING,, vol. 26, no. 1,, 31 January 2017 (2017-01-31), XP080686975, ISSN: 1057-7149, DOI: 20200313145741Y * |
SUN, YUBAO ET AL.: "Hypergraph Embedding for Spatial-Spectral Joint Feature Extraction in Hyperspectral Images", REMOTE SENSING,, vol. 9, no. 5, 22 May 2017 (2017-05-22), XP055772535, ISSN: 2072-4292 * |
Cited By (28)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112801028A (zh) * | 2021-02-09 | 2021-05-14 | 北京工业大学 | 基于归纳表示学习网络的光谱和空间图高光谱图像分类方法 |
CN113111834A (zh) * | 2021-04-23 | 2021-07-13 | 大连海事大学 | 一种基于注意力的全局局部光谱权重的波段选择方法 |
CN113111834B (zh) * | 2021-04-23 | 2024-04-05 | 大连海事大学 | 一种基于注意力的全局局部光谱权重的波段选择方法 |
CN113191287A (zh) * | 2021-05-08 | 2021-07-30 | 大连海事大学 | 一种面向分类的高光谱图像波段选择方法 |
CN113191287B (zh) * | 2021-05-08 | 2023-12-26 | 大连海事大学 | 一种面向分类的高光谱图像波段选择方法 |
CN113139513B (zh) * | 2021-05-14 | 2023-11-10 | 辽宁工程技术大学 | 基于超像素轮廓和改进pso-elm的空谱主动学习高光谱分类方法 |
CN113139513A (zh) * | 2021-05-14 | 2021-07-20 | 辽宁工程技术大学 | 基于超像素轮廓和改进pso-elm的空谱主动学习高光谱分类方法 |
CN113688697A (zh) * | 2021-08-06 | 2021-11-23 | 南京审计大学 | 一种基于局部相似性保持特征表示的掌纹识别方法 |
CN113836996A (zh) * | 2021-08-10 | 2021-12-24 | 中国地质大学(武汉) | 高光谱遥感图像部分迁移方法及*** |
CN113836996B (zh) * | 2021-08-10 | 2024-02-02 | 中国地质大学(武汉) | 高光谱遥感图像部分迁移方法及*** |
CN113837258A (zh) * | 2021-09-17 | 2021-12-24 | 华中师范大学 | 基于局部相关熵矩阵的高光谱图像分类方法和*** |
CN113837258B (zh) * | 2021-09-17 | 2023-09-08 | 华中师范大学 | 基于局部相关熵矩阵的高光谱图像分类方法和*** |
CN113807438A (zh) * | 2021-09-17 | 2021-12-17 | 中国石油大学(华东) | 基于语义保留的对抗混合半监督高光谱图像分类方法 |
CN113807438B (zh) * | 2021-09-17 | 2023-08-29 | 中国石油大学(华东) | 基于语义保留的对抗混合半监督高光谱图像分类方法 |
CN114332986A (zh) * | 2021-12-07 | 2022-04-12 | 北京理工大学 | 一种基于子空间学习的小样本人脸识别方法 |
CN114863291A (zh) * | 2022-04-20 | 2022-08-05 | 重庆市地理信息和遥感应用中心 | 基于mcl和光谱差异度量的高光谱影像波段选择方法 |
CN114863291B (zh) * | 2022-04-20 | 2023-08-08 | 重庆市地理信息和遥感应用中心 | 基于mcl和光谱差异度量的高光谱影像波段选择方法 |
CN115019368B (zh) * | 2022-06-09 | 2023-09-12 | 南京审计大学 | 一种审计调查中人脸识别特征提取方法 |
CN115019368A (zh) * | 2022-06-09 | 2022-09-06 | 南京审计大学 | 一种基于2desdlpp的审计调查中人脸识别特征提取方法 |
CN116310459A (zh) * | 2023-03-28 | 2023-06-23 | 中国地质大学(武汉) | 基于多视图空谱联合的高光谱图像子空间聚类方法 |
CN116310459B (zh) * | 2023-03-28 | 2024-01-19 | 中国地质大学(武汉) | 基于多视图空谱联合的高光谱图像子空间聚类方法 |
CN116755397A (zh) * | 2023-05-26 | 2023-09-15 | 北京航空航天大学 | 一种基于图卷积策略梯度的多机协同任务调度方法 |
CN116755397B (zh) * | 2023-05-26 | 2024-01-23 | 北京航空航天大学 | 一种基于图卷积策略梯度的多机协同任务调度方法 |
CN116611001B (zh) * | 2023-07-19 | 2023-10-03 | 中国海洋大学 | 基于多维自适应增量图的近红外光谱数据分类方法 |
CN116611001A (zh) * | 2023-07-19 | 2023-08-18 | 中国海洋大学 | 基于多维自适应增量图的近红外光谱数据分类方法 |
CN117173350A (zh) * | 2023-08-09 | 2023-12-05 | 中国科学技术大学 | 基于主动领域适应学习的地质建模方法、***及介质 |
CN116824282A (zh) * | 2023-08-31 | 2023-09-29 | 中国石油大学(华东) | 融合多跳图卷积和注意力的半监督高光谱图像分类方法 |
CN116824282B (zh) * | 2023-08-31 | 2023-11-10 | 中国石油大学(华东) | 融合多跳图卷积和注意力的半监督高光谱图像分类方法 |
Also Published As
Publication number | Publication date |
---|---|
CN110399909A (zh) | 2019-11-01 |
CN110399909B (zh) | 2021-07-27 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2021003951A1 (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 | |
WO2022001159A1 (zh) | 无监督高光谱图像隐低秩投影学习特征提取方法 | |
Ding et al. | Semi-supervised locality preserving dense graph neural network with ARMA filters and context-aware learning for hyperspectral image classification | |
Wang et al. | Beyond low-rank representations: Orthogonal clustering basis reconstruction with optimized graph structure for multi-view spectral clustering | |
CN106203523B (zh) | 基于梯度提升决策树半监督算法融合的高光谱图像分类方法 | |
CN104751191B (zh) | 一种稀疏自适应半监督多流形学习的高光谱影像分类方法 | |
WO2018010434A1 (zh) | 一种图像分类方法及装置 | |
CN107563442B (zh) | 基于稀疏低秩正则图张量化嵌入的高光谱图像分类方法 | |
CN111695636B (zh) | 一种基于图神经网络的高光谱图像分类方法 | |
Liu et al. | Hyperspectral image classification with kernel-based least-squares support vector machines in sum space | |
CN109190511B (zh) | 基于局部与结构约束低秩表示的高光谱分类方法 | |
CN104077612B (zh) | 一种基于多特征稀疏表示技术的害虫图像识别方法 | |
WO2022228348A1 (zh) | 基于高阶关联保持的缺失多视图子空间聚类方法及*** | |
CN109241816B (zh) | 一种基于标签优化的图像再识别***及损失函数确定方法 | |
Chu et al. | Hyperspectral image classification with discriminative manifold broad learning system | |
Levin et al. | Out-of-sample extension of graph adjacency spectral embedding | |
CN111460966B (zh) | 基于度量学习和近邻增强的高光谱遥感图像分类方法 | |
Guo et al. | Dual graph U-Nets for hyperspectral image classification | |
Han et al. | Deep low-rank graph convolutional subspace clustering for hyperspectral image | |
CN114758224A (zh) | 一种基于深度学习的垃圾分类检测方法 | |
Ren et al. | Research on infrared small target segmentation algorithm based on improved mask R-CNN | |
CN105023239B (zh) | 基于超像素和最大边界分布的高光谱数据降维方法 | |
CN110070485A (zh) | 一种高光谱图像降维方法 | |
CN111639686B (zh) | 基于维度加权和视角特征一致性的半监督分类方法 | |
CN110188821B (zh) | 多参数边缘保持滤波和多特征学习的高光谱图像分类方法 |
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: 19936663 Country of ref document: EP Kind code of ref document: A1 |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 19936663 Country of ref document: EP Kind code of ref document: A1 |