CN108596154B - Remote sensing image classification method based on high-dimensional feature selection and multilevel fusion - Google Patents

Remote sensing image classification method based on high-dimensional feature selection and multilevel fusion Download PDF

Info

Publication number
CN108596154B
CN108596154B CN201810455398.6A CN201810455398A CN108596154B CN 108596154 B CN108596154 B CN 108596154B CN 201810455398 A CN201810455398 A CN 201810455398A CN 108596154 B CN108596154 B CN 108596154B
Authority
CN
China
Prior art keywords
feature
dimensional
remote sensing
features
sensing image
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
CN201810455398.6A
Other languages
Chinese (zh)
Other versions
CN108596154A (en
Inventor
王鑫
熊星南
李可
石爱业
吕国芳
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.)
Hohai University HHU
Original Assignee
Hohai University HHU
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 Hohai University HHU filed Critical Hohai University HHU
Priority to CN201810455398.6A priority Critical patent/CN108596154B/en
Publication of CN108596154A publication Critical patent/CN108596154A/en
Application granted granted Critical
Publication of CN108596154B publication Critical patent/CN108596154B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G06V20/13—Satellite images
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/211—Selection of the most significant subset of features
    • 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
    • G06V10/40—Extraction of image or video features
    • G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional 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
    • G06V10/40—Extraction of image or video features
    • G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
    • G06V10/462—Salient features, e.g. scale invariant feature transforms [SIFT]
    • G06V10/464—Salient features, e.g. scale invariant feature transforms [SIFT] using a plurality of salient features, e.g. bag-of-words [BoW] representations

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Multimedia (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Computation (AREA)
  • Evolutionary Biology (AREA)
  • General Engineering & Computer Science (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Astronomy & Astrophysics (AREA)
  • Remote Sensing (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a remote sensing image classification method based on high-dimensional feature selection and multilevel fusion. Firstly, on the basis of analyzing a special imaging mechanism of a remote sensing image, extracting various heterogeneous characteristics such as the shape, the spectrum, the texture and the like of the remote sensing image. Secondly, aiming at the extracted high-dimensional heterogeneous features, on one hand, a class information subset feature selection algorithm is adopted to perform feature selection on each high-dimensional feature to obtain a corresponding group of optimal low-dimensional feature subsets; and on the other hand, fusing every two types of high-dimensional heterogeneous features by using a discrimination correlation analysis algorithm to obtain corresponding low-dimensional fusion features. And then, respectively inputting each optimal low-dimensional feature subset and low-dimensional fusion feature into an SVM classifier, and initially classifying the remote sensing image. And finally, designing a decision-level fusion classifier based on SVM weighting, and fusing the classification result of each SVM classifier to obtain the final remote sensing image classification result.

