CN105389597B - A kind of more sorting techniques of high-spectral data based on Chernoff distances and SVM - Google Patents

A kind of more sorting techniques of high-spectral data based on Chernoff distances and SVM Download PDF

Info

Publication number
CN105389597B
CN105389597B CN201510969347.1A CN201510969347A CN105389597B CN 105389597 B CN105389597 B CN 105389597B CN 201510969347 A CN201510969347 A CN 201510969347A CN 105389597 B CN105389597 B CN 105389597B
Authority
CN
China
Prior art keywords
chernoff
class
distances
classification
svm
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201510969347.1A
Other languages
Chinese (zh)
Other versions
CN105389597A (en
Inventor
张淼
沈飞
林喆祺
沈毅
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Harbin Institute of Technology
Original Assignee
Harbin Institute of Technology
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Harbin Institute of Technology filed Critical Harbin Institute of Technology
Priority to CN201510969347.1A priority Critical patent/CN105389597B/en
Publication of CN105389597A publication Critical patent/CN105389597A/en
Application granted granted Critical
Publication of CN105389597B publication Critical patent/CN105389597B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

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

Landscapes

  • Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Theoretical Computer Science (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Artificial Intelligence (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
  • Sorting Of Articles (AREA)
  • Investigating Or Analysing Materials By Optical Means (AREA)

Abstract

The invention discloses a kind of more sorting techniques of the high-spectral data based on Chernoff distances and SVM, and its step are as follows：One, input data is pre-processed, obtains normalization data；Two, the Chernoff distances between any two classification are calculated, Chernoff distance matrixs are obtained；Three, it determines the execution sequence list under more classification task OAA strategies, obtains each wave band and full wave dissociable basis；Four, sub-classifier of the structure based on Chernoff distances instructs coefficient；Five, classification task determined by entire OAA strategies is executed using the weighed SVM grader based on Chernoff distances, until obtaining the final single category attribute of each test sample.The present invention not only increases the accuracy of traditional SVM methods, and is greatly improved to the nicety of grading of small sample classification, is suitable for the high spectrum image application of pattern recognition based on OAA classification policies.

