CN114519769A - Low-redundancy hyperspectral band selection method and device considering representativeness and information content - Google Patents
Low-redundancy hyperspectral band selection method and device considering representativeness and information content Download PDFInfo
- Publication number
- CN114519769A CN114519769A CN202111637324.2A CN202111637324A CN114519769A CN 114519769 A CN114519769 A CN 114519769A CN 202111637324 A CN202111637324 A CN 202111637324A CN 114519769 A CN114519769 A CN 114519769A
- Authority
- CN
- China
- Prior art keywords
- band
- subset
- candidate
- hyperspectral
- redundancy
- 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.)
- Pending
Links
- 238000010187 selection method Methods 0.000 title claims abstract description 34
- 238000000034 method Methods 0.000 claims abstract description 20
- 238000011156 evaluation Methods 0.000 claims description 20
- 238000012549 training Methods 0.000 claims description 19
- 238000010367 cloning Methods 0.000 claims description 13
- 238000004364 calculation method Methods 0.000 claims description 12
- 238000005259 measurement Methods 0.000 claims description 7
- 230000000903 blocking effect Effects 0.000 claims description 6
- 238000013377 clone selection method Methods 0.000 claims description 6
- 230000035772 mutation Effects 0.000 claims description 6
- 238000000691 measurement method Methods 0.000 claims description 4
- 230000010076 replication Effects 0.000 claims description 4
- 238000010845 search algorithm Methods 0.000 claims description 4
- 238000000638 solvent extraction Methods 0.000 claims description 4
- 238000010276 construction Methods 0.000 claims description 3
- 238000001514 detection method Methods 0.000 claims description 3
- 230000002068 genetic effect Effects 0.000 claims description 3
- 230000003362 replicative effect Effects 0.000 claims description 3
- 238000012163 sequencing technique Methods 0.000 claims description 3
- 239000000126 substance Substances 0.000 claims description 3
- 230000017105 transposition Effects 0.000 claims description 3
- 238000005516 engineering process Methods 0.000 claims description 2
- 238000005457 optimization Methods 0.000 claims description 2
- 239000002245 particle Substances 0.000 claims description 2
- 238000001228 spectrum Methods 0.000 claims description 2
- 230000006870 function Effects 0.000 abstract description 11
- 238000012545 processing Methods 0.000 abstract description 8
- 238000013135 deep learning Methods 0.000 abstract 1
- 235000008331 Pinus X rigitaeda Nutrition 0.000 description 7
- 235000011613 Pinus brutia Nutrition 0.000 description 7
- 241000018646 Pinus brutia Species 0.000 description 7
- 230000004913 activation Effects 0.000 description 6
- 230000000694 effects Effects 0.000 description 6
- 238000002474 experimental method Methods 0.000 description 3
- 238000010606 normalization Methods 0.000 description 3
- 238000000605 extraction Methods 0.000 description 2
- 230000009467 reduction Effects 0.000 description 2
- 238000012935 Averaging Methods 0.000 description 1
- 238000010521 absorption reaction Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000003595 spectral effect Effects 0.000 description 1
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Chemical compound O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T17/00—Three dimensional [3D] modelling, e.g. data description of 3D objects
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/004—Artificial life, i.e. computing arrangements simulating life
- G06N3/006—Artificial life, i.e. computing arrangements simulating life based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T3/00—Geometric image transformations in the plane of the image
- G06T3/40—Scaling of whole images or parts thereof, e.g. expanding or contracting
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A40/00—Adaptation technologies in agriculture, forestry, livestock or agroalimentary production
- Y02A40/10—Adaptation technologies in agriculture, forestry, livestock or agroalimentary production in agriculture
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Software Systems (AREA)
- Biomedical Technology (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Computer Graphics (AREA)
- Biophysics (AREA)
- Geometry (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention relates to the field of hyperspectral remote sensing image processing, and discloses a low-redundancy hyperspectral band selection method and device considering both representativeness and information content. The method comprises the following steps: (1) constructing a 3D reconstruction network, and measuring the representativeness of the candidate waveband subset to the original hyperspectral image; (2) measuring the redundancy of the candidate wave band subset; (3) measuring the information quantity contained in the candidate waveband subset; (4) designing a band subset scoring function considering band representativeness, redundancy and information quantity, and evaluating candidate band subsets; (5) several candidate band subsets are generated, and the candidate band subset with the highest score is selected as the selected band subset. According to the method, based on the characteristics of the hyperspectral images, the inherent nonlinear correlation relationship among the wave bands of the hyperspectral images is excavated, the spatial information of the hyperspectral images is fully utilized, advanced deep learning knowledge is combined, a wave band selection method considering wave band representativeness, redundancy and information content is provided, and the precision of pixel classification of the hyperspectral images can be improved.
Description
Technical Field
The invention relates to the field of image processing, in particular to a low-redundancy hyperspectral band selection method and device considering both representativeness and information content.
Background
The hyperspectral image consists of hundreds of continuous wave bands and contains abundant spatial and spectral information. However, high-dimensional data also poses some challenges, such as information redundancy, heavy computational burden, and "hounsfield". Therefore, the method for reducing the dimension of the hyperspectral image and reserving effective information through a reasonable mode is very key for subsequent processing of the hyperspectral image. The methods of dimension reduction are generally divided into feature extraction and band selection. The former dimension reduction method can cause the loss of the original data physical information, and the latter can keep the original data physical information. Therefore, the band selection method is widely concerned by the scholars.
The existing band selection methods at home and abroad are roughly divided into the following four types: (1) a band selection method based on point-by-point search; (2) a band selection method based on grouping search; (3) a rank-based band selection method; (4) a band selection method based on an advanced machine learning algorithm. These methods select directly from the original hyperspectral image the subset of bands that contains the most useful information. However, the existing band selection method mainly faces three problems:
(1) the existing band selection method cannot give consideration to the representativeness, redundancy and information quantity of the band. For example, rank-based band selection methods typically only consider the amount of information or representativeness of the bands, and ignore the correlation between the bands.
(2) Most of the existing band selection methods cannot well analyze the inherent nonlinear correlation relationship between bands, and generally only simply consider the linear correlation between bands or the nonlinear correlation based on a predefined kernel function.
(3) The existing wave band selection method based on the self-encoder and the genetic algorithm cannot accurately reflect the representativeness of wave bands to an original image, and has the problem that spatial information of a hyperspectral image cannot be utilized.
Disclosure of Invention
Aiming at the defects in the prior art, the invention aims to provide a low-redundancy hyperspectral band selection method and device which take account of representativeness and information content, discloses the inherent nonlinear correlation among the hyperspectral image bands, fully utilizes the spatial information of the hyperspectral image, researches a band selection strategy which can take account of band representativeness, redundancy and information content, improves the extraction effect of band subsets containing abundant valuable information, and solves the problem of low pixel classification precision caused by the fact that the band subset with the most valuable information cannot be selected.
In order to achieve the purpose, the invention provides the following technical scheme:
the invention provides a low-redundancy hyperspectral band selection method considering both representativeness and information content, which comprises the following steps of:
step 1) constructing and training a 3D reconstruction network:
partitioning an original hyperspectral image, wherein each hyperspectral image block is used as a sample; taking the hyperspectral image blocks as the input of a 3D reconstruction network and training;
step 2) measuring band representativeness:
passing the hyperspectral image data block through a sparse binary mask, the calculation formula is expressed as:
wherein, mu belongs to [0,1 ]]LRepresenting a sparse binary mask indicating whether each band is contained in a subset of candidate bands, L representing the total number of bands,
representing multiplication symbols by band, XPRepresenting a hyperspectral image block;
rescaling the data passing through the mask; reconstructing original hyperspectral data by taking the rescaled data as the input of a trained 3D reconstruction network to obtain the representative measurement of a candidate waveband subset;
step 3), measuring the band redundancy; the band redundancy measurement method adopts Pearson correlation coefficient or vector subspace projection technology;
step 4), measuring the information content of the wave band; the wave band information measurement method adopts information divergence or information entropy;
step 5) constructing a comprehensive evaluation index giving consideration to band representativeness, redundancy and information content, wherein a calculation formula is represented as follows:
wherein α and β represent equilibrium coefficients, R (X)S) Representing a subset X of candidate bandsSRedundancy measure of, I (X)S) Representing a subset X of candidate bandsSThe magnitude of the information metric of (a),
representing a subset X of candidate bandsSIs a representative metric of (c), score (X)S) Representing a subset X of candidate bandsSThe score of the comprehensive evaluation index;
step 6) searching the desired waveband subset:
calculating the comprehensive evaluation index score of each candidate waveband subset by using a subset search strategy of a grouping search algorithm, and searching the waveband subset with the highest score as the selected waveband subset; the grouping type search algorithm adopts an immune clone selection algorithm, a genetic algorithm or a particle swarm optimization algorithm.
Further, the rescaling in the step 2) is to scale the data after passing through the mask according to a ratio between the total number of the bands and the number of the selected bands, and the calculation formula is expressed as:
wherein k represents the number of selected bands, YPRepresenting the rescaled data.
Further, the reconstructing original hyperspectral data to obtain a representative metric of the candidate waveband subset specifically comprises:
reconstructing original hyperspectral data, wherein the formula is as follows:
wherein the content of the first and second substances,
representing a reconstructed hyperspectral image block, F (-) represents a 3D reconstruction network, and theta represents a trainable parameter in the 3D reconstruction network;
obtaining a representative metric of the candidate band subset, wherein the formula is as follows:
wherein n represents the number of samples,
representing the ith input hyper-spectral image block,
representing the ith reconstructed hyperspectral image block; i | · | purple windFRepresents the number of the F norm,
representative metric values are represented.
Further, the redundancy measure value of the candidate band subset is calculated by using the Pearson correlation coefficient according to the following formula:
wherein R (X)S) Representing a subset X of candidate bandsSRedundancy measure of (x)s(i)Represents the ith band in the candidate band subset, | | | · | | represents a 2-norm, and the superscript T represents transposition.
Further, the calculation formula for calculating the information metric values of the candidate band subsets by using the information divergence is as follows:
wherein I (.) represents the information amount, xs(i)Representing the ith band, X, of a subset of the selected bandsSRepresents a subset of the candidate bands of the spectrum,
representative band xs(i)Normalized ith element, N stands for pixel number, qiIs according to the band xs(i)The mean and variance of the first element in the probability distribution are initialized randomly to generate Gaussian distribution and normalized, and k represents the number of the selected bands.
Further, searching the desired band subset by using an immune clone selection algorithm, specifically:
(1) constructing an initial population:
dividing all wave bands in an original hyperspectral image into k groups according to wave band index serial numbers, randomly selecting one wave band from each group of wave bands to form an initial antibody, wherein each antibody represents a candidate wave band subset containing k wave bands; repeating the steps m times to generate an initial population consisting of m initial antibodies, and calculating the comprehensive evaluation index score of the candidate waveband subset corresponding to each antibody in the initial population;
(2) cloning operation:
number of antibody replications n for each antibody in the replicating populationc(XS(i)) The overall evaluation index score depending on it, namely:
wherein, Floor (·) represents rounding-down, nc(XS(i)) Represents the ith antibody XS(i)The number of copies of; score (X)S(i)) Represents the ith antibody XS(i)The score of the comprehensive evaluation index;
(3) mutation operation:
randomly selecting some elements from each replicated antibody and replacing them with equal amounts of other candidate bands;
(4) selecting operation:
selecting m antibodies with highest affinity from all antibodies to form a new population;
(5) repeating (2) - (4) until the maximum value of the comprehensive evaluation index scores of the candidate waveband subsets in the new population changes by less than a threshold value tau.
The invention also provides a low-redundancy hyperspectral band selection device which gives consideration to representativeness and information content, and is used for realizing the hyperspectral band selection method; the hyperspectral band selection device comprises:
the 3D reconstruction network training module is used for constructing a 3D reconstruction network and training the 3D reconstruction network;
the wave band representativeness measuring module is used for measuring the representativeness of the candidate wave band subset to the original hyperspectral image;
the band redundancy measurement module is used for measuring the redundancy of the candidate band subset;
a band information metric module for measuring the amount of information contained in the subset of candidate bands;
the comprehensive evaluation index construction module is used for designing a wave band subset scoring function and evaluating candidate wave band subsets;
the optimal waveband subset searching module is used for generating a certain number of candidate waveband subsets, scoring the candidate waveband subsets and selecting the waveband subset with the highest score as a selected waveband subset;
and the wave band selection result output module is used for outputting the selected optimal wave band subset result.
Further, the band selection device further comprises an application module, and the application module uses the band selection result to perform hyperspectral image classification or target detection.
Further, the 3D reconstruction network training module includes:
the image blocking module is used for blocking the hyperspectral image, and each hyperspectral image block is used as a sample;
and the 3D reconstruction network module is used for constructing a 3D reconstruction network and training the constructed 3D reconstruction network.
Further, the optimal band subset search module includes:
the initial population building module is used for obtaining an initial population;
the cloning module is used for cloning the antibodies in the population;
the mutation module is used for carrying out mutation operation on the antibodies in the population;
the selection module is used for selecting the antibodies in the population;
the iteration stop condition judgment module is used for repeating the cloning module, the variation module and the selection module and judging whether the iteration is stopped or not according to the condition of the generated new population;
and the waveband selection module is used for sequencing the affinities of the antibodies in the population and selecting the waveband subset corresponding to the antibody with the highest affinity as the selected waveband subset.
The invention has the beneficial effects that:
1) aiming at the problem that the representativeness, the information content and the redundancy of the selected wave band subset cannot be considered in the prior art, the invention provides the comprehensive evaluation criterion of the wave band subset considering the representativeness, the information content and the redundancy of the wave band, so that the wave band subset which can well represent the original hyperspectral image, contains rich information and has lower redundancy can be selected, and the implementation of downstream tasks is facilitated.
2) The invention provides the method for measuring the representativeness of the wave band subsets by using the 3D reconstruction network, can solve the problem that the existing nonlinear relation between the wave bands can not be usually found in the prior art, and further improves the implementation effect of the downstream task.
3) The invention provides the technical scheme that the data passing through the mask are rescaled, the rescaled data are used as the input of a trained 3D reconstruction network to reconstruct the original hyperspectral image, the problem that the reconstruction error cannot accurately represent the representativeness of the wave band subset in the prior art can be solved, and the pixel classification precision can be improved.
Drawings
FIG. 1 is a flowchart illustrating basic steps of an embodiment of a low-redundancy hyperspectral band selection method considering both representativeness and information content according to the invention.
FIG. 2 is a schematic structural diagram of a hyperspectral image band selection apparatus according to the present invention.
FIG. 3 is a hyperspectral image from an Indian Pines dataset.
FIG. 4 is a graph of classification accuracy for different band selection methods when using SVM classifiers on Indian Pines datasets.
FIG. 5 is a graph of classification accuracy for different band selection methods using the EPF-G-G classifier on Indian Pines datasets.
Detailed Description
The technical solution in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
As shown in fig. 1, a flowchart of basic steps of an embodiment of the inventive hyperspectral image band selection method specifically includes the following steps.
Inputting: original hyperspectral image X belongs to RW×H×LWhere W × H is the number of pixels, and L represents the number of bands; the number of bands k is selected.
The method comprises the following steps: a 3D reconstruction network is trained.
(1) Partitioning an original hyperspectral image, and representing the hyperspectral image as XP∈Ra×a×LThe a x a is a pixel after partitioning, each hyperspectral image block is used as a sample, and n samples are obtained in total;
(2) and taking the hyperspectral image blocks as input, constructing a 3D reconstruction network, and properly training the network.
In this embodiment, the specific implementation step of training the 3D reconstruction network is to input the hyperspectral image blocks into the constructed convolutional auto-encoder, take the mean square error between the input data and the output data obtained through reconstruction as a loss function, and train 10 epochs.
The output channels of the first three two-dimensional convolution layers of the convolution self-encoder are respectively 128, 64 and 32, the sizes of convolution kernels are all 3 x 3, and batch normalization processing and ReLU activation function layer processing are carried out after each convolution layer; subsequently, the obtained data is passed through three two-dimensional deconvolution layers with a kernel size of 3 × 3, the sizes of output channels are 64, 128 and the number L of original bands, respectively, and after passing through each deconvolution layer, batch normalization processing and activation function layer processing are performed, in which the remaining activation function layers use ReLU activation functions except the last activation function layer using Sigmoid activation functions.
Step two: band representativeness is measured based on 3D reconstructed networks.
The hyperspectral image block XPWith a sparse binary mask, the calculation formula is expressed as:
wherein, mu belongs to [0,1 ]]LRepresenting a sparse binary mask indicating whether each band is contained in a subset of candidate bands, L representing the total number of bands,
representing multiplication symbols by band, XPRepresenting a hyperspectral image block.
Rescaling the data passing through the mask according to the ratio of the total wave band number to the selected wave band number:
wherein k represents the number of selected bands, YPRepresenting the rescaled data.
And (3) reconstructing original hyperspectral data by taking the rescaled data as the input of a trained 3D reconstruction network, wherein the expression is as follows:
wherein the content of the first and second substances,
represents a reconstructed hyperspectral image block, F (-) represents a 3D reconstruction network, and theta represents a trainable parameter in the 3D reconstruction network.
And calculating a reconstruction error to obtain a representative measure of the candidate waveband subset, wherein a calculation formula is represented as:
wherein n represents the number of samples,
representing the ith input hyper-spectral image block,
representing the ith reconstructed hyperspectral image block; i | · | purple windFRepresents the number of the F norm,
representative metric values are represented.
Step three: the redundancy of the candidate band subset is measured.
The embodiment adopts Pearson correlation coefficients to calculate the redundancy of the candidate waveband subsets. Specifically, the Pearson correlation coefficient between every two wave bands in the candidate wave band subset is calculated first, and then the average is taken to obtain the redundancy measure value of the candidate wave band subset, wherein the calculation formula is represented as:
wherein, XS＝[xs(1),xs(2),...,xs(k)]∈RN×kRepresenting two dimensions obtained by expanding an original hyperspectral image X according to a spatial dimensionAccording to a candidate band subset, N represents the number of pixels, R (X)S) Represents a subset X of candidate bandsSRedundancy measure of (x)s(i)Represents the ith band in the candidate band subset, | | | · | | represents a 2-norm, and the superscript T represents transposition.
Step four: the amount of information of the candidate band subset is measured.
The present embodiment employs information divergence as the amount of information of the candidate band subset. Specifically, the information divergence for each band in the subset of candidate bands is first calculated:
wherein I (x) represents the amount of information in band x,
represents the ith element of the band x after normalization, wherein the normalized band x is the sum of each element in the band x divided by all elements, and q is [ q ═ q1,q2,...,qN]T∈RN×1Is probability distribution obtained by initializing Gaussian distribution generated according to the mean value and variance of x and normalizing the Gaussian distribution, qiIs the ith element in the probability distribution.
Averaging the information quantities of all the wave bands in the candidate wave band subset to obtain the information measurement value of the candidate wave band subset, wherein the calculation formula is represented as:
step five: constructing a comprehensive evaluation index giving consideration to band representativeness, redundancy and information content, wherein a calculation formula is represented as follows:
where α and β represent equilibrium coefficients.
Step six: searching for a desired subset of bands based on an immune clonal selection algorithm
(1) Constructing an initial population:
assuming a population consists of m antibodies; an antibody represents a candidate band subset comprising k bands, denoted XS(i)(i＝1,...,m)。
Firstly, all wave bands (L wave bands) in an original hyperspectral image X are uniformly divided into k groups according to wave band index serial numbers. If the total number of bands L is not evenly divisible by k, then each group in the first k-1 group is made to contain
The bands, the remaining bands, are divided into a final group, where Round (-) represents the rounded up symbol. One band from each set of bands was then randomly selected for composing the initial antibody. The procedure of generating the initial antibody was repeated m times to generate an initial population.
After the initial population is generated, the comprehensive evaluation index of the candidate wave band subset is used as the affinity A (X) of the antibodyS(i)) And generating a new population through three operations of cloning, mutation and selection.
(2) Cloning operation:
number of antibody replications n for each antibody in the replicating populationc(XS(i)) Depending on its affinity, i.e.:
wherein, Floor (. cndot.) represents rounding-down, nc(XS(i)) Represents the number of replications of the ith antibody.
(3) Mutation operation:
elements were randomly selected from each replicated antibody and replaced with equal amounts of the other candidate bands.
(4) Selecting operation:
the m antibodies with the highest affinity are selected from all antibodies to form a new population.
(5) Repeating (2) - (4) until the maximum change in affinity of the subset of candidate bands in the new population is less than the threshold τ.
(6) And taking the candidate waveband subset with the highest affinity in the population as the finally selected waveband subset.
Corresponding to the embodiment of the low-redundancy hyperspectral band selection method considering both representativeness and information content, the application also provides an embodiment of a low-redundancy hyperspectral band selection device considering both representativeness and information content, which includes:
the 3D reconstruction network training module is used for constructing a 3D reconstruction network and carrying out proper training on the 3D reconstruction network;
the wave band representativeness measuring module is used for measuring the representativeness of the candidate wave band subset to the original hyperspectral image;
the band redundancy measurement module is used for measuring the redundancy of the candidate band subset;
a band information metric module for measuring the amount of information contained in the subset of candidate bands;
the comprehensive evaluation index construction module is used for designing a wave band subset scoring function and evaluating candidate wave band subsets;
the band subset searching module is used for generating a considerable number of candidate band subsets, scoring the candidate band subsets and selecting the band subset with the highest score as a selected band subset;
and the wave band selection result output module is used for outputting the selected optimal wave band subset result.
In an implementation of the present invention, the band selection apparatus further includes an application module, and the application module performs hyperspectral image classification or target detection using the band selection result.
In an embodiment of the invention, the 3D reconstructed network training module includes:
the image blocking module is used for blocking the hyperspectral image, and each hyperspectral image block is used as a sample;
and the 3D reconstruction network module is used for constructing a 3D reconstruction network and properly training the constructed 3D reconstruction network.
In one embodiment of the present invention, the band subset search module is implemented based on an immune clone selection algorithm, and includes:
the initial population building module is used for obtaining an initial population;
the cloning module is used for cloning the antibodies in the population;
the variation module is used for performing variation operation on the antibodies in the population;
the selection module is used for selecting the antibodies in the population;
the iteration stop condition judgment module is used for repeating the cloning module, the variation module and the selection module and judging whether the iteration is stopped or not according to the condition of the generated new population;
and the waveband selection module is used for sequencing the affinities of the antibodies in the population and selecting the waveband subset corresponding to the antibody with the highest affinity as the selected waveband subset.
With regard to the apparatus in the above-described embodiments, the specific manner in which each unit or module performs operations has been described in detail in the embodiments related to the method, and will not be described in detail herein.
For the device embodiments, since they substantially correspond to the method embodiments, reference may be made to the partial description of the method embodiments for relevant points. The above-described embodiments of the apparatus are merely illustrative, and the band subset search module, which is a 3D reconstruction network training module and an immune clone selection algorithm, may or may not be physically separate. In addition, each functional module in the present invention may be integrated into one processing unit, or each module may exist alone physically, or two or more modules may be integrated into one unit. The integrated modules or units can be implemented in the form of hardware, or in the form of software functional units, so that part or all of the modules can be selected according to actual needs to achieve the purpose of the scheme of the application.
In order to verify the effect of the invention, experiments were performed on real hyperspectral images. Specific embodiments are illustrated below using the IndianPines dataset as an example, and the experiments are as follows:
the image adopted in this embodiment is an Indian Pines dataset, which is a hyperspectral image captured by an AVIRIS sensor, as shown in fig. 3, the size is 145 × 145 pixels, the water vapor absorption band and the band with low signal-to-noise ratio are removed, and the remaining 185 bands participate in the experiment.
In order to further verify the application effect of the invention, the results obtained by the method of the invention and other methods are used for pixel classification. In order to more intuitively compare the influence of various band selection methods on the accuracy of the downstream classification task, fig. 4 shows the variation curve of the overall classification accuracy of different band selection methods with the number of bands when an SVM classifier is used on an Indian Pines dataset, with the abscissa being the number of selected bands and the ordinate being the classification accuracy. As can be seen from fig. 4, for the SVM classifier, the specific example of the present invention can achieve a classification effect significantly better than other band selection methods when selecting different numbers of bands.
FIG. 5 shows the overall classification accuracy versus the number of bands for different band selection methods using the EPF-G-G classifier on an Indian Pines dataset, with the abscissa being the number of bands selected and the ordinate being the classification accuracy. As can be seen from fig. 5, when different numbers of bands are selected, for the EPF-G classifier, the classification accuracy obtained when different numbers of bands are selected is higher than that obtained by other band selection methods.
TABLE 1 Classification Performance (%) comparison of different band selection methods on SVM classifiers
| Method | OA | AA |
| 1.MVPCA | 64.81 | 50.83 |
| 2.LCMVBCC | 58.95 | 49.74 |
| 3.LCMVBCM | 66.90 | 60.98 |
| 4.ECA | 75.16 | 65.25 |
| 5.MR | 78.42 | 71.24 |
| 6.BS-Net-Conv | 78.91 | 72.27 |
| 7. The invention | 80.36 | 74.63 |
Table 1 gives the classification accuracy achieved when using SVM classifiers on Indian Pines datasets for different band selection methods, taking k 15 as an example. In the table, oa (overall accuracy) is the ratio of correctly classified samples to total samples, and aa (average accuracy) is the average of the accuracy obtained for each class. The results in table 1 show that the low-redundancy hyperspectral band selection method considering both representativeness and information content can improve the classification effect of band selection.
Specific embodiments are described to simplify the present disclosure. It is to be understood that the invention is not limited to the embodiments described and that various modifications thereof are possible without departing from the basic concept, and such equivalents are intended to fall within the scope of the invention as defined in the appended claims.
Claims (10)
1. A low-redundancy hyperspectral band selection method giving consideration to both representativeness and information quantity is characterized by comprising the following steps of:
step 1) constructing and training a 3D reconstruction network:
partitioning an original hyperspectral image, wherein each hyperspectral image block is used as a sample; taking the hyperspectral image blocks as the input of a 3D reconstruction network and training;
step 2) measuring band representativeness:
passing the hyperspectral image data block through a sparse binary mask, the calculation formula is expressed as:
wherein, mu belongs to [0,1 ]]LRepresenting a sparse binary mask indicating whether each band is contained in a subset of candidate bands, L representing the total number of bands,
representing multiplication symbols by band, XPRepresenting a hyperspectral image block;
rescaling the data passing through the mask; reconstructing original hyperspectral data by taking the rescaled data as the input of a trained 3D reconstruction network to obtain the representative measurement of a candidate waveband subset;
step 3), measuring the band redundancy; the band redundancy measurement method adopts Pearson correlation coefficient or vector subspace projection technology;
step 4), measuring the information content of the wave band; the wave band information measurement method adopts information divergence or information entropy;
step 5) constructing a comprehensive evaluation index giving consideration to band representativeness, redundancy and information content, wherein a calculation formula is represented as follows:
score(XS)＝-l-αR(XS)+βI(XS)
wherein α and β represent equilibrium coefficients, R (X)S) Representing a subset X of candidate bandsSRedundancy measure of, I (X)S) Representing a subset X of candidate bandsSOf the candidate band subset X, l represents the candidate band subset XSIs a representative metric of (c), score (X)S) Representing a subset X of candidate bandsSThe score of the comprehensive evaluation index;
step 6) searching the desired waveband subset:
calculating the comprehensive evaluation index score of each candidate waveband subset by using a subset search strategy of a grouping search algorithm, and searching the waveband subset with the highest score as the selected waveband subset; the grouping type search algorithm adopts an immune clone selection algorithm, a genetic algorithm or a particle swarm optimization algorithm.
2. The method for selecting hyperspectral bands with low redundancy and both representativeness and information content according to claim 1, wherein the rescaling in the step 2) is to scale the data after passing through the mask according to the ratio of the total number of bands to the number of selected bands, and the calculation formula is represented as:
wherein k represents the number of selected bands, YPRepresenting the rescaled data.
3. The method for selecting a low-redundancy hyperspectral band with consideration of representativeness and information content according to claim 1, wherein the step of reconstructing original hyperspectral data to obtain a representative measurement of a candidate band subset comprises the following steps:
reconstructing original hyperspectral data, wherein the formula is as follows:
wherein the content of the first and second substances,
representing a reconstructed hyperspectral image block, F (-) represents a 3D reconstruction network, and theta represents a trainable parameter in the 3D reconstruction network;
obtaining a representative metric of the candidate band subset, wherein the formula is as follows:
wherein n represents the number of samples,
representing the ith input hyper-spectral image block,
representing the ith reconstructed hyperspectral image block; i | · | purple windFRepresenting the F-norm and l representing a representative metric value.
4. The method for selecting a low-redundancy hyperspectral band with consideration to both representativeness and information content according to claim 1, wherein the redundancy metric value of the candidate band subset is calculated by using a Pearson correlation coefficient according to the following formula:
wherein R (X)S) Representing a subset X of candidate bandsSRedundancy measure of (x)s(i)Represents the ith band in the candidate band subset, | | | · | | represents a 2-norm, aboveThe corner mark T indicates transposition.
5. The method for selecting hyperspectral bands with low redundancy and both representativeness and information content according to claim 1, wherein the formula for calculating the information metric values of the candidate band subsets by using the information divergence is as follows:
wherein I (.) represents the information amount, xs(i)Representing the ith band, X, of a subset of the selected bandsSRepresents a subset of the candidate bands of the spectrum,
representative band xs(i)Normalized ith element, N represents pixel number, qiAccording to the band xs(i)The mean and variance of the first element in the probability distribution are initialized randomly to generate Gaussian distribution and normalized, and k represents the number of the selected bands.
6. The method for selecting a low-redundancy hyperspectral band with consideration to representativeness and information content according to claim 1, wherein an immune clone selection algorithm is used to search for a desired band subset, specifically:
(1) constructing an initial population:
dividing all wave bands in an original hyperspectral image into k groups according to wave band index serial numbers, randomly selecting one wave band from each group of wave bands to form an initial antibody, wherein each antibody represents a candidate wave band subset containing k wave bands; repeating the steps m times to generate an initial population consisting of m initial antibodies, and calculating the comprehensive evaluation index score of the candidate waveband subset corresponding to each antibody in the initial population;
(2) cloning operation:
number of antibody replications n for each antibody in the replicating populationc(XS(i)) The overall evaluation index score depending on it, namely:
wherein, Floor (·) represents rounding-down, nc(XS(i)) Represents the ith antibody XS(i)The number of copies of; score (X)S(i)) Represents the ith antibody XS(i)The score of the comprehensive evaluation index;
(3) mutation operation:
randomly selecting some elements from each replicated antibody and replacing them with equal amounts of other candidate bands;
(4) selecting operation:
selecting m antibodies with highest affinity from all antibodies to form a new population;
(5) repeating (2) - (4) until the maximum value of the comprehensive evaluation index scores of the candidate waveband subsets in the new population changes by less than a threshold value tau.
7. A low-redundancy hyperspectral band selection device taking into account representativeness and information quantity is characterized by being used for realizing the hyperspectral band selection method of claim 1; the hyperspectral band selection device comprises:
the 3D reconstruction network training module is used for constructing a 3D reconstruction network and training the 3D reconstruction network;
the wave band representativeness measuring module is used for measuring the representativeness of the candidate wave band subset to the original hyperspectral image;
the band redundancy measurement module is used for measuring the redundancy of the candidate band subset;
a band information metric module for measuring the amount of information contained in the subset of candidate bands;
the comprehensive evaluation index construction module is used for designing a wave band subset scoring function and evaluating candidate wave band subsets;
the optimal waveband subset searching module is used for generating a certain number of candidate waveband subsets, scoring the candidate waveband subsets and selecting the waveband subset with the highest score as a selected waveband subset;
and the wave band selection result output module is used for outputting the selected optimal wave band subset result.
8. The low-redundancy hyperspectral band selection apparatus considering both representativeness and information content of claim 7, wherein the band selection apparatus further comprises an application module, and the application module uses the band selection result to perform hyperspectral image classification or target detection.
9. The representative and information-quantity compatible low-redundancy hyperspectral band selection device according to claim 7, wherein the 3D reconstruction network training module comprises:
the image blocking module is used for blocking the hyperspectral image, and each hyperspectral image block is used as a sample;
and the 3D reconstruction network module is used for constructing a 3D reconstruction network and training the constructed 3D reconstruction network.
10. The apparatus according to claim 7, wherein the optimal band subset search module comprises:
the initial population building module is used for obtaining an initial population;
the cloning module is used for cloning the antibodies in the population;
the variation module is used for performing variation operation on the antibodies in the population;
the selection module is used for selecting the antibodies in the population;
the iteration stop condition judgment module is used for repeating the cloning module, the variation module and the selection module and judging whether the iteration is stopped or not according to the condition of the generated new population;
and the waveband selection module is used for sequencing the affinities of the antibodies in the population and selecting the waveband subset corresponding to the antibody with the highest affinity as the selected waveband subset.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111637324.2A CN114519769A (en) | 2021-12-29 | 2021-12-29 | Low-redundancy hyperspectral band selection method and device considering representativeness and information content |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111637324.2A CN114519769A (en) | 2021-12-29 | 2021-12-29 | Low-redundancy hyperspectral band selection method and device considering representativeness and information content |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114519769A true CN114519769A (en) | 2022-05-20 |
Family
ID=81596809
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111637324.2A Pending CN114519769A (en) | 2021-12-29 | 2021-12-29 | Low-redundancy hyperspectral band selection method and device considering representativeness and information content |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114519769A (en) |
-
2021
- 2021-12-29 CN CN202111637324.2A patent/CN114519769A/en active Pending
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Yang et al. | Deephoyer: Learning sparser neural network with differentiable scale-invariant sparsity measures | |
| Abiri et al. | Establishing strong imputation performance of a denoising autoencoder in a wide range of missing data problems | |
| CN110348399B (en) | Hyperspectral intelligent classification method based on prototype learning mechanism and multidimensional residual error network | |
| CN112052755B (en) | Semantic convolution hyperspectral image classification method based on multipath attention mechanism | |
| CN111191514A (en) | Hyperspectral image band selection method based on deep learning | |
| CN109190511B (en) | Hyperspectral classification method based on local and structural constraint low-rank representation | |
| CN107679539B (en) | Single convolution neural network local information and global information integration method based on local perception field | |
| CN111914728A (en) | Hyperspectral remote sensing image semi-supervised classification method and device and storage medium | |
| CN113902622B (en) | Spectrum super-resolution method based on depth priori joint attention | |
| CN112613536A (en) | Near infrared spectrum diesel grade identification method based on SMOTE and deep learning | |
| CN107301643A (en) | Well-marked target detection method based on robust rarefaction representation Yu Laplce's regular terms | |
| CN110991247B (en) | Electronic component identification method based on deep learning and NCA fusion | |
| CN114913379A (en) | Remote sensing image small sample scene classification method based on multi-task dynamic contrast learning | |
| CN115311502A (en) | Remote sensing image small sample scene classification method based on multi-scale double-flow architecture | |
| CN106250918A (en) | A kind of mixed Gauss model matching process based on the soil-shifting distance improved | |
| Zeng et al. | Solving OSCAR regularization problems by fast approximate proximal splitting algorithms | |
| CN113516019B (en) | Hyperspectral image unmixing method and device and electronic equipment | |
| CN111079011A (en) | Deep learning-based information recommendation method | |
| CN115641503A (en) | Hyperspectral band selection method and device based on Transformer network | |
| CN111738086A (en) | Composition method and system for point cloud segmentation and point cloud segmentation system and device | |
| CN109325140B (en) | Method and device for extracting hash code from image and image retrieval method and device | |
| CN114519769A (en) | Low-redundancy hyperspectral band selection method and device considering representativeness and information content | |
| CN115496950A (en) | Neighborhood information embedded semi-supervised discrimination dictionary pair learning image classification method | |
| CN113378979B (en) | Hyperspectral band selection method and device based on band attention reconstruction network | |
| CN112308151A (en) | Weighting-based classification method for hyperspectral images of rotating forest |
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 |