Description

Remote sensing image classification method based on high-dimensional feature selection and multilevel fusion
Technical Field
The invention relates to a remote sensing image classification method based on high-dimensional feature selection and multilevel fusion, and belongs to the technical field of image processing and pattern recognition.
Background
Remote sensing image classification is a research hotspot in the field of remote sensing image processing and analysis, and people tend to extract increasingly rich feature information in order to obtain increasingly improved classification results, so that the image feature dimension extracted in many remote sensing image classification researches is higher and higher. However, as the feature dimension grows, redundant and even negative correlation information is highly likely to be generated between feature attributes, which not only increases the calculation amount of the classifier sharply, but also affects the performance of classification.
In order to solve the above problems, the conventional solution is to use Principal Component Analysis (PCA) algorithm to perform dimensionality reduction on high-dimensional features, or use Relief algorithm to design 'correlation statistics' to measure the importance of features so as to implement dimensionality reduction. However, these conventional feature selection or dimension reduction methods often only select a group of solutions that are considered to be optimal, and therefore, it is likely that the dimension is reduced while the information of the features is weakened.
In addition, in order to improve the accuracy of image classification, different fusion strategies are receiving more and more attention. For example, the pixel-level fusion can enhance information carried by an image, the feature-level fusion can enhance cross-correlation information among different types of features so as to improve the classification capability of the features, and the decision-level fusion can comprehensively judge a plurality of classification results without affecting the classification capability of each feature so as to obtain a better result. The document (Haghighat M, Abdel-Mottalbeb M, Alhalabi W. characterization Analysis: Real-Time Feature Level Fusion for Multimodal biometrical Recognition [ M ]. IEEE Press,2016.) shows that Feature Level Fusion strategies generate Feature data that can achieve better classification results than pixel Level Fusion or decision Level Fusion, and therefore Feature Level Fusion is essential. According to the literature (Kuncheva L I, Bezdek J C, Duin R P W. precision templates for multiple classifier fusion: an experimental composition [ J ]. Pattern Recognition,2001,34(2):299-314.), the result is classified by fusing a plurality of classifiers, which is often better than the result classified by using a single classifier.
In summary, how to comprehensively extract various heterogeneous features of an image in remote sensing image classification, how to effectively reduce dimensions of high-dimensional heterogeneous features, how to mine the relationship among the heterogeneous features to enable the heterogeneous features to be subjected to feature level fusion, and design a decision level classifier still remain challenging problems at present.
Disclosure of Invention
The purpose of the invention is as follows: aiming at the problems in the prior art, the invention provides a remote sensing image classification method based on high-dimensional feature selection and multilevel fusion, which firstly provides various heterogeneous features for extracting a remote sensing image; then, aiming at the defects that the Feature dimension extracted by the Feature extraction method in the classification of the remote sensing images is generally too high and important Feature information is easily lost only by adopting a common dimension reduction algorithm, an algorithm (SISFS) based on class information Subset Feature Selection is provided for carrying out effective Feature Selection on the high-dimensional features; meanwhile, in order to solve the problem that the existing fusion algorithm is difficult to acquire the characteristics with identification information when the multi-class heterogeneous characteristics are fused, a heterogeneous characteristic fusion method based on a Differential Correlation Analysis (DCA) is provided; and finally, a satisfactory classification performance is achieved by combining a decision-level fusion strategy.
The technical scheme is as follows: a remote sensing image classification method based on high-dimensional feature selection and multilevel fusion comprises the following steps:
step 1: on the basis of analyzing the special imaging mechanism of the remote sensing image, various heterogeneous characteristics such as the shape, the spectrum, the texture and the like of the remote sensing image are extracted. The specific process is as follows:
firstly, extracting shape features of the remote sensing image. Firstly, extracting dense SURF (speeded Up Robust of features) features and sparse SURF features of the image, wherein the dense SURF features are extracted by adopting a grid division image, a grid center point is taken as a dense SURF feature point, and the sparse SURF feature points are detected by a traditional SURF feature extraction algorithm. Then, a SURF algorithm is used to extract features in each feature point region. And then, clustering the extracted dense SURF features according to a K-Means algorithm to obtain a class center as an encoding dictionary. Finally, the extracted features are subjected to local Constrained Linear Coding (LLC).
Wherein X denotes a set of d-dimensional features obtained by describing one image, and X ═ X1,x2,...,xn]∈Rd×nAnd a feature dictionary B obtained by K-Means algorithm from the bottom features of all the training images, wherein the dictionary B is [ B ]1,b2,...,bm]∈Rd×mThe objective function encoded with the LLC algorithm is as follows:
Figure GDA0003116345720000021
wherein the lines indicate multiplication between corresponding pixels, diAn adapter representing the ith position, which assigns a weight to each atom based on its similarity to the input descriptor, λ being a parameter greater than 0, α ═ α1,α2,...,αn]，αiCorresponding feature xiThe result after LLC encoding, α is the set of encoding results for all features X.
And performing LLC feature coding on the dense SURF to obtain dense SURF-LLC features, then layering according to a Spatial Pyramid Matching model (SPM), performing LLC coding on the sparse SURF features to extract sparse SURF-LLC features, and connecting the dense SURF-LLC features and the sparse SURF-LLC features in series to obtain final shape features which are recorded as DS-SURF-LLC.
Secondly, extracting the spectral characteristics of the remote sensing image. First, the remote sensing image on the original RGB space is transformed into HSV and XYZ spaces. Then, the remote sensing image is divided on the 9 subspaces according to grids, and the mean value and the variance of each divided subregion are obtained. And finally, coding the extracted Mean value and variance information by adopting LLC, layering by adopting an SPM model to obtain final spectral characteristics, and marking as Mean-Std-LLC.