Description

A kind of more sorting techniques of high-spectral data based on Chernoff distances and SVM
Technical field
The invention belongs to area of pattern recognition, are related to a kind of the SVM based on Chernoff distances optimizing training data points Class method.
Background technology
Hyperspectral image data is usually made of hundreds of continuously distributed wave bands.Entire data are regarded as three dimensions Degrees of data cuboid, wherein bidimensional determine that position of the object in plane of vision, the third dimension determine object in spectral wavelength Position.By taking AVIRIS (Airborne Visible/Infrared Imaging Spectrometer) high-spectral data as an example, Normally only it is separated by 10nm or so between two neighboring wave band.Since correlation is very strong between the wave band of image space adjacent band, So that conventional sorting methods must carry out just continuing with data after dimensionality reduction, common method is waveband selection.Because of core Method (kernel method) is influenced very little by the high dimension of the input space, so more and more researchers select kernel method. Regardless of whether selection waveband selection, kernel method all have excellent classification performance, such as our very familiar support vector machines (Support Vector Machines, SVM) grader.But few researchs are dedicated to extension SVM methods and keep it more suitable Close more classification applications of high-spectral data.
Classification error probability is the best quantitive measure of characteristic validity in pattern-recognition, and the dreamboat of feature selecting makes to reach Classification error probability is minimum.But this point is often difficult to accomplish.Therefore Upper bound of error probability minimum is often a kind of rational selection. The Upper bound of error probability that Chernoff is proposed is minimum, claims the upper bounds Chernoff.The upper bounds Chernoff can effectively improve list The nicety of grading of a core grader, and have certain directive function for the classification policy of core grader.
Since SVM can be only done two classification tasks in itself, the more classification applications of typical case of high-spectral data generally require Multi-categorizer is built by multiple SVM and certain strategy.SVM is changed currently with high-spectral data third dimension information Into work be concentrated mainly on following two aspects：First, being filtered to EO-1 hyperion input data, second is that generating customization The kernel function of change.But these methods are all just for unified SVM.For setting up each SVM of multi-categorizer, previous research is all Take the simple scheme unanimously treated.More accurate method is to take a kind of more favorably more classification schemes, i.e. basis The characteristics of two class object handled by each sub-classifier, carries out it independent customization of kernel function.On more classification policies, extensively Application is two kinds of strategies of OAA (One-Against-All, one-to-many) and OAO (One-Against-One, one-to-one), they Respectively there are advantage and disadvantage, the former sub-classifier negligible amounts, but the training time of each sub-classifier is longer, the sub-classifier number of the latter Measure more, but the training time of each sub-classifier is shorter.And in terms of nicety of grading, the two by difference after parameter optimization simultaneously Less.
Invention content
It is an object of the invention to propose a kind of improved svm classifier method based on Chernoff distances, pass through introducing The distance weighted matrixes of Chernoff so that grader makes full use of information between the class of sample during training, and utilizes Chernoff distances are given OAA policy class sequential selections and are instructed, and the accuracy of traditional SVM methods is not only increased, and And be greatly improved to the nicety of grading of small sample classification, it is suitable for the high spectrum image pattern based on OAA classification policies and knows It does not apply.
The purpose of the present invention is what is be achieved through the following technical solutions：
A kind of more sorting techniques of high-spectral data based on Chernoff distances and SVM, including following five steps：
One, input data is pre-processed, obtains normalization data；
Two, the Chernoff distances between any two classification are calculated, Chernoff distance matrixs are obtained；
Three, it determines the execution sequence list under more classification task OAA strategies, obtains each wave band and full wave separability Measurement；
Four, sub-classifier of the structure based on Chernoff distances instructs coefficient；
Five, classification determined by entire OAA strategies is executed using the weighed SVM grader based on Chernoff distances to appoint Business, until obtaining the final single category attribute of each test sample.
Compared with prior art, the present invention having the following advantages that：
