CN106960221A - A kind of hyperspectral image classification method merged based on spectral signature and space characteristics and system - Google Patents
A kind of hyperspectral image classification method merged based on spectral signature and space characteristics and system Download PDFInfo
- Publication number
- CN106960221A CN106960221A CN201710150041.2A CN201710150041A CN106960221A CN 106960221 A CN106960221 A CN 106960221A CN 201710150041 A CN201710150041 A CN 201710150041A CN 106960221 A CN106960221 A CN 106960221A
- Authority
- CN
- China
- Prior art keywords
- high spectrum
- space characteristics
- spectrum image
- spectral
- spectral signature
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/30—Noise filtering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The present invention provides a kind of hyperspectral image classification method merged based on spectral signature and space characteristics and system, belongs to remote sensing image classification field.The present invention includes carrying out space characteristics extraction to it according to the connectivity of region of high spectrum image；And feature denoising is carried out to spectral signature and space characteristics respectively；Then Fusion Features are carried out to spectral signature and space characteristics；The classification of unknown classification pixel on high spectrum image is predicted by disaggregated model again；Present invention additionally comprises space characteristics extraction module, feature denoising module, Fusion Features module and sort module.The beneficial effects of the invention are as follows：Low-rank expression and Fusion Features are carried out by the spectral signature to high spectrum image and space characteristics, the nicety of grading of high spectrum image is lifted, there is higher accuracy rate to the prediction of unknown classification pixel classification.
Description
Technical field
The invention belongs to Classification in Remote Sensing Image field, more particularly to a kind of EO-1 hyperion merged based on spectral signature and space characteristics
Image classification method and system.
Background technology
The data that high spectrum resolution remote sensing technique is collected have contained substantial amounts of information, can be by observed object by these information
Spectral characteristic closely connected with spatial geographic information, and the potential feature of many materials is by analyzing high-spectral data
Just it is easy to be mined out, these are the major reasons that high spectrum image is favored by increasing remote sensing expert.Bloom
Spectrum remote sensing technology not only obtains the spectral signature of earth object, while remaining observed object and other earth objects of surrounding again
Relation, produced image contained abundant information, and strong support is provided using data and analyze data for us.
The identification of ground target object and the prevailing scenario that classification is high-spectral data application.The major function of ground object target classification
It is the unique classification of pixel point prediction one to unknown classification on image, this classification represents the capped content in this block region,
Such as building, vegetation, crops etc., researcher or related practitioner can have found atural object from sorted image
The regularity of distribution.
For object identification and classification problem, traditional method is visual interpretation method.This method needs expertise and reality
Border geography information is investigated and the outstanding space associative ability of binding personnel constructs the geographical distribution of entire image.But simultaneously
Not all is engaged in the priori in terms of the researchers of remote sensing images has abundant remote sensing and geography, completes such one
This needs to spend substantial amounts of human and material resources and time for individual work, and this method is not a kind of wise selection；In addition, evaluation one
The information that width remote sensing images are reflected, it is only far from enough by visual interpretation.It is continuous with subjects such as remote sensing technology and computers
Intersect and merge, therefore the required precision classified to remote sensing images using computer technology also more and more higher, is improved or sought
It is one of study hotspot in the last few years to look for new Classifying Method in Remote Sensing Image.
High spectrum image is during image is formed, and can receive is influenceed by extraneous factor, such as：The irradiation of the sun
Solar radiation is gushed in angle, the angle of perceptron vision, atmospheric scattering, in the scene the secondary photograph of the light of adjacent objects reflection
It is bright, the shade of irradiation, the scattering and absorption of atmospheric reflectance radiation, the spatially and spectrally distortion of sensor.In the different time and
Under meteorological condition, the spectral signature of same geographic area may produce larger difference, so high spectrum image is used alone
Spectral signature can produce larger error as the foundation result of image classification.
The content of the invention
For defect present in prior art or deficiency, the technical problems to be solved by the invention are：One kind is provided to melt
The spectral signature of high spectrum image and the hyperspectral image classification method and system of space characteristics are closed, classification hyperspectral imagery is improved
Accuracy rate.
To achieve these goals, the technical scheme that the present invention takes is a kind of based on spectral signature and space characteristics to provide
The hyperspectral image classification method and system of fusion.
A. space characteristics extraction is carried out to high spectrum image according to the connectivity of region of high spectrum image；
B. the spectral signature and space characteristics to high spectrum image carry out feature denoising respectively；
C. Fusion Features are carried out to the spectral signature and space characteristics of high spectrum image by characteristic weighing；
D. unknown classification pixel on high spectrum image is classified by disaggregated model.
This method divides the image into fritter, and each fritter uses matrix low rank decomposition and combines the synchronization of dictionary matrix
Study, can go back the low-rank of original matrix, that is, reduce the pollution of noise, multiple features fusion is carried out by way of weighting, is obtained
Classification model construction is carried out in new feature, feeding support vector machine classifier, it is higher accurate to have for pixel class prediction
Rate.
As a further improvement on the present invention, expressed by the matrix low-rank based on high spectrum image region division to bloom
The spectral signature and space characteristics of spectrogram picture carry out feature denoising respectively.
As a further improvement on the present invention, the step A comprises the following steps：
A1. the size and attribute of attribute filter are selected；
A2. the attribute filter is slided on high spectrum image, changes the pixel value of attribute filter overlay area, and
The property value of computation attribute filter overlay area, the property value is average, variance, area or girth；
A3. { T }={ M is passed throughmin,Mmin+δ,Mmin+2δ,...,MmaxThreshold vector T is calculated, and compare property value a successively
() extracts the space characteristics of high spectrum image, wherein M with threshold vector TminFor the minimum value of average, MmaxFor the maximum of average
Value, δ is spaced to increase, δ=Mmean× 2.5%, MmeanFor the average of all averages, each classification in the high spectrum image
Under have the pixel of known class, the pixel of known class corresponding variance in each category is calculated respectively.
As a further improvement on the present invention, the step B comprises the following steps：
B1. by the characteristics of adjacent area pixel is more likely to belong to same atural object classification in high spectrum image, by phase
The pixel in neighbouring region is divided into a spatial group, and selects the square area division size of fixed size to be divided；
B2. dictionary matrix is fixed, low rank sparse matrix and noise matrix are calculated by method of Lagrange multipliers；
B3. fixed coefficient matrix and noise matrix, pass through coordinate descent (Block-Coordinate Desent
Approach, BCD) Dictionary of Computing matrix；
B4. judge whether to be less than set threshold value.
As a further improvement on the present invention, the step C includes comparing A=[Aspectral；Aspatial], A=
Aspectral+Aspatial, A=μ Aspectral+(1-μ)Aspatial, A=μ Aspectral+(1-μ)AspatialWithFive kinds of Feature Fusion Algorithms, and a kind of Feature Fusion Algorithm is selected wherein to high spectrum image
Spectral signature and space characteristics carry out Fusion Features.Pass through A=μ Aspectral+(1-μ)AspatialTo the spectrum of institute's high spectrum image
Feature and space characteristics carry out Fusion Features.The AspectralFor spectral signature, AspatialFor space characteristics, μ (0≤μ≤1) is
Weight coefficient, determines AspectralAnd AspatialWeight shared by two kinds of features.
As a further improvement on the present invention, the step D is to be also easy to produce Hughes's effect according to high spectrum image
(Hughes) the characteristics of, disaggregated model is set up using support vector machine classifier.
The classification hyperspectral imagery model system merged based on spectral signature and space characteristics, including：
Including space characteristics extraction module, the property for the connectivity of region according to high spectrum image extracts high-spectrum
The feature of picture so that possess belong to the same area tendency pixel it is more like, should not belong to and be same as the similitude in region not
It is similar；
Feature denoising module, passes through spectral signature and sky of the matrix low-rank expression based on region division to high spectrum image
Between feature carry out denoising, the low-rank part for extracting eigenmatrix；
Fusion Features module, the spectral signature and space characteristics of the high spectrum image to extracting are added using characteristic weighing
Carry out Fusion Features；
Sort module, disaggregated model is set up by support vector machine classifier, for predicting unknown class on high spectrum image
The classification of other pixel, the characteristics of sort module is also easy to produce Hughes's effect (Hughes) according to high spectrum image, selection branch
Vector machine sorting algorithm is held for last grader, i.e. support vector machine classifier, the support vector machine classifier sets up classification
Model, disaggregated model is classified to the pixel of unknown classification on high spectrum image.
As a further improvement on the present invention, the feature denoising module includes：
Region division, the square area for selecting fixed size divides size and the pixel of adjacent area is drawn
Point, it is divided into a spatial group；
Low rank sparse matrix is solved, fixed dictionary matrix solves low rank sparse matrix and low-rank noise matrix；
Dictionary matrix, fixed low rank sparse matrix and low-rank noise matrix are solved, Block-Coordinate is used
Descent Approach (BCD) Algorithm for Solving dictionary matrix.
As a further improvement on the present invention, the Fusion Features module is from A=[Aspectral；Aspatial], A=Aspectral
+Aspatial, A=μ Aspectral+(1-μ)Aspatial, A=μ Aspectral+(1-μ)AspatialWithThis five
A Feature Fusion Algorithm is selected in Feature Fusion Algorithm.The AspectralFor spectral signature, AspatialFor space characteristics, μ (0
≤ μ≤1) is weight coefficient, determines AspectralAnd AspatialWeight shared by two kinds of features.
As a further improvement on the present invention, the Fusion Features module passes through A=μ Aspectral+(1-μ)AspatialTo height
The spectral signature and space characteristics of spectrum picture carry out Fusion Features.
As a further improvement on the present invention, the sort module is also easy to produce Hughes's effect according to high spectrum image
(Hughes) the characteristics of, selection support vector cassification algorithm is last grader, i.e. support vector machine classifier, the support
Vector machine classifier sets up disaggregated model, and disaggregated model is predicted to the pixel of unknown classification on high spectrum image.
The beneficial effects of the invention are as follows：The present invention by high spectrum image by being divided into fritter, and each fritter uses matrix
Low-rank decomposition and combination dictionary matrix, go back the low-rank of original matrix, that is, reduce the pollution of noise；To through and being based on region division
Cross two kinds of low-rank expression algorithm process multiple features fusion is carried out by way of weighting summation, obtain new feature, send into
Classification model construction is carried out in support vector machine classifier, is experimentally confirmed, accuracy rate of the present invention for pixel class prediction
Up to 95%.
Brief description of the drawings
Fig. 1 is flow chart of the method for the present invention；
Wherein numeral is represented：A- steps A；B- steps B；C- steps C；D- steps D；
Embodiment
The present invention is further described for explanation and embodiment below in conjunction with the accompanying drawings.
As shown in figure 1, the present invention includes：
Space characteristics extraction module, the property for the connectivity of region according to high spectrum image extracts high spectrum image
Feature so that possess belong to the same area tendency pixel it is more like, should not belong to be same as region similitude it is dissimilar；
Feature denoising module, passes through spectral signature and sky of the matrix low-rank expression based on region division to high spectrum image
Between feature carry out denoising, the low-rank part for extracting eigenmatrix；
Fusion Features module, the spectral signature and space characteristics of the high spectrum image to extracting are added using characteristic weighing
Carry out Fusion Features；
Sort module, disaggregated model is set up by support vector machine classifier, for predicting unknown class on high spectrum image
The classification of other pixel；
A. space characteristics extraction is carried out to high spectrum image according to the connectivity of region of high spectrum image；Its specific steps is such as
Under：
A1. the size and attribute of attribute filter are selected, the attribute filter is the forms of a fixed dimension；
A2. the attribute filter is slided on high spectrum image, changes the pixel value of forms overlay area according to rule,
And the property value of computation attribute filter overlay area, the property value a () is average, variance, area or girth；
A3. { T }={ M is passed throughmin,Mmin+δ,Mmin+2δ,...,MmaxThreshold vector T is calculated, and compare property value successively
The space characteristics of high spectrum image, wherein M are extracted with threshold vector TminFor the minimum value of average, MmaxFor the maximum of average
Value, δ is spaced to increase, δ=Mmean× 2.5%, MmeanFor the average of all averages, each classification in the high spectrum image
Under have the pixel of known class, the pixel of known class in each classification corresponding side in each category is calculated respectively
Difference.
In space characteristics extraction module, slided by the slip forms of a fixed size on high spectrum image, according to
Rule changes the pixel value of the forms overlay area, and the slip forms are referred to as attribute filter (Attribute
Filters, AFs), the attributive character refers to the category according to specified by attribute filter (Attribute Filters, AFs)
Property, the value of computation attribute filter (Attribute Filters, AFs) overlay area, the value of the attribute can be average, side
Difference, area or girth etc.；The attribute a () that calculating is obtained is compared successively with threshold vector T, if a (Ci) ＞ Tk, then Ci
Attribute a (the C in the region of coveringi) constant, if a (Ci)≤Tk, then CiThe central value in the high spectrum image region of covering is adjusted
For the immediate gray value of close region, if gray value becomes big or diminished, referred to as thickening or thinning；Wherein
CiRepresent capped high spectrum image region, TkRepresent k-th of threshold value.
For capped region f, according to n given threshold value, the possible value of 2n+1 kinds is obtained：
AP (f)={ Φn(f),...,Φ1(f),f,γ1(f),...,γn(f)}
Feature Dimension Reduction, preceding C dimension are carried out to high spectrum image for high spectrum image same use feature dimension reduction method
Feature be used for calculate, the attributive character (Extended Attribute Profiles, EAPs) being expanded：
EAP={ AP (PC1),AP(PC2),...,AP(PCC)}
Furthermore it is also possible to which multiple attributes are joined together to use, it is assumed that there is m attribute, many attributive character being expanded
(Extended Multi-Attribute Profiles, EMAPs)：
Each ai, i=1 ..., m represents an attribute.
And threshold vector T computing formula is as follows：
{ T }={ Mmin,Mmin+δ,Mmin+2δ,...,Mmax}
Assuming that image has 5 passages, and 9 classifications, for each passage (i.e. each feature), its threshold vector T is
{ T }={ Mmin,Mmin+δ,Mmin+2δ,...,Mmax, obtain the corresponding variance of 9 classifications, M by calculatingminRepresent average most
Small value, MmaxThe maximum of average is represented, δ increases interval, δ=Mmean× 2.5%, MmeanIt is the average of all averages.
In feature denoising module, it is more likely to belong to same atural object classification by adjacent area pixel in high spectrum image
The characteristics of, the pixel of adjacent area is divided into well a spatial group, and select the square area of fixed size to divide size
Divided；Fixed dictionary matrix, low rank sparse matrix and noise matrix are solved by method of Lagrange multipliers；Fixed coefficient square
Battle array and noise matrix, word is solved by Block-Coordinate Desent Approach (BCD) algorithm, i.e. coordinate descent
Allusion quotation matrix.
Matrix low-rank expresses algorithm combination dictionary learning, can effectively represent each vectorial under new space in original matrix
Linear expression (these linear expressions constitute low-rank matrix), and constantly update by dictionary learning the vector basis in space, reach
Separate the purpose that sparse noise solves low-rank matrix；For high spectrum image, in high spectrum image, the pixel of adjacent area is more
Tend to belong to same atural object classification, if the present invention proposes the pixel of adjacent area being divided into a spatial group, a figure
As multiple area of space can be divided into, so as to propose the matrix low-rank expression algorithm based on region division.Specific solution is as follows：
Matrix low-rank expression model based on region division：
S.t.D=ZA+E, A >=0, | | zi||2≤1,zi≥0
Assuming that representing the matrix that high spectrum image all pixels point is arranged in matrix D, b represents the logical of high spectrum image
The dimension of road number, i.e. characteristic vector, n represents the number of pixel.{g1,g2,...,gmRepresent divide m area of space,
Assuming that selection square area is 5 × 5；The order of rank () representing matrix, | | | |00 norm of representing matrix.
Wherein Z is space vector base,Representing matrixCorresponding low-rank expression matrix；Restrictive condition illustrates low-rank square
Battle array A is necessary for nonnegative matrix, ziDictionary Z i-th of base vector is represented, is also non-negative, and the mould of base vector is less than or equal to
1.Because the data of real-life data, such as high spectrum image are all largely the numbers of non-negative, so matrix Z and matrix
A is nonnegative matrix.
Because optimal problem above is a np hard problem, so being converted to following form：
S.t.D=ZA+E, A >=0, | | zi||2≤1,zi≥0
Wherein | | | |*The nuclear norm of representing matrix, | | | |11 norm of representing matrix.
Solve low rank sparse matrix A and low-rank noise matrix E：
Dictionary matrix Z is fixed first, updates low rank sparse matrix A and low rank sparse matrix E, can by optimization problem above
Know, optimization problem can be decomposed into multiple subproblems, and each problem is solvedWithIn order to not lose one
As property, each optimize subproblem be expressed as form：
s.t.Dg=ZAg+Eg,Ag≥0
The problem of above-mentioned formula is described is a convex optimization problem, can use uncertain augmented vector approach
(Inexact ALM) Algorithm for Solving, adds matrix J in restrictive condition, and optimization problem changes into following form：
s.t.Dg=ZAg+Eg, J=Ag,J≥0
Solve dictionary matrix Z：
For dictionary learning step, target is to make original matrix D and dictionary matrix Z and low rank sparse matrix A product
Gap it is less and less, object function is as follows：
s.t.||zi||2≤1,||zi||2≥0
Formula Central Plains matrix D and low rank sparse matrix A are fixed, utilize iterative algorithm Block-Coordinate
Descent Approach (BCD) are solved, it is therefore an objective to restrain the value of optimization function, finally less than given threshold value,
The advantage of the algorithm is that need not preset any hyper parameter, it is not required that learning rate.
Fusion Features module, as described above, the space characteristics of high spectrum image are extracted before this, respectively to high spectrum image
Spectral signature and space characteristics carry out low-rank processing, and spectral signature and space characteristics then again to high spectrum image carry out feature
Fusion, it is comprised the following steps that：
A kind of spectral signature to high spectrum image and space characteristics are selected to carry out in following five kinds of Feature fusions
Fusion Features, this five kinds of Feature fusions are：
1. two parts feature is directly cascaded：
A=[Aspectral；Aspatial]
2. two parts feature correspondence is added：
A=Aspectral+Aspatial
3. two parts characteristic weighing is added：
A=μ Aspectral+(1-μ)Aspatial
4. by two parts feature is squared and evolution again：
5. two parts feature is multiplied evolution again：
Fusion Features mean that the information that two kinds of features are provided is combined into new feature all of getting up is used to set up
Disaggregated model, the method for five kinds of Fusion Features of the above is all considered comprehensively using two kinds of features, and calculates simple, versatility
It is high；When the dimension of two kinds of features is unequal, it is defined by the dimension of the small feature of dimension, it is unnecessary that the big feature of dimension is directly subtracted
Part.
Wherein in the third Feature fusion, μ (0≤μ≤1) represents weight coefficient, determines AspectralAnd Aspatial
Weight shared by two kinds of features, on μ determination, the present invention selects optimal by the way of 5 folding cross validations, eventually through reality
Test and determine optimal μ values, μ interval is [0,1.0].The third fusion method is finally selected as the side of Fusion Features
Method.
In sort module, selection SVMs (SVM) sets up disaggregated model as grader, because SVMs point
Class device can effectively mitigate Hughes's effect (Hughes), lift the nicety of grading of high spectrum image；Finally unknown classification is predicted again
The classification of pixel.
The hyperspectral image classification method merged based on spectral signature and space characteristics and system that the present invention is provided, in original
Space characteristics are extracted on the high spectrum image of beginning, this method is slided using attribute filter on each passage of high spectrum image
It is dynamic, attribute calculating is carried out to the region that each pixel is covered according to attribute filter, the space characteristics of the pixel are obtained
Value, finally obtains the corresponding characteristic vector of pixel；Because original image may be made an uproar during collection by a series of
The interference of sound, so, it is necessary to handle spectral signature and space characteristics after space characteristics are obtained, that is, reducing making an uproar for data
Sound and dimensionality reduction.The present invention has higher accuracy rate for the prediction of unknown classification pixel classification.
Above content is to combine specific preferred embodiment further description made for the present invention, it is impossible to assert
The specific implementation of the present invention is confined to these explanations.For general technical staff of the technical field of the invention,
On the premise of not departing from present inventive concept, some simple deduction or replace can also be made, should all be considered as belonging to the present invention's
Protection domain.
Claims (10)
1. a kind of hyperspectral image classification method merged based on spectral signature and space characteristics, it is characterised in that methods described
Comprise the following steps：
A. space characteristics extraction is carried out to high spectrum image according to the connectivity of region of high spectrum image；
B. the spectral signature and space characteristics to high spectrum image carry out feature denoising respectively；
C. Fusion Features are carried out to the spectral signature and space characteristics of high spectrum image by characteristic weighing；
D. unknown classification pixel on high spectrum image is classified by disaggregated model.
2. the hyperspectral image classification method merged according to claim 1 based on spectral signature and space characteristics, its feature
It is that the step A comprises the following steps：
A1. the size and attribute of attribute filter are selected；
A2. the attribute filter is slided on high spectrum image, changes the pixel value of attribute filter overlay area, and is calculated
The property value of attribute filter overlay area, the property value is average, variance, area or girth；
A3. { T }={ M is passed throughmin,Mmin+δ,Mmin+2δ,...,MmaxCalculate threshold vector T, and compare successively property value a () with
Threshold vector T extracts the space characteristics of high spectrum image, wherein MminFor the minimum value of average, MmaxFor the maximum of average, δ is
Increase interval, δ=Mmean× 2.5%, MmeanFor the average of all averages, have under each classification in the high spectrum image
The pixel of known class, calculates the pixel corresponding variance in each category of known class respectively.
3. the hyperspectral image classification method merged according to claim 1 based on spectral signature and space characteristics, its feature
It is, the step B to handle the light of high spectrum image respectively by the matrix low-rank expression based on high spectrum image region division
Spectrum signature and space characteristics, and feature denoising is carried out to it.
4. the hyperspectral image classification method merged according to claim 3 based on spectral signature and space characteristics, its feature
It is that the step B comprises the following steps：
B1. by the characteristics of adjacent area pixel is more likely to belong to same atural object classification in high spectrum image, by adjacent region
The pixel in domain is divided into a spatial group, and selects the square area division size of fixed size to be divided；
B2. dictionary matrix is fixed, low rank sparse matrix and noise matrix are calculated by method of Lagrange multipliers；
B3. fixed coefficient matrix and noise matrix, pass through coordinate descent Dictionary of Computing matrix；
B4. judge whether to be less than set threshold value.
5. the hyperspectral image classification method merged according to claim 1 based on spectral signature and space characteristics, its feature
It is that the step C includes：Compare A=[Aspectral；Aspatial], A=Aspectral+Aspatial, A=μ Aspectral+(1-μ)
Aspatial, A=μ Aspectral+(1-μ)AspatialWithFive kinds of Feature Fusion Algorithms, and select wherein
A kind of Feature Fusion Algorithm carries out Fusion Features, the A to the spectral signature and space characteristics of high spectrum imagespectralFor spectrum
Feature, AspatialFor space characteristics.
6. the hyperspectral image classification method merged according to claim 5 based on spectral signature and space characteristics, its feature
It is to pass through A=μ Aspectral+(1-μ)AspatialSpectral signature and space characteristics to high spectrum image carry out Fusion Features,
The AspectralFor spectral signature, AspatialFor space characteristics, μ (0≤μ≤1) is weight coefficient, determines AspectralWith
AspatialWeight shared by two kinds of features.
7. the hyperspectral image classification method merged according to claim 1 based on spectral signature and space characteristics, its feature
It is that the step D also includes, the characteristics of being also easy to produce Hughes's effect according to high spectrum image uses support vector machine classifier
Set up disaggregated model.
8. a kind of realize the classification hyperspectral imagery merged described in claim any one of 1-7 based on spectral signature and space characteristics
The system of model method, it is characterised in that including space characteristics extraction module, for the connectivity of region according to high spectrum image
Property extract high spectrum image feature so that possess belong to the same area tendency pixel it is more like, should not belong to
The similitude for being same as region is dissimilar；
Feature denoising module, it is special to the spectral signature of high spectrum image and space by the matrix low-rank expression based on region division
Carry out denoising is levied, the low-rank part for extracting eigenmatrix；
Fusion Features module, spectral signature and the space characteristics progress of the high spectrum image to extracting are added using characteristic weighing
Fusion Features；
Sort module, disaggregated model is set up by support vector machine classifier, for predicting unknown classification picture on high spectrum image
The classification of vegetarian refreshments, the characteristics of sort module is also easy to produce Hughes's effect according to high spectrum image selects support vector cassification
Algorithm is last grader, i.e. support vector machine classifier, and the support vector machine classifier sets up disaggregated model, disaggregated model
The pixel of unknown classification on high spectrum image is classified.
9. the classification hyperspectral imagery model system merged according to claim 8 based on spectral signature and space characteristics, its
It is characterised by, the feature denoising module includes：
Region division, the square area for selecting fixed size divides size and the pixel of adjacent area is divided,
It is divided into a spatial group；
Low rank sparse matrix is calculated, fixed dictionary matrix solves low rank sparse matrix and low-rank noise matrix；
Dictionary of Computing matrix, fixed low rank sparse matrix and low-rank noise matrix, dictionary matrix is solved using coordinate descent.
10. the classification hyperspectral imagery model system merged according to claim 8 based on spectral signature and space characteristics, its
It is characterised by, the Fusion Features module passes through A=μ Aspectral+(1-μ)AspatialTo the spectral signature and sky of high spectrum image
Between feature carry out Fusion Features, the AspectralFor spectral signature, AspatialFor space characteristics.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710150041.2A CN106960221A (en) | 2017-03-14 | 2017-03-14 | A kind of hyperspectral image classification method merged based on spectral signature and space characteristics and system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710150041.2A CN106960221A (en) | 2017-03-14 | 2017-03-14 | A kind of hyperspectral image classification method merged based on spectral signature and space characteristics and system |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN106960221A true CN106960221A (en) | 2017-07-18 |
Family
ID=59471333
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710150041.2A Pending CN106960221A (en) | 2017-03-14 | 2017-03-14 | A kind of hyperspectral image classification method merged based on spectral signature and space characteristics and system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106960221A (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109146003A (en) * | 2018-09-30 | 2019-01-04 | 合刃科技（深圳）有限公司 | A kind of classifying identification method based on high light spectrum image-forming, apparatus and system |
| CN109978838A (en) * | 2019-03-08 | 2019-07-05 | 腾讯科技（深圳）有限公司 | Image-region localization method, device and Medical Image Processing equipment |
| CN111263614A (en) * | 2017-10-24 | 2020-06-09 | 百合医疗科技株式会社 | Ultrasonic diagnostic system and ultrasonic diagnostic method |
| CN112465076A (en) * | 2021-01-28 | 2021-03-09 | 知行汽车科技（苏州）有限公司 | Multi-feature fusion method, device, system and storage medium |
| CN112633045A (en) * | 2019-10-09 | 2021-04-09 | 华为技术有限公司 | Obstacle detection method, device, equipment and medium |
| CN113688845A (en) * | 2021-08-19 | 2021-11-23 | 南京诺源医疗器械有限公司 | Feature extraction method and device suitable for hyperspectral remote sensing image and storage medium |
| CN113971747A (en) * | 2021-12-24 | 2022-01-25 | 季华实验室 | Raman spectrum data processing method, device, equipment and readable storage medium |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20130135609A1 (en) * | 2010-06-17 | 2013-05-30 | Chemimage Corporation | Targeted Agile Raman System for Detection of Unknown Materials |
| CN103971123A (en) * | 2014-05-04 | 2014-08-06 | 南京师范大学 | Hyperspectral image classification method based on linear regression Fisher discrimination dictionary learning (LRFDDL) |
| CN104584326A (en) * | 2012-05-09 | 2015-04-29 | 杜克大学 | Metamaterial devices and methods of using the same |
| US9052290B2 (en) * | 2012-10-15 | 2015-06-09 | Chemimage Corporation | SWIR targeted agile raman system for detection of unknown materials using dual polarization |
| CN105761234A (en) * | 2016-01-28 | 2016-07-13 | 华南农业大学 | Structure sparse representation-based remote sensing image fusion method |
| CN105894030A (en) * | 2016-04-01 | 2016-08-24 | 河海大学 | High-resolution remote sensing image scene classification method based on layered multi-characteristic fusion |
| CN106022391A (en) * | 2016-05-31 | 2016-10-12 | 哈尔滨工业大学深圳研究生院 | Hyperspectral image characteristic parallel extraction and classification method |
-
2017
- 2017-03-14 CN CN201710150041.2A patent/CN106960221A/en active Pending
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20130135609A1 (en) * | 2010-06-17 | 2013-05-30 | Chemimage Corporation | Targeted Agile Raman System for Detection of Unknown Materials |
| CN104584326A (en) * | 2012-05-09 | 2015-04-29 | 杜克大学 | Metamaterial devices and methods of using the same |
| US9052290B2 (en) * | 2012-10-15 | 2015-06-09 | Chemimage Corporation | SWIR targeted agile raman system for detection of unknown materials using dual polarization |
| CN103971123A (en) * | 2014-05-04 | 2014-08-06 | 南京师范大学 | Hyperspectral image classification method based on linear regression Fisher discrimination dictionary learning (LRFDDL) |
| CN105761234A (en) * | 2016-01-28 | 2016-07-13 | 华南农业大学 | Structure sparse representation-based remote sensing image fusion method |
| CN105894030A (en) * | 2016-04-01 | 2016-08-24 | 河海大学 | High-resolution remote sensing image scene classification method based on layered multi-characteristic fusion |
| CN106022391A (en) * | 2016-05-31 | 2016-10-12 | 哈尔滨工业大学深圳研究生院 | Hyperspectral image characteristic parallel extraction and classification method |
Non-Patent Citations (3)
| Title |
|---|
| 张春森等: "高光谱影像光谱-空间多特征加权概率融合分类", 《测绘学报》 * |
| 张静妙等: "基于低秩字典学习的高光谱遥感图像去噪", 《控制工程》 * |
| 马秀丹等: "一种优化的高光谱图像特征提取方法", 《信息技术》 * |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111263614A (en) * | 2017-10-24 | 2020-06-09 | 百合医疗科技株式会社 | Ultrasonic diagnostic system and ultrasonic diagnostic method |
| CN109146003A (en) * | 2018-09-30 | 2019-01-04 | 合刃科技（深圳）有限公司 | A kind of classifying identification method based on high light spectrum image-forming, apparatus and system |
| CN109146003B (en) * | 2018-09-30 | 2022-06-07 | 合刃科技（深圳）有限公司 | Hyperspectral imaging-based classification identification method, device and system |
| CN109978838A (en) * | 2019-03-08 | 2019-07-05 | 腾讯科技（深圳）有限公司 | Image-region localization method, device and Medical Image Processing equipment |
| CN110458813A (en) * | 2019-03-08 | 2019-11-15 | 腾讯科技（深圳）有限公司 | Image-region localization method, device and Medical Image Processing equipment |
| CN109978838B (en) * | 2019-03-08 | 2021-11-30 | 腾讯科技（深圳）有限公司 | Image area positioning method and device and medical image processing equipment |
| CN112633045A (en) * | 2019-10-09 | 2021-04-09 | 华为技术有限公司 | Obstacle detection method, device, equipment and medium |
| CN112465076A (en) * | 2021-01-28 | 2021-03-09 | 知行汽车科技（苏州）有限公司 | Multi-feature fusion method, device, system and storage medium |
| CN113688845A (en) * | 2021-08-19 | 2021-11-23 | 南京诺源医疗器械有限公司 | Feature extraction method and device suitable for hyperspectral remote sensing image and storage medium |
| CN113971747A (en) * | 2021-12-24 | 2022-01-25 | 季华实验室 | Raman spectrum data processing method, device, equipment and readable storage medium |
| CN113971747B (en) * | 2021-12-24 | 2022-04-26 | 季华实验室 | Raman spectrum data processing method, device, equipment and readable storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106960221A (en) | A kind of hyperspectral image classification method merged based on spectral signature and space characteristics and system | |
| US11783569B2 (en) | Method for classifying hyperspectral images on basis of adaptive multi-scale feature extraction model | |
| Shi et al. | Incorporating spatial information in spectral unmixing: A review | |
| CN103729848B (en) | High-spectrum remote sensing small target detecting method based on spectrum saliency | |
| Coburn et al. | A multiscale texture analysis procedure for improved forest stand classification | |
| CN102646200B (en) | Image classifying method and system for self-adaption weight fusion of multiple classifiers | |
| CN106503739A (en) | The target in hyperspectral remotely sensed image svm classifier method and system of combined spectral and textural characteristics | |
| CN110084159A (en) | Hyperspectral image classification method based on the multistage empty spectrum information CNN of joint | |
| Wang et al. | Spatio-temporal spectral unmixing of time-series images | |
| CN105427309A (en) | Multiscale hierarchical processing method for extracting object-oriented high-spatial resolution remote sensing information | |
| CN107392130A (en) | Classification of Multispectral Images method based on threshold adaptive and convolutional neural networks | |
| CN110728197B (en) | Single-tree-level tree species identification method based on deep learning | |
| Bonnet et al. | Comparison of UAS photogrammetric products for tree detection and characterization of coniferous stands | |
| CN108830312B (en) | Integrated learning method based on sample adaptive expansion | |
| CN104331698A (en) | Remote sensing type urban image extracting method | |
| CN105469098A (en) | Precise LINDAR data ground object classification method based on adaptive characteristic weight synthesis | |
| CN110309780A (en) | High resolution image houseclearing based on BFD-IGA-SVM model quickly supervises identification | |
| CN110390255A (en) | High-speed rail environmental change monitoring method based on various dimensions feature extraction | |
| CN104881868A (en) | Method for extracting phytocoenosium spatial structure | |
| Xu et al. | A new spectral-spatial sub-pixel mapping model for remotely sensed hyperspectral imagery | |
| CN112258523B (en) | Method for finely extracting enteromorpha coverage information of medium-low resolution remote sensing image | |
| CN110363246A (en) | A kind of fusion method of high-spatial and temporal resolution vegetation index NDVI | |
| CN110060273A (en) | Remote sensing image landslide plotting method based on deep neural network | |
| CN107895136A (en) | A kind of colliery area recognizing method and system | |
| CN112766161B (en) | Hyperspectral target detection method based on integrated constraint multi-example learning |
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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20170718 |