And thirdly, extracting texture features of the remote sensing image. Firstly, Gabor filtering and scale transformation are carried out on the remote sensing image. Then, for the transformed image at each scale, a Complete Local Binary Pattern (CLBP) feature is extracted. And finally, connecting the CLBP features under multiple scales in series to obtain the final multi-scale CLBP texture feature which is recorded as MS-CLBP.
Step 2: and (3) for the extracted high-dimensional heterogeneous features (the shape Feature of the remote sensing image, the spectral Feature of the remote sensing image and the texture Feature of the remote sensing image), performing Feature Selection on each high-dimensional Feature by adopting a similar information Subset Feature Selection algorithm (SISFS) to obtain a corresponding group of optimal low-dimensional Feature subsets. The specific process is as follows:
firstly, for each extracted high-dimensional heterogeneous feature, in order to correctly evaluate the classification capability of the optimal low-dimensional feature subset to be selected, the high-dimensional feature needs to be sent to an SVM classifier, and then the optimal parameter of the SVM classifier corresponding to the feature is obtained through grid search and cross validation.
Second, four objective functions of the SISFS algorithm are defined, and an optimal set of low-dimensional feature subsets is obtained by optimizing the four objective functions.
Let S be a subset selected from the high-dimensional features X, Y be a class label set corresponding to S, and we define the following four objective functions for S:
(1) correlation:
Figure GDA0003116345720000031
wherein x isiThe ith column attribute representing S. I (x)i,y)＝H(xi)+H(y)-H(xi,y)。H(xi) Is xiEntropy of (2). H (y) is the entropy of y. H (x)iY) is xiAnd the joint entropy of y.
(2) Redundancy:
Figure GDA0003116345720000041
(3) feature subset dimension:
f3(S)＝dim(S)
where dim (·) represents the dimensioning function.
(4) Cross validation average accuracy:
Figure GDA0003116345720000042
wherein, trNumber of samples correctly predicted for testing, tallFor testing the total number of samples, training samples are sent to an SVM classifier and subjected to 5-fold cross validationAnd calculating to obtain the average accuracy.
Third, a subset of class information for the SISFS algorithm is defined. Is provided with
Figure GDA0003116345720000043
Represents the average classification accuracy of S, SiAnd SjTwo different subsets of information are selected, if
Figure GDA0003116345720000044
Then call SiIs SjIs used to determine the class information subset.
Fourth, feature subsets are selected. The process mainly comprises two steps:
(1) firstly, a subset is selected as an initial subset S arbitrarily, and the initial subset is optimized by adopting a heuristic global optimization algorithm based on an ethnic group. Namely: first, four objective function values of the initial subset are calculated, and the four objective function values are optimized (i.e., the correlation f is optimized)1(S) and Cross-validation average accuracy f4(S) is maximized while making the redundancy f2(S) and feature subset dimension f3(S) minimize) to continually update the subset until an optimization termination condition is reached. It should be noted that after the process is completed, a set of optimized subsets is generated.
(2) Searching the sub-set with the optimal cross validation average accuracy in the optimized sub-sets1Then selecting sub based on the preset threshold value delta1Subsets sub with differences in average accuracy within a threshold δ2,...,subnAnd finally, all optimized subsets are sub ═ sub1,sub2,...,subn}。
And step 3: and (3) aiming at the high-dimensional heterogeneous features extracted in the step (1), fusing every two types of high-dimensional heterogeneous features by using a Discrimination Correlation Analysis (DCA) algorithm to obtain corresponding low-dimensional fusion features. The specific process is as follows:
first, for any set of high-dimensional heterogeneous feature sets X (given the dimension p × n), first, for each feature setCalculating the average characteristic vector of the characteristics of all training samples of each target class
Figure GDA0003116345720000051
Then calculate the average vector of the whole feature set X
Figure GDA0003116345720000052
Figure GDA0003116345720000053
Wherein x isjie.X denotes the feature vector of the ith sample of the jth class object. Class j target samples have njAnd (4) respectively. n represents the number of all training samples. J denotes the number of object classes.
Second, for X, find its interspecies scatter matrix Sbx，SbxIs defined as follows:
Figure GDA0003116345720000054
wherein,
Figure GDA0003116345720000055
third, to SbxCarrying out diagonalization:
Figure GDA0003116345720000056
where P is a matrix of orthogonal eigenvectors.
Figure GDA0003116345720000057
A diagonal matrix composed of non-negative eigenvalues.
Fourthly, solving the low-dimensional transformation result of X. Firstly, selecting the eigenvectors corresponding to the first r nonzero eigenvalues from the matrix P to form a matrix Q, and obtaining the matrix Q
Figure GDA0003116345720000058
Then, an inter-class scatter matrix S is usedbxThe feature vectors of the first r are mapped by: q → phibxQ, to give (phi)bxQ)TSbx(ΦbxQ)＝Λr×r. Then let Wbx＝ΦbxQΛ-1/2Then there is
Figure GDA0003116345720000059
Finally, by
Figure GDA00031163457200000510
That is, the dimension of the high-dimensional heterogeneous feature set X can be reduced from p × n to r × n, and the specific operations are as follows:
Figure GDA00031163457200000511
where X' is the result of a low-dimensional transformation of X (with dimensions r n).
Fifthly, for any two groups of different types of high-dimensional feature sets X1 and X2, according to the first step to the fourth step, low-dimensional transformation results of the feature sets X1 and X2 can be obtained:
Figure GDA0003116345720000061
sixth, to make the correlation between different types of features of the same target class stronger, first, their covariance matrices are calculated for the low-dimensional transformed results X1' and X2
Figure GDA0003116345720000062
Then, to
Figure GDA0003116345720000063
Using singular value decomposition to diagonalize, one can obtain:
Figure GDA0003116345720000064
u and V have no specific physical meaning here. U and V are singular value decomposition results and represent singular vectors corresponding to the non-square matrix, and the importance degree of the singular vectors is determined by the size of the corresponding singular value. Then, let
Figure GDA0003116345720000065
Figure GDA0003116345720000066
Where Σ is a diagonal matrix with main diagonal elements not zero, then
Figure GDA0003116345720000067
Finally, X1 'and X2' are further transformed:
Figure GDA0003116345720000068
Figure GDA0003116345720000069
seventh, the
Figure GDA00031163457200000610
And
Figure GDA00031163457200000611
carrying out fusion:
Figure GDA00031163457200000612
Fus12namely the final DCA fusion feature vector.
And 4, step 4: the sub-set of the low-dimensional features of the class information and all the low-dimensional fusion features Fus are combined12Respectively inputting the data into an SVM classifier, initially classifying the remote sensing image, and obtaining the probability P that each sample belongs to different classes under each different characteristicij，PijIndicating the probability that the sample belongs to the ith class based on the feature j.
And 5: and designing a decision-level fusion classifier based on SVM weighting, fusing classification results of each SVM classifier, and realizing remote sensing image classification. The specific process is as follows:
first, let α be given for each sample imagejWeight (f) representing the jth feature
Figure GDA00031163457200000613
J represents the number of features), given a set αjThen, the probability that the sample belongs to the i-th class target is calculated
Figure GDA00031163457200000614
Second, for each sample image, P is calculatedm＝max{P1,P2,...,PJAt this time, the sample is determined as the mth class.
Thirdly, the prediction categories and the real categories of all the samples are compared, so that the total accuracy of the whole training sample is obtained
Figure GDA0003116345720000071
trNumber of samples correctly predicted for the entire training sample, tallIs the total number of samples.
Fourthly, the first to the third steps are repeated, and all possible alpha is traversedjAnd combining to obtain the corresponding total accuracy. Selecting the weight combination with the highest accuracy as the optimal weight beta, and constructing a decision-level fusion classifier:
class＝max{Q1,Q2,...,QJ}
wherein,
Figure GDA0003116345720000072
representing the probability that the test sample belongs to the class i object.
Has the advantages that: the technical scheme provided by the invention is characterized by comprising the following related actions: extracting various heterogeneous characteristics of the remote sensing image; aiming at the defects that the feature dimension extracted by the feature extraction method in the classification of the remote sensing image is generally too high and important feature information is easily lost only by adopting a common dimension reduction algorithm, the algorithm based on class information subset feature selection is provided for carrying out effective feature selection on high-dimensional features; in order to solve the problem that the existing fusion algorithm is difficult to obtain the characteristics with identification information when the various heterogeneous characteristics are fused, a heterogeneous characteristic fusion method based on identification correlation analysis is provided; and a decision-level fusion strategy and the like are provided, so that the classification performance is improved in various aspects, and a satisfactory classification result is finally obtained.
Drawings
FIG. 1 is a block diagram of the method of the present invention.
Detailed Description
The present invention is further illustrated by the following examples, which are intended to be purely exemplary and are not intended to limit the scope of the invention, as various equivalent modifications of the invention will occur to those skilled in the art upon reading the present disclosure and fall within the scope of the appended claims.
A remote sensing image classification method based on high-dimensional feature selection and multilevel fusion is shown in a block diagram of a method in figure 1 and mainly comprises the following steps.
Step 1: on the basis of analyzing the special imaging mechanism of the remote sensing image, various heterogeneous characteristics such as the shape, the spectrum, the texture and the like of the remote sensing image are extracted. The specific process is as follows:
firstly, extracting shape features of the remote sensing image. Firstly, extracting dense SURF (speeded Up Robust of features) features and sparse SURF features of the image, wherein the dense SURF features are extracted by adopting a grid division image, a grid center point is taken as a dense SURF feature point, and the sparse SURF feature points are detected by a traditional SURF feature extraction algorithm. Then, a SURF algorithm is used to extract features in each feature point region. And then, clustering the extracted dense SURF features according to a K-Means algorithm to obtain a class center as an encoding dictionary. Finally, the extracted features are subjected to local Constrained Linear Coding (LLC).
Wherein X denotes a set of d-dimensional features obtained by describing one image, and X ═ X1,x2,...,xn]∈Rd×nAnd a feature dictionary B obtained by K-Means algorithm from the bottom features of all the training images, wherein the dictionary B is [ B ]1,b2,...,bm]∈Rd×mThe objective function encoded with the LLC algorithm is as follows:
Figure GDA0003116345720000081
wherein the lines indicate multiplication between corresponding pixels, diAn adapter representing the ith position, which assigns a weight to each atom based on its similarity to the input descriptor, λ being a parameter greater than 0, α ═ α1,α2,...,αn]，αiCorresponding feature xiThe result after LLC encoding, α is the set of encoding results for all features X.
And after the dense SURF extraction features are coded, the dense SURF-LLC features are obtained, then layering is carried out according to a Spatial Pyramid Matching model (SPM), the sparse SURF-LLC features are extracted in the same way, and the dense SURF-LLC features and the sparse SURF-LLC features are connected in series to obtain final shape features which are recorded as DS-SURF-LLC.
Secondly, extracting the spectral characteristics of the remote sensing image. First, the remote sensing image on the original RGB space is transformed into HSV and XYZ spaces. Then, the remote sensing image is divided on the 9 subspaces according to grids, and the mean value and the variance of each divided subregion are obtained. And finally, coding the extracted Mean value and variance information by adopting LLC, layering by adopting an SPM model to obtain final spectral characteristics, and marking as Mean-Std-LLC.
And thirdly, extracting texture features of the remote sensing image. Firstly, Gabor filtering and scale transformation are carried out on the remote sensing image. Then, for the transformed image at each scale, a Complete Local Binary Pattern (CLBP) feature is extracted. And finally, connecting the CLBP features under multiple scales in series to obtain the final multi-scale CLBP texture feature which is recorded as MS-CLBP.
Step 2: and (3) for the extracted high-dimensional heterogeneous features, performing Feature Selection on each high-dimensional Feature by adopting a similar information Subset Feature Selection algorithm (SISFS) to obtain a corresponding group of optimal low-dimensional Feature subsets. The specific process is as follows:
firstly, for each extracted high-dimensional heterogeneous feature, in order to correctly evaluate the classification capability of the optimal low-dimensional feature subset to be selected, the high-dimensional feature needs to be sent to an SVM classifier, and then the optimal parameter of the SVM classifier corresponding to the feature is obtained through grid search and cross validation.
Second, four objective functions of the SISFS algorithm are defined, and an optimal set of low-dimensional feature subsets is obtained by optimizing the four objective functions.
Let S be a subset selected from the high-dimensional features X, Y be a class label set corresponding to S, and we define the following four objective functions for S:
(1) correlation:
Figure GDA0003116345720000091
wherein x isiThe ith column attribute representing S. I (x)i,y)＝H(xi)+H(y)-H(xi,y)。H(xi) Is xiEntropy of (2). H (y) is the entropy of y. H (x)iY) is xiAnd the joint entropy of y.
(2) Redundancy:
Figure GDA0003116345720000092
(3) feature subset dimension:
f3(S)＝dim(S)
where dim (·) represents the dimensioning function.
(4) Cross validation average accuracy:
Figure GDA0003116345720000093
wherein, trNumber of samples correctly predicted for testing, tallIn order to test the total number of samples, the training samples are sent to an SVM classifier, and the average accuracy is calculated through 5-fold cross validation.
Third, a subset of class information for the SISFS algorithm is defined. Is provided with
Figure GDA0003116345720000094
Represents the average classification accuracy of S, SiAnd SjTwo different subsets of information are selected, if
Figure GDA0003116345720000095
Then call SiIs SjIs used to determine the class information subset.
Fourth, feature subsets are selected. The process mainly comprises two steps:
(1) firstly, a subset is selected as an initial subset S arbitrarily, and the initial subset is optimized by adopting a heuristic global optimization algorithm based on an ethnic group. Namely: first, four objective function values of the initial subset are calculated, and the four objective function values are optimized (i.e., the correlation f is optimized)1(S) and Cross-validation average accuracy f4(S) is maximized while making the redundancy f2(S) and feature subset dimension f3(S) minimize) to continually update the subset until an optimization termination condition is reached. It should be noted that after the process is completed, a set of optimized subsets is generated.
(2) Searching the sub-set with the optimal cross validation average accuracy in the optimized sub-sets1Then selecting sub based on the preset threshold value delta1Subsets sub with similar average accuracy2,...,subnFinally all optimized subsets, i.e. sub ═ sub }1,sub2,...,subn}。
And 3, fusing every two types of high-dimensional heterogeneous features by using a Discriminative Correlation Analysis (DCA) aiming at the high-dimensional heterogeneous features extracted in the step 1 to obtain corresponding low-dimensional fusion features. The specific process is as follows:
first, for any set of high-dimensional heterogeneous feature sets X (with dimension p × n), first, an average feature vector is calculated for the features of all training samples of each target class in the feature set
Figure GDA0003116345720000101
Then calculate the average vector of the whole feature set X
Figure GDA0003116345720000102
Figure GDA0003116345720000103
Wherein x isjie.X denotes the feature vector of the ith sample of the jth class object. Class j target samples have njAnd (4) respectively. n represents the number of all training samples. J denotes the number of object classes.
Second, for X, find its interspecies scatter matrix Sbx，SbxIs defined as follows:
Figure GDA0003116345720000104
wherein,
Figure GDA0003116345720000105
third, to SbxCarrying out diagonalization:
Figure GDA0003116345720000106
where P is a matrix of orthogonal eigenvectors.
Figure GDA0003116345720000107
A diagonal matrix composed of non-negative eigenvalues.
Fourthly, solving the low-dimensional transformation result of X. Firstly, selecting the eigenvectors corresponding to the first r nonzero eigenvalues from the matrix P to form a matrix Q, and obtaining the matrix Q
Figure GDA0003116345720000111
Then, an inter-class scatter matrix S is usedbxThe first r heaviestThe desired feature vector is obtained by mapping: q → phibxQ, to give (phi)bxQ)TSbx(ΦbxQ)＝Λr×r. Then let Wbx＝ΦbxQΛ-1/2Then there is
Figure GDA0003116345720000112
Finally, by
Figure GDA0003116345720000113
That is, the dimension of the high-dimensional heterogeneous feature set X can be reduced from p × n to r × n, and the specific operations are as follows:
Figure GDA0003116345720000114
where X' is the result of a low-dimensional transformation of X (with dimensions r n).
Fifthly, for any two groups of different types of high-dimensional feature sets X1 and X2, according to the first step to the fourth step, low-dimensional transformation results of the feature sets X1 and X2 can be obtained:
Figure GDA0003116345720000115
sixth, to make the correlation between different types of features of the same target class stronger, first, their covariance matrices are calculated for the low-dimensional transformed results X1' and X2
Figure GDA0003116345720000116
Then, to
Figure GDA0003116345720000117
Using singular value decomposition to diagonalize, one can obtain:
Figure GDA0003116345720000118
then, let
Figure GDA0003116345720000119
Where Σ is a diagonal matrix with main diagonal elements not zero, then
Figure GDA00031163457200001110
Finally, X1 'and X2' are further transformed:
Figure GDA00031163457200001111
seventh, the
Figure GDA00031163457200001112
And
Figure GDA00031163457200001113
carrying out fusion:
Figure GDA00031163457200001114
Fus12namely the final DCA fusion feature vector.
And 4, step 4: the sub-set of the low-dimensional features of the class information and all the low-dimensional fusion features Fus are combined12Respectively input into an SVM classifier to carry out initial classification on the remote sensing images, and the probability P that each sample belongs to different classes under each different characteristic is obtainedij，PijIndicating the probability that the sample belongs to the ith class based on the feature j.
And 5: and designing a decision-level fusion classifier based on SVM weighting, fusing classification results of each SVM classifier, and realizing remote sensing image classification. The specific process is as follows:
first, let α be given for each sample imagejWeight (f) representing the jth feature
Figure GDA0003116345720000121
J represents the number of features), given a set αjThen, the probability that the sample belongs to the i-th class target is calculated
Figure GDA0003116345720000122
Second, for each sample image, P is calculatedm＝max{P1,P2,...,PJAt this time, the sample is determined as the mth class.
Thirdly, the prediction categories and the real categories of all the samples are compared, so that the total accuracy of the whole training sample is obtained
Figure GDA0003116345720000123
trNumber of samples correctly predicted for the entire training sample, tallIs the total number of samples.
Fourthly, the first to the third steps are repeated, and all possible alpha is traversedjAnd combining to obtain the corresponding total accuracy. Selecting the weight combination with the highest accuracy as the optimal weight beta, and constructing a decision-level fusion classifier:
class＝max{Q1,Q2,...,QJ}
wherein,
Figure GDA0003116345720000124
representing the probability that the test sample belongs to the class i object.

Claims (6)

1. A remote sensing image classification method based on high-dimensional feature selection and multilevel fusion is characterized by comprising the following steps:
step 1: on the basis of analyzing a special imaging mechanism of the remote sensing image, extracting various high-dimensional heterogeneous characteristics of the remote sensing image, including the shape, spectrum and texture characteristics of the remote sensing image;
step 2: aiming at the extracted high-dimensional heterogeneous feature set, performing feature selection on each high-dimensional heterogeneous feature by adopting a class information subset feature selection algorithm SISFS to obtain a corresponding group of optimal low-dimensional feature subsets;
and step 3: aiming at the extracted high-dimensional heterogeneous feature set, fusing every two types of high-dimensional heterogeneous features by using a discrimination correlation analysis algorithm DCA to obtain corresponding low-dimensional fusion features; the specific process is as follows:
firstly, for any group of high-dimensional heterogeneous feature set X, setting the dimension of the high-dimensional heterogeneous feature set X as p multiplied by n, firstly, aiming at the features of all training samples of each target class in the high-dimensional heterogeneous feature set X, calculating the average feature vector of the training samples
Figure FDA0003116345710000011
Then, the average vector of the whole high-dimensional heterogeneous feature set X is calculated
Figure FDA0003116345710000012
Figure FDA0003116345710000013
Wherein x isjiE X represents the feature vector of the ith sample of the jth class target, and the jth class target sample has njN represents the number of all training samples, and J represents the number of target categories;
second, for X, find its interspecies scatter matrix Sbx，SbxIs defined as follows:
Figure FDA0003116345710000014
wherein,
Figure FDA0003116345710000015
third, to SbxCarrying out diagonalization:
Figure FDA0003116345710000016
wherein P is a matrix of orthogonal eigenvectors,
Figure FDA0003116345710000017
a diagonal matrix composed of non-negative eigenvalues;
fourthly, solving a low-dimensional transformation result of X; firstly, selecting the eigenvectors corresponding to the first r nonzero eigenvalues from the matrix P to form a matrix Q, and obtaining the matrix Q
Figure FDA0003116345710000018
Then, an inter-class scatter matrix S is usedbxThe first r most significant feature vectors of (a) are mapped by: q → phibxQ, to give (phi)bxQ)TSbx(ΦbxQ)＝Λr×r(ii) a Then let Wbx＝ΦbxQΛ-1/2Then there is
Figure FDA0003116345710000019
Finally, by
Figure FDA0003116345710000021
That is, the dimension of the high-dimensional heterogeneous feature set X can be reduced from p × n to r × n, and the specific operations are as follows:
Figure FDA0003116345710000022
wherein X' is the low-dimensional transformation result of X, and the dimension is r multiplied by n;
fifthly, aiming at any two groups of different types of high-dimensional heterogeneous feature sets X1 and X2, according to the first step to the fourth step, low-dimensional transformation results of the high-dimensional heterogeneous feature sets X1 and X2 can be obtained:
Figure FDA0003116345710000023
Figure FDA0003116345710000024
sixth, to make the correlation between different types of features of the same target class stronger, first, their covariance matrices are calculated for the low-dimensional transformed results X1' and X2
Figure FDA0003116345710000025
Then, to
Figure FDA0003116345710000026
Using singular value decomposition to diagonalize, one can obtain:
Figure FDA0003116345710000027
u and V have no specific physical meaning herein; u and V are singular value decomposition results, represent singular vectors corresponding to the non-square matrix, and the importance degree of the singular vectors is determined by the size of the corresponding singular value; then, let
Figure FDA0003116345710000028
Figure FDA0003116345710000029
Where Σ is a diagonal matrix with main diagonal elements not zero, then
Figure FDA00031163457100000210
Finally, X1 'and X2' are further transformed:
Figure FDA00031163457100000211
Figure FDA00031163457100000212
seventh, the
Figure FDA00031163457100000213
And
Figure FDA00031163457100000214
carrying out fusion:
Figure FDA00031163457100000215
Fus12the final DCA fusion feature vector is obtained; and 4, step 4: respectively inputting each optimal low-dimensional feature subset and low-dimensional fusion feature into an SVM classifier, and initially classifying the remote sensing image;
and 5: and designing a decision-level fusion classifier based on SVM weighting, fusing classification results of each SVM classifier, and realizing remote sensing image classification.
2. The remote sensing image classification method based on high-dimensional feature selection and multilevel fusion as claimed in claim 1, characterized in that the specific process of step 1 is as follows:
firstly, extracting shape features of a remote sensing image; firstly, extracting dense SURF features and sparse SURF features of an image, wherein the dense SURF features are extracted by adopting a grid division image, the central point of the grid is taken as a dense SURF feature point, and the sparse SURF feature points are detected by a traditional SURF feature extraction algorithm; then, extracting features in each feature point region by using an SURF algorithm; secondly, clustering the extracted dense SURF characteristics according to a K-Means algorithm to obtain a class center as a coding dictionary; finally, performing local constraint linear coding LLC on the extracted features; after extracting the feature codes, layering the dense SURF-LLC features according to a spatial pyramid matching model SPM, and connecting the dense SURF-LLC features and the sparse SURF-LLC features in series to obtain final shape features which are recorded as DS-SURF-LLC;
secondly, extracting spectral features of the remote sensing image; firstly, converting a remote sensing image on an original RGB space into HSV and XYZ spaces; then, dividing the remote sensing image on the 9 subspaces according to grids respectively, and solving the mean value and the variance of each divided subregion; finally, similar to the method in the first step, the LLC is adopted to encode the extracted Mean value and variance information, and an SPM model is adopted for layering to obtain the final spectral characteristics which are marked as Mean-Std-LLC;
thirdly, extracting texture features of the remote sensing image; firstly, carrying out Gabor filtering and scale transformation on a remote sensing image; then, extracting the CLBP characteristic of a complete local binary mode aiming at the image under each scale after transformation; and finally, connecting the CLBP features under multiple scales in series to obtain the final multi-scale CLBP texture feature which is recorded as MS-CLBP.
3. The remote sensing image classification method based on high-dimensional feature selection and multi-level fusion as claimed in claim 1, characterized in that, for the extracted high-dimensional heterogeneous feature set, a class information subset feature selection algorithm SISFS is adopted to perform feature selection on each high-dimensional heterogeneous feature to obtain a corresponding group of optimal low-dimensional feature subsets; the specific process is as follows:
firstly, aiming at each high-dimensional heterogeneous feature set, in order to correctly evaluate the classification capability of an optimal low-dimensional feature subset to be selected, the high-dimensional heterogeneous feature set is firstly sent to an SVM classifier, and then the optimal parameters of the SVM classifier corresponding to the high-dimensional heterogeneous feature set are obtained through grid search and cross validation;
secondly, defining four objective functions of a SISFS algorithm, and obtaining an optimal low-dimensional feature subset by optimizing the four objective functions;
setting S as a subset selected from the high-dimensional heterogeneous feature set X, setting Y as a class label set corresponding to S, and defining the following four objective functions aiming at S: correlation objective function f1(S), redundancy objective function f2(S), feature subset dimension objective function f3(S), cross-validation of average accuracy target function f4(S)；
Thirdly, defining a class information subset of the SISFS algorithm; is provided with
Figure FDA0003116345710000031
Represents the average classification accuracy of S, SiAnd SjTwo different subsets of information are selected, if
Figure FDA0003116345710000041
Then call SiIs SjA subset of class information of;
fourthly, selecting a feature subset; the process mainly comprises two steps:
(1) firstly, randomly selecting a subset as an initial subset S, and optimizing the initial subset by adopting a heuristic global optimization algorithm based on an ethnic group, wherein the specific process is as follows: firstly, four objective function values of an initial subset are calculated, and the subset is continuously updated by optimizing the four objective function values until an optimization termination condition is reached; after the process is finished, a group of optimized subsets is generated; four objective function value optimization means to make correlation objectiveStandard function f1(S) and cross validation average accuracy objective function f4(S) is maximized while simultaneously making the redundancy objective function f2(S) and feature subset dimension objective function f3(S) minimization;
(2) searching the sub-set with the optimal cross validation average accuracy in the optimized sub-sets1Then selecting sub based on the preset threshold value delta1Several subsets with similar average accuracy.
4. The remote sensing image classification method based on high-dimensional feature selection and multilevel fusion as claimed in claim 3, characterized in that the four objective functions are specifically:
(1) correlation objective function:
Figure FDA0003116345710000042
wherein x isiThe ith column attribute, I (x), representing Si,y)＝H(xi)+H(y)-H(xi,y)；H(xi) Is xiH (y) is the entropy of y, H (x)iY) is xiAnd the joint entropy of y;
(2) redundancy objective function:
Figure FDA0003116345710000043
(3) feature subset dimension objective function:
f3(S)＝dim(S)
wherein dim (·) represents the dimensionality function;
(4) cross validation average accuracy objective function:
Figure FDA0003116345710000051
wherein, trWhen it is a testNumber of correctly predicted samples, tallIn order to test the total number of samples, the training samples are sent to an SVM classifier, and the average accuracy is calculated through 5-fold cross validation.
5. The remote sensing image classification method based on high-dimensional feature selection and multilevel fusion of claim 1, characterized in that each optimal low-dimensional feature subset and low-dimensional fusion feature are respectively input into an SVM classifier, the remote sensing image is initially classified, and the probability P that each sample belongs to different classes under each different feature is obtainedjk，PjkIndicating the probability that the sample belongs to the jth class based on the kth feature.
6. The remote sensing image classification method based on high-dimensional feature selection and multi-level fusion as claimed in claim 1, wherein decision-level fusion classifiers based on SVM weighting are designed, and classification results of each SVM classifier are fused to realize remote sensing image classification; the specific process is as follows:
first, let α be given for each sample imagekThe weight of the kth feature is represented,
Figure FDA0003116345710000052
k represents the number of features, given a set of akThen, the probability that the sample belongs to the j-th class target is calculated
Figure FDA0003116345710000053
Second, for each sample image, P is calculatedm＝max{P1,P2,...,PKAt this time, the sample is determined as the mth type;
thirdly, the prediction categories and the real categories of all the samples are compared, so that the total accuracy of the whole training sample is obtained
Figure FDA0003116345710000054
trNumber of samples correctly predicted for the entire training sample, tallThe total number of samples is obtained;
fourthly, the first to the third steps are repeated, and all possible alpha is traversedkCombining to obtain the corresponding total accuracy; selecting the weight combination with the highest accuracy as the optimal weight betakAnd constructing a decision-level fusion classifier:
class＝max{Q1,Q2,...,QJ}
wherein,
Figure FDA0003116345710000055
representing the probability that the test sample belongs to the j-th class of targets.
CN201810455398.6A 2018-05-14 2018-05-14 Remote sensing image classification method based on high-dimensional feature selection and multilevel fusion Active CN108596154B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201810455398.6A CN108596154B (en) 2018-05-14 2018-05-14 Remote sensing image classification method based on high-dimensional feature selection and multilevel fusion

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201810455398.6A CN108596154B (en) 2018-05-14 2018-05-14 Remote sensing image classification method based on high-dimensional feature selection and multilevel fusion

Publications (2)

Publication Number Publication Date
CN108596154A CN108596154A (en) 2018-09-28
CN108596154B true CN108596154B (en) 2021-09-24

Family

ID=63637000

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810455398.6A Active CN108596154B (en) 2018-05-14 2018-05-14 Remote sensing image classification method based on high-dimensional feature selection and multilevel fusion

Country Status (1)

Country Link
CN (1) CN108596154B (en)

Families Citing this family (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109492700B (en) * 2018-11-21 2020-09-08 西安中科光电精密工程有限公司 Complex background target identification method based on multi-dimensional information fusion
CN109670418B (en) * 2018-12-04 2021-10-15 厦门理工学院 Unsupervised object identification method combining multi-source feature learning and group sparsity constraint
CN109800812A (en) * 2019-01-24 2019-05-24 山东大学第二医院 CT image classification feature selection approach and system based on counterfeit filter
CN110008925A (en) * 2019-04-15 2019-07-12 中国医学科学院皮肤病医院 A kind of skin automatic testing method based on integrated study
CN110210534B (en) * 2019-05-21 2022-09-02 河海大学 Multi-packet fusion-based high-resolution remote sensing image scene multi-label classification method
CN111340061A (en) * 2020-01-18 2020-06-26 中国人民解放军国防科技大学 Multi-mode data fusion and classification method based on SVM model parameter optimization
CN111598131B (en) * 2020-04-17 2023-08-25 北京百度网讯科技有限公司 Image processing method, device, electronic equipment and storage medium
CN112464977A (en) * 2020-10-15 2021-03-09 深圳先进技术研究院 Object classification method, computer equipment and storage medium
CN112364859B (en) * 2020-11-05 2024-03-19 哈尔滨工业大学 Multi-spectrum infrared target identification method based on feature optimization and fusion
CN113011465B (en) * 2021-02-25 2021-09-03 浙江净禾智慧科技有限公司 Household garbage throwing intelligent supervision method based on grouping multi-stage fusion
CN113378941B (en) * 2021-06-16 2022-05-03 中国石油大学(华东) Multi-decision fusion small sample image classification method
CN113673563A (en) * 2021-07-15 2021-11-19 浙江大华技术股份有限公司 Image classification method and device, electronic equipment and computer-readable storage medium
CN113673441A (en) * 2021-08-23 2021-11-19 王彬 Quantitative variation texture-driven high-resolution remote sensing image classification method
CN114461804B (en) * 2022-02-10 2023-04-07 电子科技大学 Text classification method, classifier and system based on key information and dynamic routing

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102799856A (en) * 2012-06-15 2012-11-28 天津大学 Human action recognition method based on two-channel infrared information fusion
CN104036293A (en) * 2014-06-13 2014-09-10 武汉大学 Rapid binary encoding based high resolution remote sensing image scene classification method
CN104751477A (en) * 2015-04-17 2015-07-01 薛笑荣 Space domain and frequency domain characteristic based parallel SAR (synthetic aperture radar) image classification method
CN106127225A (en) * 2016-06-13 2016-11-16 西安电子科技大学 Semi-supervised hyperspectral image classification method based on rarefaction representation
CN106203487A (en) * 2016-06-30 2016-12-07 北京航空航天大学 A kind of image classification method based on Multiple Kernel Learning Multiple Classifier Fusion and device
CN107480620A (en) * 2017-08-04 2017-12-15 河海大学 Remote sensing images automatic target recognition method based on heterogeneous characteristic fusion

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9710730B2 (en) * 2011-02-11 2017-07-18 Microsoft Technology Licensing, Llc Image registration

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102799856A (en) * 2012-06-15 2012-11-28 天津大学 Human action recognition method based on two-channel infrared information fusion
CN104036293A (en) * 2014-06-13 2014-09-10 武汉大学 Rapid binary encoding based high resolution remote sensing image scene classification method
CN104751477A (en) * 2015-04-17 2015-07-01 薛笑荣 Space domain and frequency domain characteristic based parallel SAR (synthetic aperture radar) image classification method
CN106127225A (en) * 2016-06-13 2016-11-16 西安电子科技大学 Semi-supervised hyperspectral image classification method based on rarefaction representation
CN106203487A (en) * 2016-06-30 2016-12-07 北京航空航天大学 A kind of image classification method based on Multiple Kernel Learning Multiple Classifier Fusion and device
CN107480620A (en) * 2017-08-04 2017-12-15 河海大学 Remote sensing images automatic target recognition method based on heterogeneous characteristic fusion

Also Published As

Publication number Publication date
CN108596154A (en) 2018-09-28

Similar Documents

Publication Publication Date Title
CN108596154B (en) Remote sensing image classification method based on high-dimensional feature selection and multilevel fusion
CN110443143B (en) Multi-branch convolutional neural network fused remote sensing image scene classification method
US8374442B2 (en) Linear spatial pyramid matching using sparse coding
Tharwat et al. A biometric-based model for fish species classification
CN113657425B (en) Multi-label image classification method based on multi-scale and cross-modal attention mechanism
CN110321830B (en) Chinese character string picture OCR recognition method based on neural network
KR101777601B1 (en) Distinction method and system for characters written in caoshu characters or cursive characters
CN102902979B (en) A kind of method of synthetic-aperture radar automatic target detection
Lin et al. Biomedical image classification based on a cascade of an SVM with a reject option and subspace analysis
CN107085731B (en) Image classification method based on RGB-D fusion features and sparse coding
CN107480620A (en) Remote sensing images automatic target recognition method based on heterogeneous characteristic fusion
CN108520281A (en) A kind of semi-supervised dimension reduction method of high spectrum image kept based on overall situation and partial situation
CN110569860A (en) Image interesting binary classification prediction method combining discriminant analysis and multi-kernel learning
CN103886334A (en) Multi-index fused hyperspectral remote sensing image dimensionality reduction method
Jiang et al. Joint kernel-based supervised hashing for scalable histopathological image analysis
Wilber et al. Exemplar codes for facial attributes and tattoo recognition
CN105389588A (en) Multi-semantic-codebook-based image feature representation method
Lin et al. LLC encoded BoW features and softmax regression for microscopic image classification
Li et al. Texture analysis of remote sensing imagery with clustering and Bayesian inference
Abdullah et al. Ensembles of novel visual keywords descriptors for image categorization
CN100416592C (en) Human face automatic identifying method based on data flow shape
CN104361354B (en) A kind of large nuber of images sorting technique based on sparse coding K arest neighbors histograms
Bhagyalakshmi et al. Image retrieval using color and texture binary patterns
Risojević Analysis of learned features for remote sensing image classification
Sánchez et al. Robust multiband image segmentation method based on user clues

Legal Events

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