1, in traditional svm classifier method, the distance weighted matrixes of Chernoff are introduced so that grader is in trained mistake Information between the class of sample is made full use of in journey, so as to improve classification effectiveness.Since this method is adopted in the optimization of each sub-classifier It has taken and has had more targetedly way, it is thus possible to the nicety of grading of each sub-classifier has been effectively improved, to improve more classification The final classification precision of device.
2, it is utilized compared with traditional OAA classification policies, in this method and sequence is executed to each sub-classifier under OAA strategies It gives and instructs, measured particular by the separability of the collectivity being calculated using Chernoff to determine the execution sequence list, this It allows for the good classification of separability to separate from test sample first, the bad classification of separability is then successively placed on last progress It distinguishes.Moreover, when each sub-classifier is classified, used weighting coefficient, i.e. guidance to different types of areas coefficient are all roots Carry out COMPREHENSIVE CALCULATING according to remaining classification, the classification separated is there is no considering, this also enhances the specific aim of weighting classification, To which the error in classification of each sub-classifier be further decreased.
Description of the drawings
Fig. 1 is the flow chart of the present invention；
Fig. 2 is hyperspectral image data original graph；
Fig. 3 is hyperspectral image data label figure；
Fig. 4 is standard SVM kernel function classifying quality figures；
Fig. 5 is the kernel function classifying quality figure based on Chernoff distances.
Specific implementation mode
Technical scheme of the present invention is further described below in conjunction with the accompanying drawings, however, it is not limited to this, every to this Inventive technique scheme is modified or replaced equivalently, and without departing from the spirit of the technical scheme of the invention and range, should all be covered In protection scope of the present invention.
Specific implementation mode one：Present embodiments provide for a kind of high-spectral data based on Chernoff distances and SVM More sorting techniques calculate Chernoff distances and the profit between any two classification by pretreatment by input data normalization It instructs OAA (One-Against-All, one-to-many) policy class sequence to improve nicety of grading with dissociable basis, utilizes Chernoff distance vectors are weighted single sub-classifier, and weighed SVM grader is then used for multiple times to sample classification.
As shown in Figure 1, being divided into five steps, it is as follows：
Step 1：Training sample and test sample are set.
1) for the multi-spectral remote sensing image of shootingWherein Row, Column indicate multi-spectral remote sensing image Width and length, B indicate multi-spectral remote sensing image wave band number, if xz∈RB, z=1,2 ..., N are B dimension samples, yz∈ 1, 2 ..., L } be and xzRelevant class label, wherein N are numbers of samples, and L is class number.
2) same category of pixel is all brought together, to involved in all pixels arrive class categories pixel all this Sample processing.
3) grouping coefficient Group assignment, while the broken number of " a few folding cross validations " that the coefficient is namely often said are given.
4) original sample is divided into training sample and test sample, respectively use two-dimensional matrix TrainSamples and TestSamples indicates that column vector corresponds to the spectral information data of each wave band of single pixel, wherein training sample accounts for gross sample This 1/Group, remaining (1-Group)/Group residue sample are test sample.
5) training sample and test sample are normalized.
Step 2：Chernoff distances between calculating two is different classes of.
1) two different classifications are selected, pth class and q classes are set as.
If npAnd nqThe number of pth class and q class samples respectively in training sample.For the kth -1 of sample, k and k+1 Three wave bands, the data of these three wave bands are normalized first, then by the pth class and q on these three wave bands The data of class sample are stored to two temporary variable matrix Ds according to the form of column vectorpAnd DqIn, wherein DpFor npThe square of × 3 dimensions Battle array, DqFor nqThe matrix of × 3 dimensions.
2) D is calculatedpWith DqThe mean value of each row obtains two 1 × 3 Mean Matrix MeanpWith Meanq。
3) D is calculatedpWith DqCovariance, obtain two 3 × 3 covariance matrix CovpWith Covq。
4) Chernoff distances are carried out to all classifications different between two to calculate：
Wherein, Ωk Q, qFor the Chernoff distances between pth class and q classes, k is current band number, MeanpWith Meanq The respectively Mean Matrix of pth class and q classes on tri- kth -1, k and k+1 wave bands, CovpWith CovqRespectively pth class and Covariance matrix of the q classes on tri- kth -1, k and k+1 wave bands, β are Chernoff apart from adjustment parameter, and 0 ＜ β ＜ 1.It is more Secondary change β can be that grader finds suitable upper error.
5) pass through 4) the middle Chernoff distance structure Chernoff distance matrixs calculated：
Since Chernoff distances are two parameters between different classes of, so the element on diagonal line is 0, CP, qFor B × 1 vector,P={ 1,2 ..., L }, q={ 1,2 ..., L } and p ≠ q.By pth class The distance between q classes are merged into a distance with the distance between q classes and pth class.
6) the is repeated 1) to the 5) step, until all dimensions to training sample any two classification all carry out Until the calculating process of Chernoff distances.
Step 3：Determine the execution sequence list under more classification task OAA strategies.
More classification tasks consider the case where classification sum is more than or equal to 3.Different from common OAA strategies, this method utilizes Chernoff distances instruct OAA classified orders.
2) willIn each wave band be added, obtain pth class on all wave bands to the separability of the collectivity of other all categories Measurement
3) separability of the collectivity of all categories is measured and is ranked sequentially by what numerical values recited was successively decreased, then holding under OAA strategies Row sequence list is that this puts in order, and uses ordered set<Class1, Class2... Classl..., ClassL>It indicates, wherein Classl∈ { 1,2 ..., L }, l=1,2 ..., L.
Step 4：Sub-classifier of the structure based on Chernoff distances instructs coefficient.
After the execution sequence under OAA strategies determines, we can (each subclassification be by L sub-classifier Two graders) complete the classification task of all L classes, specific execution sequence is：Pass through Class first1With all remaining classifications Two classification determine in test sample that label sequence number is equal to Class1Sample, then pass through Class2With all residue classes (attention has not contained Class at this time1Classification) two classification determine in test sample that label sequence number is equal to Class2's Sample executes successively, the last one two grader will determine that ClassL-1And ClassLClassification ownership.
Each specific sub-classifier is used for containing there are one the sub-classifiers based on Chernoff distances to instruct coefficient The classification performance of each sub-classifier is improved, the higher classification results of confidence rate are obtained, which instructs the calculating of coefficient Journey is specific as follows：
1) the classification Class for classifying at first1If Class1=i, wherein i ∈ { 1,2 ..., L }, i.e. Class1It is corresponding Classification be the i-th class, the i-th row being located in Chernoff matrixes, then sub-classifier instruct the coefficient to be：
Wherein, CI, qFor the i-th row in Chernoff matrixes, the element of q row.
3) finally to ClassL-1And ClassLClassification only need to carry out a subseries.
Step 5：It is executed using the weighed SVM grader based on Chernoff distances and is classified determined by entire OAA strategies Task, until obtaining the final single category attribute of each test sample.
1) suitable kernel function is selected first.Since RBF is checked in the more adaptable of numerous types of data, and to OAA The larger classification problem of quantity variance that a kind of classification and plurality of classes are be easy to cause under strategy has good result of calculation, so The present invention selects RBF cores as the kernel function of SVM classifier；Meanwhile we be obtained using Chernoff distances before in order to incite somebody to action The guidance to different types of areas coefficient for pth classIt is introduced into each sub-classifier, it would be desirable to using the improved RBF cores letter of weighting Number：
Wherein σ is width constant, and σ is smaller, and function more has selectivity.
In order to ensure that the correct execution of SVM sub-classifiers, the sample data that input is also coped with after weighting are normalized Processing.
2) according to classification policy determined by step 3, each step all uses constructed based on Chernoff in step 1) The weighed SVM grader of distance goes to complete training and test.
3) execute step 2) repeatedly, until classification task determined by all OAA strategies all after the completion of, that is, each survey Sample originally all by final decision is single category attribute, at this moment can entirely be terminated for more classification tasks of test sample.
Specific implementation mode two：Present embodiment is by improved based on the SVM approach applications of Chernoff distances to bloom It composes in image data assorting process, nicety of grading is further increased in conjunction with OAA strategies.
The description of hyperspectral image data is provided first：
Experimental subjects is high-spectrum sheet data.The data include 520 continuous wave bands, and experimental subjects is EO-1 hyperion picture Data.The data include 75x75 pixel, and each pixel includes 520 continuous wave bands, wave-length coverage 400nm-1000nm, Spectral resolution is 8nm.
Present embodiment has chosen 4 classes of pixel number at most in shooting picture and refers to table 1 as experiment sample.
The corresponding experiment sample number of all categories of table 1
Mark Classification Sample number
A Sky background 4369
B Tree 1027
C Cloud 222
D Aircraft 7
Fig. 2 is the original image of high-spectral data, wherein sky, and tree and cloud occupy the overwhelming majority；Aircraft is in whole picture figure In only account for several pixels.Fig. 3 is the label figure of sample.During labelling, due to the relationship of resolution ratio, treetop A part be denoted as sky.But since the spectral differences of treetop and sky are enough away from larger and two classes sample points, so This part has little effect the effect of classification, and can be as the verification of classification results.
Present embodiment takes the classification experiments that two kinds of verification high-spectral data set carry out 4 classes respectively, they are marks Quasi- RBF cores and weighting RBF cores based on Chernoff distances.Both kernel functions all construct multi-categorizer by OAA strategies, 3 sub-classifiers are needed for the classification experiments of 4 class atural objects.
Execute step 1：Input hyperspectral image dataWith corresponding label, if need calculate Chernoff distances Wave band B be 520.Experiment packet coefficient selects Group=2, i.e. 2 folding cross validations.Class number L=4.To input data It is normalized, gives training sample label corresponding with its classification.
Execute step 2：β numerical value is typically chosen during Chernoff distances calculateFor smaller or The β value of person's bigger, since the complexity of calculating and Singular Value problem that may be present are not considered.This method selectsThe different classes of Chernoff distances of each wave band any two are calculated, Matrix C hernoff distance matrixs C is obtained.
Execute step 3：Calculate the separability of the collectivity measurement of each classificationThe separability of the collectivity of all categories is measuredBy being ranked sequentially of successively decreasing of numerical values recited sequence is executed as under OAA strategies.Obtain the execution sequence list under OAA strategies< Class1, Class2, Class3, Class4>。
Execute step 4：It calculates each sub-classifier based on Chernoff distances and instructs coefficientI ∈ { 1,2,3,4 }.
Execute step 5：Each sub-classifier is weighted using guidance to different types of areas coefficient, then training sample is normalized again And svm classifier is carried out to it.Since RBF has preferable parameter adaptation in hyperspectral data processing field, we take Penalty factor is 100, and parameter σ is 0.4.
Addition standard SVM kernel functions are tested as a comparison.It keeps input sample constant, single sub-classifier is not added Power, and OAA strategies are that category is classified successively.
Conclusion：Contrast test the results are shown in Table 2.More sorting techniques based on Chernoff distances are compared with standard SVM methods Average nicety of grading have 10% or so promotion, overall accuracy also has 2% promotion, but the latter relative to the former promotion simultaneously It is not fairly obvious.From the point of view of supporting vector sum, the supporting vector number of improved kernel method is slightly more than the kernel function of script, it is seen that More sorting techniques based on Chernoff distances can make trained process slightly more complexization, but training time and testing time subtract It is few.Although improving the product calculation that kernel method both increases each element in kernel function weighting coefficient, either instruct Practice or classify, the time that improved kernel method is consumed all can be less than standard SVM methods.But calculate Chernoff weightings The time of matrix is much larger than training and the time classified, and time of this part is not included in.So this comprehensive two side The time of face factor, this method actual consumption is greater than standard SVM methods.
The comparison of nicety of grading, supporting vector sum and time loss when table 2 is using RBF cores
Parameter index Standard SVM Chernoff is apart from kernel method
Mean accuracy 55.8688 65.0868
Overall accuracy 88.0697 90.1323
Supporting vector number 4046 4634
Training time 16.8302 9.5793
Testing time 26.4759 23.0401
Fig. 4 is the classification results figure of standard SVM kernel functions, and Fig. 5 is classification results figures of the CHernoff apart from kernel function.By We are it can be found that for sky, for tree and three classifications of cloud, since the quantity of training sample is enough, so classification in figure Effect difference it is little, the difference only put individually.However for aircraft, since sample point itself is considerably less, so Trained difficulty is very big.And for the SVM methods of standard, more points based on Chernoff distances of this method proposition The classifying quality of class method is more preferable.

Claims (5)

1. a kind of more sorting techniques of high-spectral data based on Chernoff distances and SVM, it is characterised in that the method step It is as follows：
One, input data is pre-processed, obtains normalization data；
Two, the Chernoff distances between any two classification are calculated, obtain Chernoff distance matrixs, the Chernoff away from It is as follows from computational methods：
Wherein, Ωk P, qFor the Chernoff distances between pth class and q classes, k is current band number, MeanpWith MeanqRespectively For the Mean Matrix of pth class and q classes on tri- kth -1, k and k+1 wave bands, CovpWith CovqRespectively pth class and q classes Covariance matrix on tri- kth -1, k and k+1 wave bands, β are Chernoff apart from adjustment parameter, and 0 ＜ β ＜ 1；
Chernoff distance matrix computational methods are as follows：
CP, qFor the vector of B × 1, B indicates the wave band number of multi-spectral remote sensing image,P={ 1,2 ..., L }, q={ 1,2 ..., L } and p ≠ q；
Three, it determines the execution sequence list under more classification task OAA strategies, obtains each wave band and full wave separability degree Amount；
Four, sub-classifier of the structure based on Chernoff distances instructs coefficient, is as follows：
1) the classification Class for classifying at first1If Class1=i, wherein i ∈ { 1,2 ..., L }, L is class number, i.e., Class1Corresponding classification is the i-th class, the i-th row being located in Chernoff matrixes, then sub-classifier instructs the coefficient to be
2) according to the execution sequence list under OAA strategies<Class1, Class2..., Classl..., ClassL>, proceed to ClasslWhen, if Classl=j, wherein j ∈ { 1,2 ..., L }, when calculating the Chernoff distances of class categories to Measure CP, qIt forecloses, then sub-classifier instructs the coefficient to be
3) finally to CIassL-1And ClassLClassification only need to carry out a subseries；
Five, classification task determined by entire OAA strategies is executed using the weighed SVM grader based on Chernoff distances, directly To obtaining the final single category attribute of each test sample.
2. the more sorting techniques of the high-spectral data according to claim 1 based on Chernoff distances and SVM, feature exist It is as follows in the step 1：
1) for the multi-spectral remote sensing image of shootingWherein Row, Column indicate multi-spectral remote sensing image width and Long, B indicates the wave band number of multi-spectral remote sensing image, if xz∈RB, z=1,2 ..., N are B dimension samples, yz∈ 1,2 ..., L } be and xzRelevant class label, wherein N are numbers of samples, and L is class number；
2) same category of pixel is all brought together, the pixel to arriving class categories involved in all pixels all carries out together Sample processing；
3) grouping coefficient Group assignment is given；
4) original sample is divided into training sample and test sample, respectively use two-dimensional matrix TrainSamples and TestSamples indicates that column vector corresponds to the spectral information data of each wave band of single pixel；
5) training sample and test sample are normalized.
3. the more sorting techniques of the high-spectral data according to claim 1 based on Chernoff distances and SVM, feature exist It is selected in the β numerical value
4. the more sorting techniques of the high-spectral data according to claim 1 based on Chernoff distances and SVM, feature exist It is as follows in the step 3：
1) it sums, is obtained on each wave band to other institutes to the Chernoff distance vectors of pth class training sample and other classifications There are the Chernoff distance vectors of classificationAs the coefficient that instructs of sub-classifier, pth class is instructed on each wave band Practice sample and the Chernoff distances of other classifications are
2) willIn each wave band be added, obtain pth class and the separability of the collectivity of other all categories measured on all wave bands
3) separability of the collectivity of all categories is measured and is ranked sequentially by what numerical values recited was successively decreased, then the execution under OAA strategies is suitable Sequence table is that this puts in order, and uses ordered set<Class1, Class2... Classl..., ClassL>It indicates, wherein Classl ∈ { 1,2 ..., L }, l=1,2 ..., L.
5. the more sorting techniques of the high-spectral data according to claim 1 based on Chernoff distances and SVM, feature exist It is as follows in the step 5：
1) selection weights improved RBF kernel functionsAs the weighed SVM grader based on Chernoff distances Kernel function, whereinFor the guidance to different types of areas coefficient for pth class, and it is normalized after weighting；
2) according to classification policy determined by step 3, using the weighting based on Chernoff distances constructed in step 1) SVM classifier goes to complete training and test；
3) execute step 2) repeatedly, until classification task determined by all OAA strategies all after the completion of, it is at this moment entire for test More classification tasks of sample can terminate.
CN201510969347.1A 2015-12-22 2015-12-22 A kind of more sorting techniques of high-spectral data based on Chernoff distances and SVM Active CN105389597B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201510969347.1A CN105389597B (en) 2015-12-22 2015-12-22 A kind of more sorting techniques of high-spectral data based on Chernoff distances and SVM

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201510969347.1A CN105389597B (en) 2015-12-22 2015-12-22 A kind of more sorting techniques of high-spectral data based on Chernoff distances and SVM

Publications (2)

Publication Number Publication Date
CN105389597A CN105389597A (en) 2016-03-09
CN105389597B true CN105389597B (en) 2018-07-27

Family

ID=55421867

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201510969347.1A Active CN105389597B (en) 2015-12-22 2015-12-22 A kind of more sorting techniques of high-spectral data based on Chernoff distances and SVM

Country Status (1)

Country Link
CN (1) CN105389597B (en)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110008990A (en) * 2019-02-22 2019-07-12 上海拉扎斯信息科技有限公司 More classification methods and device, electronic equipment and storage medium

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101414378A (en) * 2008-11-24 2009-04-22 罗向阳 Hidden blind detection method for image information with selective characteristic dimensionality
CN102122349A (en) * 2011-02-16 2011-07-13 哈尔滨工业大学 Method for building multi-classification support vector machine classifier based on Bhattacharyya distance and directed acyclic graph
CN102831436A (en) * 2012-08-02 2012-12-19 哈尔滨工业大学 High spectrum image classifying method based on reference diagram mutual information wave band selection and EMD (Empirical Mode Decomposition)

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101414378A (en) * 2008-11-24 2009-04-22 罗向阳 Hidden blind detection method for image information with selective characteristic dimensionality
CN102122349A (en) * 2011-02-16 2011-07-13 哈尔滨工业大学 Method for building multi-classification support vector machine classifier based on Bhattacharyya distance and directed acyclic graph
CN102831436A (en) * 2012-08-02 2012-12-19 哈尔滨工业大学 High spectrum image classifying method based on reference diagram mutual information wave band selection and EMD (Empirical Mode Decomposition)

Non-Patent Citations (6)

* Cited by examiner, † Cited by third party
Title
Combining Support Vector Machines With a Pairwise Decision Tree;Jin Chen等;《IEEE》;20080731;第409-413页 *
Extending Kernel Fisher Discriminant Analysis with the Weighted Pairwise Chernoff Criterion;Guang Dai等;《Springerlink》;20061231;第308-320页 *
基于Bhattacharyya距离的典型地物波谱特征差异性分析;程熙等;《遥感技术与应用》;20130831;第707-713页 *
基于Chernoff上界的特征选择;宣国荣等;《模式识别与人工智能》;19960331;第26-30页 *
基于损失最小化的SVM多类网页分类算法;邵浩然等;《计算机应用与软件》;20050731;第16-17页 *
结肠癌基因表达谱IDE特征选取研究;潘冬寅等;《山东大学学报》;20120430;第23-29页 *

Also Published As

Publication number Publication date
CN105389597A (en) 2016-03-09

Similar Documents

Publication Publication Date Title
CN106203523B (en) The hyperspectral image classification method of the semi-supervised algorithm fusion of decision tree is promoted based on gradient
CN107451614B (en) Hyperspectral classification method based on fusion of space coordinates and space spectrum features
CN105631474B (en) Based on Jeffries-Matusita distance and class to the more classification methods of the high-spectral data of decision tree
CN106503739A (en) The target in hyperspectral remotely sensed image svm classifier method and system of combined spectral and textural characteristics
CN107563442B (en) Hyperspectral image classification method based on sparse low-rank regular graph tensor embedding
CN106557782B (en) Hyperspectral image classification method and device based on class dictionary
CN106503727B (en) A kind of method and device of classification hyperspectral imagery
CN112052755A (en) Semantic convolution hyperspectral image classification method based on multi-path attention mechanism
Li et al. Hyperspectral image classification based on multiscale spatial information fusion
CN109344891A (en) A kind of high-spectrum remote sensing data classification method based on deep neural network
Nema et al. Wheat leaf detection and prevention using support vector machine
CN110210625A (en) Modeling method, device, computer equipment and storage medium based on transfer learning
CN107092921A (en) The hyperspectral image classification method filtered based on K arest neighbors
CN116595208B (en) Classification method and device for hyperspectral images and electronic equipment
CN104268556A (en) Hyperspectral image classification method based on nuclear low-rank representing graph and spatial constraint
CN105160351B (en) Semi-supervised hyperspectral classification method based on anchor point sparse graph
CN108154094A (en) The non-supervisory band selection method of high spectrum image divided based on subinterval
CN111160396A (en) Hyperspectral image classification method based on graph convolution network with multi-graph structure
CN105320970A (en) Potato disease diagnostic device, diagnostic system and diagnostic method
CN108629373A (en) A kind of image classification method, system, equipment and computer readable storage medium
Huang et al. Joint sparsity based sparse subspace clustering for hyperspectral images
CN110147725A (en) A kind of high spectrum image feature extracting method for protecting projection based on orthogonal index office
CN106529458A (en) Deep neural network space spectrum classification method for high-spectral image
CN111368940A (en) Hyperspectral image classification method based on guided filtering and kernel extreme learning machine
CN110706092A (en) Risk user identification method and device, storage medium and electronic equipment

Legal Events

Date Code Title Description
C06 Publication
PB01 Publication
C10 Entry into substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant