CN116778339B - Method and system for selecting hyperspectral wave bands by aid of local view auxiliary discrimination - Google Patents
Method and system for selecting hyperspectral wave bands by aid of local view auxiliary discriminationInfo
- Publication number
- CN116778339B CN116778339B CN202310887160.1A CN202310887160A CN116778339B CN 116778339 B CN116778339 B CN 116778339B CN 202310887160 A CN202310887160 A CN 202310887160A CN 116778339 B CN116778339 B CN 116778339B
- Authority
- CN
- China
- Prior art keywords
- band
- hyperspectral
- matrix
- hypergraph
- super
- Prior art date
- Legal status (The legal status 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 status listed.)
- Active
Links
Landscapes
- Image Processing (AREA)
- Image Analysis (AREA)
Abstract
本发明公开了一种局部视图辅助判别高光谱波段选择方法及系统。本发明首先运用超像素分割技术ERS对高光谱图像进行分割，形成一系列超像素块来捕捉不同地物分布的光谱特征差异；其次本发明针对每一超像素块，将波段视为超图顶点，构造对应超图来表达波段间的多元邻接关系，尽可能合理化波段邻接结构以指导稀疏自表示模型的优化，降低波段子集的冗余；其中共识矩阵通过迭代更新融合了各超像素块的系数矩阵；最后根据共识矩阵计算各波段重构原始数据的重要度，选择波段子集，实现用统一波段子集表征高光谱图像的目的；本发明充分捕捉了异质区域的光谱特征差异，可切实增强模型的像元空间局部约束，提高波段子集质量，从而提升后续分类的精确度。
This invention discloses a method and system for selecting hyperspectral bands with local view-assisted discrimination. First, the invention uses superpixel segmentation (ERS) technology to segment the hyperspectral image, forming a series of superpixel blocks to capture the spectral feature differences of different land cover distributions. Second, for each superpixel block, the invention treats bands as vertices of a hypergraph, constructing a corresponding hypergraph to express the multivariate adjacency relationships between bands, rationalizing the band adjacency structure as much as possible to guide the optimization of the sparse self-representation model and reduce the redundancy of band subsets. The consensus matrix iteratively updates and merges the coefficient matrices of each superpixel block. Finally, based on the consensus matrix, the importance of each band in reconstructing the original data is calculated, and a band subset is selected to achieve the goal of representing the hyperspectral image with a unified band subset. This invention fully captures the spectral feature differences of heterogeneous regions, effectively enhances the local constraints of the model's pixel space, improves the quality of band subsets, and thus improves the accuracy of subsequent classification.
Description
Technical Field
The invention belongs to the technical field of hyperspectral image dimension reduction, and particularly relates to a hyperspectral wave band selection method and a hyperspectral wave band selection system for auxiliary discrimination of a local view.
Background
The hyperspectral remote sensing image has hundreds of adjacent and long spectrum channels, can provide higher spectrum resolution than the RGB image, has rich spectrum and space information, and is more beneficial to the accurate identification of ground features. However, the nano-scale spectrum resolution of the hyperspectral image also causes problems for data processing, such as high computational complexity and information redundancy, which results in waste of storage space.
In order to solve these problems, it is necessary to perform a dimension-reduction preprocessing on the hyperspectral image. Typical dimension reduction methods have feature extraction and band selection. The feature extraction projects the original high-dimensional data into a low-dimensional space, and changes the physical properties of the original data, so that some key information is destroyed. The band selection is to select the band subset with the most discrimination from the original data, and compared with the feature extraction, the band selection can better keep the information of the original hyperspectral data without changing the physical characteristics of the band, so that the data after the dimension reduction has higher interpretability and usability.
Band selection techniques can be broadly divided into two categories, supervised and unsupervised. The supervision band selection requires a certain priori information, such as training samples and corresponding labels, and the difficulty and the cost of obtaining the labels of the labels prevent the development of the supervision band selection to a certain extent. In contrast, the unsupervised band selection method does not need explicit labels, only uses unlabeled data to develop a learning model, provides a feasible solution for a plurality of band selection methods plagued by labels, and is more convenient to apply.
In recent years, application of the sparse representation theory to the hyperspectral field is proved to be reasonable, the interpretability of the model is improved based on sparse hyperspectral band selection, the redundancy phenomenon of data is greatly reduced, meanwhile, the storage efficiency is improved, and unnecessary resource waste is avoided. However, some existing researches only expand functions of the sparse self-expression model to a certain extent, influence of pixel space information on key feature extraction is ignored in the process of band selection, and a real multi-element adjacent structure cannot be accurately expressed when a band relation is described, so that the quality of a final selected band subset is low.
Disclosure of Invention
Aiming at the problem that the traditional sparse self-representation band selection method has insufficient local constraint on pixel space in the band selection process, the invention provides a local view assisted discrimination hyperspectral band selection method so as to fully capture spectrum characteristic differences of heterogeneous areas, enhance the local constraint on pixel space of a model, improve the quality of a band subset and further improve the accuracy of subsequent classification.
In order to achieve the above purpose, the invention adopts the following technical scheme:
the hyperspectral band selection method for auxiliary discrimination of the local view comprises the following steps:
step 1, according to the distribution characteristics of ground objects, a hyperspectral image is segmented by utilizing a super-pixel segmentation technology ERS, so as to form a pixel-level super-pixel block for capturing the spectrum characteristic differences of different ground object distributions;
Step 2, constructing a local spectrum-space hypergraph of each super pixel block by combining the spatial adjacency and the band spectrum correlation so as to express a multi-element adjacency relationship among the bands and rationalize a band adjacency structure;
Step 3, constructing a local view auxiliary discrimination hyperspectral band selection model RwSSR by combining the local spectrum-space hypergraph of the super pixel block constructed in the step 2 and the sparse self-representation model;
and 4, carrying out optimization solution on the hyperspectral band selection model RwSSR for auxiliary discrimination of the local view by adopting an iterative updating method to obtain a consensus matrix, calculating the band priority and selecting a band subset.
In addition, on the basis of the hyperspectral band selecting method for auxiliary discrimination of the partial view, the invention also provides a hyperspectral band selecting system for auxiliary discrimination of the partial view, which adopts the following technical scheme:
A partial view assisted discrimination hyperspectral band selection system comprising:
The super-pixel block segmentation module is used for segmenting the hyperspectral image by utilizing a super-pixel segmentation technology ERS according to the distribution characteristics of the ground objects to form pixel-level super-pixel blocks for capturing the spectrum characteristic differences of different ground object distributions;
The local spectrum-space hypergraph construction module is used for constructing a local spectrum-space hypergraph of each super pixel block by combining the spatial proximity and the band spectrum correlation so as to express the multi-element adjacent relation among the bands and rationalize the band adjacent structure;
The hyperspectral wave band selection model construction module is used for combining the local spectrum-space hypergraph of the constructed super pixel block and the sparse self-expression model to construct a local view auxiliary discrimination hyperspectral wave band selection model RwSSR;
And the band subset selection module is used for carrying out optimization solution on the local view auxiliary discrimination hyperspectral band selection model RwSSR by adopting an iterative updating method to obtain a consensus matrix, calculating the band priority and selecting the band subset.
In addition, on the basis of the above-mentioned method for selecting hyperspectral wave bands by using the auxiliary discrimination of the partial view, the invention also provides a computer device which comprises a memory and one or more processors.
The memory stores executable codes, and the processor is used for realizing the steps of the above-mentioned method for selecting the hyperspectral wave bands by using the auxiliary discrimination of the local view when executing the executable codes.
In addition, on the basis of the above-mentioned method for selecting hyperspectral wave bands by using the auxiliary discrimination of the partial view, the invention also provides a computer readable storage medium on which a program is stored. The program, when executed by a processor, is adapted to carry out the steps of the above-mentioned method for locally view assisted discrimination of hyperspectral band selection.
The invention has the following advantages:
As described above, the invention provides a method and a system for selecting hyperspectral wavebands by using a local view for assisting discrimination. The method comprises the steps of firstly segmenting a hyperspectral image by using a superpixel segmentation technique ERS to form a series of superpixel blocks to capture spectrum characteristic differences of different ground object distributions, secondly regarding a wave band as a supergraph vertex aiming at each superpixel block, constructing a corresponding supergraph to express a multi-element adjacent relation among the wave bands, rationalizing a wave band adjacent structure as far as possible to guide optimization of a sparse self-representation model, reducing redundancy of a wave band subset, wherein the consensus matrix fuses coefficient matrixes of the superpixel blocks through an iterative updating method, finally calculating importance of reconstructed original data of each wave band according to the consensus matrix, selecting the wave band subset, and therefore achieving the purpose of representing the hyperspectral image by using the unified wave band subset.
Drawings
Fig. 1 is a flowchart of a method for selecting hyperspectral bands with assistance of a partial view in an embodiment of the present invention.
Fig. 2 is a schematic diagram of a process for segmenting hyperspectral images using the super-pixel segmentation technique ERS.
Fig. 3 is a graph showing average accuracy versus INDIAN PINES datasets for each band selection method in an embodiment of the present invention.
Fig. 4 is a graph of overall accuracy versus INDIAN PINES datasets for various band selection methods in an embodiment of the present invention.
Detailed Description
The invention is described in further detail below with reference to the attached drawings and detailed description:
Example 1
The embodiment 1 describes a method for selecting hyperspectral wave bands by using a local view auxiliary discrimination method, so as to solve the problem that the conventional sparse self-expression wave band selection method has insufficient local constraint on pixel space in the wave band selection process.
As shown in fig. 1, the method for selecting hyperspectral bands by using the partial view assisted discrimination method comprises the following steps:
and step 1, according to the distribution characteristics of the ground objects, the hyperspectral image is segmented by utilizing a super-pixel segmentation technology ERS, so as to form a pixel-level super-pixel block for capturing the spectrum characteristic differences of different ground object distributions.
The hyperspectral image is divided by using a hyperspectral image segmentation technology ERS, the hyperspectral image is divided into a series of hyperspectral blocks X s which are similar in spectral characteristics and are not overlapped with each other, different ground objects are covered, corresponding sensitive wave band sets are selected for the different ground objects, wherein S is defined to represent the number of the hyperspectral blocks, and S is more than or equal to 1 and less than or equal to S.
ERS maps the first principal component of the hyperspectral image into a graph G (V, E), wherein the vertex set V comprises all pixel points, E is an edge set connecting the pixel points, and the weight w (E ij) represents the similarity between adjacent pixel points, and the formula is as follows:
Wherein v i、vj represents pixel points, v i、vj∈V;eij represents one edge connected with the pixel points v i and v j, e ij∈E;Gs represents the S super pixel block, and S is more than or equal to 1 and less than or equal to S, namely
The similarity between pixels within the same super-pixel block is calculated with exp (- |v i-vj||/2δ2), the similarity between pixels of different super-pixel blocks is 0, where δ represents a kernel parameter.
Defining hyperspectral image X, x= [ b 1,b2,...,bL],bl ] represents the first band, L e [1, L ], L represents the band number, each band b l contains N pixels, i.e. b l＝(x1,x2,...,xN)T,xn represents the nth pixel, N e [1, N ].
The objective function definition of ERS is shown in equation (2), which finds a subset a from the edge set E, so that the graphAll Y connected subgraphs are contained, and finally, the purpose of dividing the image is achieved by removing certain edges in the original edge set E.
Wherein the figureA graph obtained by performing ERS segmentation on the first principal component of the hyperspectral image is shown.
H (A) is a graphAnd B (A) is a balance item describing cluster distribution, and the similarity of the scale of each cluster is ensured.
Μ is a variable weight factor for coordinating the proportional relationship between H (a) and B (a). N A represents a graphThe number of connected sub-graphs, Y represents the number of preset connected sub-graphs, V, A represents the graph respectivelyAnd a subset of edges.
The process of segmenting the hyperspectral image using the superpixel segmentation technique ERS is shown in fig. 2.
The connectivity of the edges between clusters can be eliminated through the objective function of ERS, so that S super-pixel blocks are generated through segmentation, and the corresponding hyperspectral image X is re-represented as x= [ X 1,X2,...Xs,...XS],1≤s≤S,Xs ] to represent the super-pixel blocks.
And 2, constructing a local spectrum-space hypergraph of each super pixel block by combining the spatial adjacency and the band spectrum correlation so as to express the multi-element adjacency relationship among the bands and rationalize the band adjacency structure.
Combining the spatial proximity and the band spectrum correlation, constructing a corresponding local spectrum-space hypergraph for each super pixel block X s To form a complete image hypergraph structure
Local spectrum-space hypergraph of each superpixel block X s Represented asWherein, the Representing a partial spectral-spatial hypergraphIs provided with a vertex set of (a),Representing a partial spectral-spatial hypergraphIs a hyperedge set of (1).
Setting a band vectorIs a vertex, a vertex setIs composed of L wave band vectors, E i denotes a superside, e i∈Es, of the K-nearest neighbor component of the band vector b i.
The weight w (e i) of the superside e i is determined by the connection relationship between all vertices in the superside e i, and the calculation formula is as follows:
wherein θ represents a balance parameter, and the expression thereof is as follows:
Wherein, the band vector b j∈Vs is a vertex, and K represents the number of neighbors of the superside e i. f (b i)、f(bj) represents the integral function of vertex b i、bj in the hypergraph, respectively.
Determined by the remaining vertices in the superedge e i formed by b i and b i as centers, i.e
Because the affinities of the rest vertexes in the superside and b i are different, the weight is set according to the affinities between the vertex b i and the vertex b j as follows:
Wherein the band vector b k∈Vs is a vertex.
The value of f ij simultaneously measures the correlation of the wave band spectrumSpatial proximityI.e.
AndAs shown in equations (4) and (5), respectively:
wherein the method comprises the steps of Is a balance parameter.
AndThe spatial index values of band vector b i and band vector b j are represented, respectively.
Construction of each local hypergraphIs a degree matrix of vertices, and is a correlation matrix H s, a weight matrix W s, and a degree matrix of verticesDegree matrix of superside
Ws＝diag(ws(e1),ws(e2)…ws(eL)) (7)
Where w s(ei) represents the weight of the superside e i, i ε [1, L ].
W s(ej) represents the weight of the superside e j, h s(bi,ej) represents the correlation matrix of the band vector b i, diag (·) represents the diagonal matrix.
So far, the local spectrum-space hypergraph is obtainedThe corresponding Laplace matrix L s is as follows:
Wherein, I is an identity matrix, L s utilizes the special spectral characteristics of the wave bands and the spatial similarity among wave band sequences to more truly reflect the local spectrum-space hypergraph Actual adjacency between bands.
And 3, constructing a local view auxiliary discrimination hyperspectral band selection model RwSSR by combining the local spectrum-space hypergraph of the super pixel block constructed in the step 2 and the sparse self-expression model.
The local view assisted discrimination hyperspectral band selection model RwSSR is obtained as follows:
Assuming that x= [ b 1,b2,...,bL ] represents a hyperspectral image, consisting of L bands, where each band contains N pixels b l＝(x1,x2,...,xN)T, and taking into account the band redundancy of the hyperspectral image, adding L 2,1 regularization sparsely performs a coefficient matrix, approximating the dataset X with as few bands as possible, i.e., dictionary columns, the sparse self-representation model can be expressed as:
Wherein A is represented as a sparse coefficient matrix, F represents an F norm, and a formula for carrying out the F norm on the sparse coefficient matrix is represented as Alpha is denoted as regularization parameter, | 2,1 denotes l 2,1 norm, A≥0 is used to ensure the non-negative properties of A, diag (A) is denoted by itself to prevent it from being represented.
Corresponding to any super-pixel block X s after super-pixel segmentation, the objective function is defined as:
Where a s is the local coefficient matrix for each super-pixel block X s, a s represents the reconstruction importance of the hyperspectral band to the super-pixel block X s. In order to mine the band reconstruction information of the whole hyperspectral image, the invention integrates all the local coefficient matrixes A s to form the consensus matrix A of the whole hyperspectral image, so that the consensus matrix A integrally constrains the hyperspectral image to integrally reflect the local characteristics of the hyperspectral image, the band selection is more accurate, and the objective function is further expressed as:
Where λ 1 represents the regularization parameters.
W s is an adaptive balance parameter for ensuringOverall minimization. In particular, ifTo be large, w s should be reduced in order to minimize this, and vice versa.
For simplicity, the parameter w s is set to 1/2I A s -A I in this embodiment. By passing throughAnd fusing the coefficient matrix A s of each super pixel block into a consensus matrix A, so that the consensus matrix A contains key band information of the super pixel block.
Because of spectral similarity between bands and spatial similarity between adjacent sequence bands, these correlations cannot be represented in equation (13). To this end, according to the definition of the spectrum-space hypergraph, optimize (13) as:
Where lambda 2 is the regularization parameter, L s is the Laplacian matrix for each superpixel block X s, A s is the coefficient matrix for each superpixel block X s, tr (·) represents the trace operator, avoiding each band to be represented by itself, |a| 2,1 denotes the L 2,1 norm of the consensus matrix a.
Therefore, on the basis of keeping spatial information of pixel level, the model introduces a band local constraint term L s of spectrum level, so that various information of pixel space and spectrum space is comprehensively considered in the optimization process, and the subsequent sparse representation model solution is improved.
And 4, carrying out optimization solution on the hyperspectral band selection model RwSSR for auxiliary discrimination of the local view by adopting an iterative updating method to obtain a consensus matrix, calculating the band priority and selecting a band subset.
The iterative updating algorithm is adopted to solve the objective function, and the process of obtaining A s、A、ws is specifically as follows:
Step 4.1. First fix A and w s, update A s, transform the fixed objective function into:
deriving A s, the updated equation for A s is:
Wherein n represents the current iteration number, A s (n+1) represents the value of the (n+1) th iteration A s, A (n) represents the value of the (n) th iteration A, Representing the value of the nth iteration w s.
Step 4.2. Fix A s and w s next, update A, transform the objective function after fixing into:
set u=diag (U 1,u2,…,uL) is an L x L diagonal matrix,
Wherein a i represents the ith row in consensus matrix A, ||a i||2 represents the l 2 norm of a i, ζ is the number in u i where zero denominator is avoided, and accordingly, formula (17) is rewritten as:
Derive A and let The updated equation for A is obtained as:
wherein a (n+1) represents a value representing the n+1th iteration a.
Step 4.3. Re-fix A s and A, update w s, get the update equation of w s as follows:
Wherein, the Representing the value of the n+1st iteration w s.
Stopping when the iteration is updated to the prescribed iteration number or is smaller than the set threshold, at this time, each row a i in the resulting consensus matrix a represents the contribution of the ith band to reconstructed raw data X.
Wherein, the larger the value of r i＝||ai2 I is, the more important the band is.
Thus, r i is ordered in descending order, selecting the first n BS bands as the final band subset, serving the subsequent classifications.
In addition, in order to verify the effectiveness of the method proposed by the present invention, the following experiments were also performed:
1. Sample data INDIAN PINES hyperspectral data were obtained from the Ind laboratory in the United states and were taken by an on-board visible infrared imaging spectrometer (AVIRIS) having 220 bands with a spectral range from 0.4 μm to 2.5 μm and a size of 145X 220, containing 16 classes of target features, including no-tillage corn, split soybeans, woods, etc. The dataset contained 21025 pixels, where the total number of target pels was 10249 and the background contained 10776 pixels.
2. Experiment setting:
the comparison method comprises the steps of classifying by using a Support Vector Machine (SVM) as a classifier, and verifying the reliability and accuracy of the method by using seven different comparison methods, wherein the comparison methods are respectively a maximum variance principal component analysis Method (MVPCA), an enhanced rapid density peak clustering method (E-FDPC), an optimal class frame aggregation method (OCF), an Adaptive Subspace Partitioning Strategy (ASPS), an extensible single-pass self-learning method (SOP-SRL), a graph regularization space-spectrum subspace clustering method (GRSC) and a Full band (Full bands).
The evaluation index is that the experiment adopts two indexes to evaluate the quality of the wave band subset, namely average precision (AA) and total precision (OA), and the larger the value of the evaluation index is, the better the classification effect is, and the selected wave band subset is more accurate. 10% of the samples were randomly selected as training set in the experiment, the remainder being used for testing. All experiments were repeated five times and the average was calculated.
3. Parameter analysis:
Five parameters in total need to be adjusted for the experiment, S, K, lambda 1、λ2 and alpha respectively. Wherein K is the number of neighbors of each superside, the set parameter range is {3,5,7,9}, and the parameter lambda 1、λ2, alpha take the value range is {1e -3,1e-2,1e-1,1,1e1,1e2,1e3 }, wherein the range of the number S of the super pixel segmentation blocks is {10,50,100}. The experiment uses voting to determine the optimal parameters. Firstly, selecting an optimal parameter set under different wave bands, then selecting a group of parameters with the largest occurrence number in a voting mode, and recording the results obtained under the group of parameters as final results. For the INDIAN PINES dataset, the final values of the five parameters are shown in table 1.
TABLE 1
| Parameters (parameters) | S | K | λ1 | λ2 | α |
| Optimum value | 100 | 3 | 1e-1 | 1e-3 | 1e-3 |
4. Experimental results:
Fig. 3 and 4 show the overall and average accuracy of the comparison at different band numbers on INDIAN PINES datasets, respectively. As can be seen from FIGS. 3 and 4, E-FDPC, MVPCA, OCF performs poorly downstream. SOP-SRL and ASPS effects are unstable. GRSC is stable in effect but does not appear to stand out. With the band number of 15 as a demarcation point, the RwSSR method gradually overtakes other comparison methods when the band number is smaller than 15, and RwSSR is far ahead of the other comparison methods and gradually becomes stable when the band number is larger than 15. This indicates RwSSR shows good performance and enables selection of a subset of bands that contribute to hyperspectral image classification. In conclusion, the method has excellent performance and stable performance in the wave band selection process.
According to the hyperspectral band selection method, firstly, a hyperspectral image is segmented by using a hyperspectral pixel segmentation technology ERS to form a series of hyperspectral blocks to capture spectrum characteristic differences of different ground object distributions, secondly, a local spectrum-space hyperspectral image of each hyperspectral pixel block is constructed to express a multi-element adjacent relation among bands, the band adjacent structure is rationalized as much as possible, and finally band information of the hyperspectral pixel blocks is integrated through a consensus matrix, so that the aim of representing the whole hyperspectral image by using a unified band subset is fulfilled, local constraint of pixel space is practically enhanced, the quality of the band subset is improved, and therefore the problem that the key characteristic extraction of pixel space information is ignored in the band selection process by a traditional sparse self-representation model is effectively solved.
Example 2
This embodiment 2 describes a partial view assisted discrimination hyperspectral band selecting system based on the same inventive concept as the partial view assisted discrimination hyperspectral band selecting method in embodiment 1 described above.
Specifically, the hyperspectral band selection system for auxiliary discrimination of the local view comprises:
The super-pixel block segmentation module is used for segmenting the hyperspectral image by utilizing a super-pixel segmentation technology ERS according to the distribution characteristics of the ground objects to form pixel-level super-pixel blocks for capturing the spectrum characteristic differences of different ground object distributions;
The local spectrum-space hypergraph construction module is used for constructing a local spectrum-space hypergraph of each super pixel block by combining the spatial proximity and the band spectrum correlation so as to express the multi-element adjacent relation among the bands and rationalize the band adjacent structure;
The hyperspectral wave band selection model construction module is used for combining the local spectrum-space hypergraph of the constructed super pixel block and the sparse self-expression model to construct a local view auxiliary discrimination hyperspectral wave band selection model RwSSR;
And the band subset selection module is used for carrying out optimization solution on the local view auxiliary discrimination hyperspectral band selection model RwSSR by adopting an iterative updating method to obtain a consensus matrix, calculating the band priority and selecting the band subset.
It should be noted that, in the hyperspectral band selection system with auxiliary discrimination of the partial view, the implementation process of the functions and roles of each functional module is specifically shown in the implementation process of the corresponding steps in the method in the above embodiment 1, and will not be described herein.
Example 3
Embodiment 3 describes a computer apparatus for implementing the steps of the partial view assisted discrimination hyperspectral band selection method described in embodiment 1 above.
The computer device includes a memory and one or more processors. Executable code is stored in the memory for implementing the steps of the partial view assisted discrimination hyperspectral band selection method when the executable code is executed by the processor.
In this embodiment, the computer device is any device or apparatus having data processing capability, which is not described herein.
Example 4
Embodiment 4 describes a computer-readable storage medium for implementing the steps of the partial view assisted discrimination hyperspectral band selection method described in embodiment 1 above.
The computer-readable storage medium of embodiment 4 has stored thereon a program which, when executed by a processor, is adapted to implement the steps of a method for localized view assisted discrimination of hyperspectral band selection.
The computer readable storage medium may be any internal storage unit of a device or apparatus having data processing capability, such as a hard disk or a memory, or may be any external storage device of a device having data processing capability, such as a plug-in hard disk, a smart memory card (SMART MEDIA CARD, SMC), an SD card, a flash memory card (FLASH CARD), or the like, provided on the device.
The foregoing description is, of course, merely illustrative of preferred embodiments of the present invention, and it should be understood that the present invention is not limited to the above-described embodiments, but is intended to cover all modifications, equivalents and alternatives falling within the spirit and scope of the present invention as defined by the appended claims.
Claims (6)
1. The hyperspectral band selection method for auxiliary discrimination of the partial view is characterized by comprising the following steps of:
step 1, according to the distribution characteristics of ground objects, a hyperspectral image is segmented by utilizing a super-pixel segmentation technology ERS, so as to form a pixel-level super-pixel block for capturing the spectrum characteristic differences of different ground object distributions;
Step 2, constructing a local spectrum-space hypergraph of each super pixel block by combining the spatial adjacency and the band spectrum correlation so as to express a multi-element adjacency relationship among the bands and rationalize a band adjacency structure;
Step 3, constructing a local view auxiliary discrimination hyperspectral band selection model RwSSR by combining the local spectrum-space hypergraph of the super pixel block constructed in the step 2 and the sparse self-representation model;
the step 3 specifically comprises the following steps:
combining the local spectrum-space hypergraph of the super pixel block and a sparse self-expression model, constructing a hyperspectral band selection model RwSSR based on local view auxiliary discrimination, and defining an objective function as follows:
Where λ 1 and λ 2 are regularization parameters, L s is the laplacian matrix of each superpixel block X s, a s is the coefficient matrix of each superpixel block X s, tr (·) represents the trace operator, avoiding each band from being represented by itself;
w s is an adaptive balance parameter for ensuring The whole is minimized, alpha represents regularization parameters, A 2,1 performs sparseness on the consensus matrix A, F represents F norms, and F norms are represented through the methodThe method comprises the steps of fusing coefficient matrixes A s of all super pixel blocks into a consensus matrix A, so that the consensus matrix A contains key band information of all the super pixel blocks;
Step 4, optimizing and solving a hyperspectral wave band selection model RwSSR for auxiliary discrimination of the local view by adopting an iterative updating method to obtain a consensus matrix, calculating the wave band priority and selecting a wave band subset;
the step 4 specifically comprises the following steps:
The iterative updating algorithm is adopted to solve the objective function, and the process of obtaining A s、A、ws is specifically as follows:
Step 4.1. First fix A and w s, update A s, transform the fixed objective function into:
deriving A s, the updated equation for A s is:
Wherein n represents the current iteration number, A s (n+1) represents the value of the (n+1) th iteration A s, A (n) represents the value of the (n) th iteration A, Representing the value of the nth iteration w s;
Step 4.2. Fix A s and w s next, update A, transform the objective function after fixing into:
Setting u=diag (U 1,u2,…,uL) to be an l×l diagonal matrix;
Where a i represents the ith row in consensus matrix A, ||a i||2 represents the l 2 norm of a i, Is a number in u i that avoids zero denominator, from which equation (14) is rewritten as:
Derive A and let The updated equation for A is obtained as:
wherein A (n+1) represents the value of the (n+1) th iteration A;
step 4.3. Re-fix A s and A, update w s, get the update equation of w s as follows:
Wherein, the Representing the value of the n+1st iteration w s;
When the iteration is updated to a prescribed number of iterations, or stopping when the value of A (n+1)-A(n) is smaller than the set threshold, at this time, each row a i in the obtained consensus matrix a represents the contribution of the ith band to reconstructed raw data X;
Wherein a larger value of r i＝||ai||2 indicates that the band is more important;
Thus, r i is ordered in descending order, selecting the first n BS bands as the final band subset.
2. The method for selecting a hyperspectral band by partial view assisted discrimination as claimed in claim 1 wherein,
The step 1 specifically comprises the following steps:
Dividing the hyperspectral image by using a hyperspectral pixel segmentation technology ERS, dividing the hyperspectral image into a series of hyperspectral regions X s which are similar in spectral characteristics and are not overlapped with each other so as to represent different ground objects to be covered, and selecting a corresponding sensitive wave band set for the different ground objects, wherein S is defined to represent the number of the hyperspectral blocks, and S is more than or equal to 1 and less than or equal to S;
ERS maps the first principal component of the hyperspectral image into a graph G (V, E), wherein the vertex set V comprises all pixel points, E is an edge set connecting the pixel points, and the weight w (E ij) represents the similarity between adjacent pixel points, and the formula is as follows:
Wherein v i、vj represents pixel points, v i、vj∈V;eij represents one edge connected with the pixel points v i and v j, e ij∈E;Gs represents the S super pixel block, and S is more than or equal to 1 and less than or equal to S, namely
The similarity between the pixel points in the same super pixel block is calculated by exp (- |v i-vj||/2δ2), and the similarity between the pixel points of different super pixel blocks is 0, wherein delta represents a kernel parameter;
Defining a hyperspectral image X, wherein X= [ b 1,b2,...,bL],bl ] represents a first wave band, L epsilon [1, L ] represents the number of wave bands, and each wave band b l comprises N pixels, namely b l＝(x1,x2,...,xN)T,xn represents an nth pixel, and N epsilon [1, N ];
the objective function definition of ERS is shown in equation (2), which finds a subset a from the edge set E, so that the graph All Y connected subgraphs are contained, and the image is segmented by removing certain edges in the original edge set E;
Wherein the figure A graph showing the hyperspectral image after ERS segmentation of the first principal component;
H (A) is a graph B (A) is a balance item describing cluster distribution, and the similarity of the scale of each cluster is ensured;
mu is a variable weight factor for coordinating the proportional relationship between H (A) and B (A), N A represents a graph The number of the connected subgraphs in the graph, Y represents the number of preset connected subgraphs, V, A represents the graph respectivelyA subset of vertices and edges of (a);
The connectivity of the edges between clusters can be eliminated through the objective function of ERS, so that S super-pixel blocks are generated through segmentation, and the corresponding hyperspectral image X is re-represented as x= [ X 1,X2,...Xs,...XS],1≤s≤S,Xs ] to represent the super-pixel blocks.
3. The method for selecting a hyperspectral band by partial view assisted discrimination as claimed in claim 2 wherein,
The step 2 specifically comprises the following steps:
combining the spatial proximity and the band spectrum correlation, constructing a corresponding local spectrum-space hypergraph for each super pixel block X s To form a complete image hypergraph structure
Local spectrum-space hypergraph of each superpixel block X s Represented asWherein, the Representing a partial spectral-spatial hypergraphIs provided with a vertex set of (a),Representing a partial spectral-spatial hypergraphIs a hyperedge set of (1);
Setting a band vector b i∈Vs as a vertex, wherein a vertex set V s consists of L band vectors, and V s＝{b1,b2,…bL},i∈[1,L],ei represents a superside consisting of K-neighbors of the band vector b i, and e i∈Es;
The weight w (e i) of the superside e i is determined by the connection relationship between all vertices in the superside e i, and the calculation formula is as follows:
wherein θ represents a balance parameter, and the expression thereof is as follows:
Wherein, the wave band vector b j∈Vs is a vertex, K represents the number of neighbors of the superside e i, f (b i)、f(bj) respectively represents the integral function of the vertex b i、bj in the hypergraph;
Determined by the remaining vertices in the superedge e i formed by b i and b i as centers, i.e
Because the affinities of the rest vertexes in the superside and b i are different, the weight is set according to the affinities between the vertex b i and the vertex b j as follows:
wherein, the band vector b k∈Vs is a vertex;
The value of f ij simultaneously measures the correlation of the wave band spectrum Spatial proximityI.e.
AndAs shown in equations (4) and (5), respectively:
Wherein, the Is a balance parameter;
And Spatial index values respectively representing band vector b i and band vector b j are respectively used for constructing each partial spectrum-space hypergraphIs a degree matrix of vertices, and is a correlation matrix H s, a weight matrix W s, and a degree matrix of verticesDegree matrix of superside
Ws＝diag(ws(e1),ws(e2)…ws(eL)) (7)
Wherein w s(ei) represents the hyperedge e i weight, i ε [1, L ];
w s(ej) represents the weight of the superside e j, h s(bi,ej) represents the correlation matrix of the band vector b i, diag (·) represents the diagonal matrix;
So far, the local spectrum-space hypergraph is obtained The corresponding Laplace matrix L s is as follows:
Wherein I is an identity matrix.
4. A partial view assisted discrimination hyperspectral band selecting system for realizing the partial view assisted discrimination hyperspectral band selecting method as claimed in any one of claims 1 to 3,
The local view assisted discrimination hyperspectral band selection system comprises:
The super-pixel block segmentation module is used for segmenting the hyperspectral image by utilizing a super-pixel segmentation technology ERS according to the distribution characteristics of the ground objects to form pixel-level super-pixel blocks for capturing the spectrum characteristic differences of different ground object distributions;
The local spectrum-space hypergraph construction module is used for constructing a local spectrum-space hypergraph of each super pixel block by combining the spatial proximity and the band spectrum correlation so as to express the multi-element adjacent relation among the bands and rationalize the band adjacent structure;
The hyperspectral wave band selection model construction module is used for combining the local spectrum-space hypergraph of the constructed super pixel block and the sparse self-expression model to construct a local view auxiliary discrimination hyperspectral wave band selection model RwSSR;
And the band subset selection module is used for carrying out optimization solution on the local view auxiliary discrimination hyperspectral band selection model RwSSR by adopting an iterative updating method to obtain a consensus matrix, calculating the band priority and selecting the band subset.
5. A computer device comprising a memory and one or more processors, the memory having executable code stored therein, wherein the processor, when executing the executable code, performs the steps of the method for partial view assisted discrimination hyperspectral band selection as claimed in any one of claims 1 to 3.
6. A computer-readable storage medium having a program stored thereon, which when executed by a processor, implements the steps of the partial view assisted discrimination hyperspectral band selection method as claimed in any one of claims 1 to 3.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310887160.1A CN116778339B (en) | 2023-07-19 | 2023-07-19 | Method and system for selecting hyperspectral wave bands by aid of local view auxiliary discrimination |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310887160.1A CN116778339B (en) | 2023-07-19 | 2023-07-19 | Method and system for selecting hyperspectral wave bands by aid of local view auxiliary discrimination |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN116778339A CN116778339A (en) | 2023-09-19 |
| CN116778339B true CN116778339B (en) | 2025-11-28 |
Family
ID=87989621
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310887160.1A Active CN116778339B (en) | 2023-07-19 | 2023-07-19 | Method and system for selecting hyperspectral wave bands by aid of local view auxiliary discrimination |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116778339B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN118072180B (en) * | 2024-04-17 | 2024-06-18 | 昆明理工大学 | Hyperspectral image classification method based on keyson hypergraph enhancement |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113486876A (en) * | 2021-09-08 | 2021-10-08 | 中国地质大学（武汉） | Hyperspectral image band selection method, device and system |
| CN115984113A (en) * | 2023-01-09 | 2023-04-18 | 青岛大学 | Spectrum-air hypergraph regularization sparse self-representation hyperspectral waveband selection method |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106778680B (en) * | 2017-01-06 | 2018-06-26 | 杭州电子科技大学 | A kind of hyperspectral image band selection method and device based on critical bands extraction |
| CN110399909B (en) * | 2019-07-08 | 2021-07-27 | 南京信息工程大学 | A Hyperspectral Image Classification Method Based on Label-constrained Elastic Net Graph Model |
-
2023
- 2023-07-19 CN CN202310887160.1A patent/CN116778339B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113486876A (en) * | 2021-09-08 | 2021-10-08 | 中国地质大学（武汉） | Hyperspectral image band selection method, device and system |
| CN115984113A (en) * | 2023-01-09 | 2023-04-18 | 青岛大学 | Spectrum-air hypergraph regularization sparse self-representation hyperspectral waveband selection method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN116778339A (en) | 2023-09-19 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US11301727B2 (en) | Efficient image classification method based on structured pruning | |
| CN111860612B (en) | Unsupervised hyperspectral image hidden low-rank projection learning feature extraction method | |
| Jia et al. | Spectral–spatial Gabor surface feature fusion approach for hyperspectral imagery classification | |
| CN111368691B (en) | Unsupervised hyperspectral remote sensing image space spectrum feature extraction method | |
| Wang et al. | Region-aware hierarchical latent feature representation learning-guided clustering for hyperspectral band selection | |
| Kanmani et al. | MRI brain images classification: a multi-level threshold based region optimization technique | |
| US20240212307A1 (en) | Hyperspectral image band selection method and system based on latent feature fusion | |
| Chen et al. | Semi-supervised adaptive pseudo-label feature learning for hyperspectral image classification in internet of things | |
| US9082043B2 (en) | Image congealing via efficient feature selection | |
| Jia et al. | Superpixel-based multitask learning framework for hyperspectral image classification | |
| US12159410B2 (en) | Image segmentation method using neural network based on mumford-shah function and apparatus therefor | |
| Wang et al. | Self-supervised low-rank representation (SSLRR) for hyperspectral image classification | |
| CN112163114B (en) | Image retrieval method based on feature fusion | |
| CN107273919B (en) | A Hyperspectral Unsupervised Classification Method Based on Constructing Generic Dictionary Based on Confidence | |
| Mei et al. | PSASL: Pixel-level and superpixel-level aware subspace learning for hyperspectral image classification | |
| CN115424138B (en) | A hyperspectral image classification method based on deep neural network | |
| CN108805061A (en) | Hyperspectral image classification method based on local auto-adaptive discriminant analysis | |
| Mignotte | MDS-based multiresolution nonlinear dimensionality reduction model for color image segmentation | |
| Tu et al. | Feature extraction using multidimensional spectral regression whitening for hyperspectral image classification | |
| CN115496933A (en) | Hyperspectral classification method and system based on space-spectrum prototype feature learning | |
| CN114743009A (en) | A kind of hyperspectral image band selection method, system and electronic equipment | |
| CN116778339B (en) | Method and system for selecting hyperspectral wave bands by aid of local view auxiliary discrimination | |
| Li et al. | A novel dimensionality reduction algorithm for Cholangiocarcinoma hyperspectral images | |
| Han et al. | Texture segmentation using independent-scale component-wise Riemannian-covariance Gaussian mixture model in KL measure based multi-scale nonlinear structure tensor space | |
| Li et al. | Subspace learning via locally constrained a-optimal nonnegative projection |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |