CN108388907B - Real-time updating method of polarized SAR data classifier based on multi-view learning - Google Patents

Real-time updating method of polarized SAR data classifier based on multi-view learning Download PDF

Info

Publication number
CN108388907B
CN108388907B CN201711482208.1A CN201711482208A CN108388907B CN 108388907 B CN108388907 B CN 108388907B CN 201711482208 A CN201711482208 A CN 201711482208A CN 108388907 B CN108388907 B CN 108388907B
Authority
CN
China
Prior art keywords
view
classifier
time
classification model
classification
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
CN201711482208.1A
Other languages
Chinese (zh)
Other versions
CN108388907A (en
Inventor
聂祥丽
黄夏渊
丁曙光
乔红
张波
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.)
Institute of Automation of Chinese Academy of Science
Original Assignee
Institute of Automation of Chinese Academy of Science
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 Institute of Automation of Chinese Academy of Science filed Critical Institute of Automation of Chinese Academy of Science
Priority to CN201711482208.1A priority Critical patent/CN108388907B/en
Publication of CN108388907A publication Critical patent/CN108388907A/en
Application granted granted Critical
Publication of CN108388907B publication Critical patent/CN108388907B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • 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
    • G06F18/217—Validation; Performance evaluation; Active pattern learning techniques

Abstract

The invention belongs to the technical field of machine learning algorithm and image processing, and particularly relates to a polarized SAR data classifier real-time updating method based on multi-view learning, aiming at solving the problem that the classifier cannot be updated in real time or can only be updated independently, and the improvement of classification precision is influenced by the consistency and complementarity between views, wherein the method comprises the following steps: s1, extracting sample polarization characteristics, color characteristics and texture characteristics based on the polarized SAR image at the time t; s2, estimating the ground feature class label of the sample through an online multi-view classification model; s3, calculating loss according to the real ground object class label, and when the loss is more than zero, solving the method of the online multi-view classification model closed solution through a Lagrange multiplier method to update the classifier; and S4, after the polarized SAR image at the t +1 moment is obtained, repeating S1 to S3 until all the polarized SAR images are processed. The method can realize online real-time classification of the polarized SAR data, and has lower online classification error rate.

Description

Real-time updating method of polarized SAR data classifier based on multi-view learning
Technical Field
The invention belongs to the technical field of machine learning algorithms and image processing, and particularly relates to a polarized SAR data classifier real-time updating method based on multi-view learning.
Background
The polarization SAR is an advanced microwave remote sensing tool, and can measure the scattering property of a ground object target under different receiving and transmitting polarization combinations, and can acquire information such as dielectric constant, physical property, geometric shape, orientation and the like of the target. Compared with a single-channel SAR, the polarized SAR can acquire richer surface feature information and classification features, and is one of the main directions of the development of microwave imaging. Therefore, the polarized SAR has wide application prospect in the fields of earth resource general survey, environmental disaster monitoring, urban planning, military reconnaissance and the like. With the development and application of the polarized SAR system, the method has important theoretical value and application significance for the research of the online classification technology of mass polarized SAR data.
The polarized SAR classification problem has gained wide attention in the last two decades and a series of classification methods including supervised, semi-supervised and unsupervised methods have emerged, such as the "polarized SAR image classification method based on sparse coding and wavelet self-encoder" (patent application No. 201610407916.8, publication No. CN 106096652A) applied by the university of electronic science and technology of western ampere "applied by the university of national defense science and technology of the people's liberty military," a polarized SAR terrain classification method combining the polarization zero-angle feature of the rotation domain "(patent application No. 201710088598.8, publication No. CN 106909939 a). However, these existing polarimetric SAR classification methods are offline learning algorithms, which require that data is available at the beginning of training, and the model is learned from the training data, and only when training is completed, classification prediction can be performed. In addition, when new samples are mistaken, the trained classifier is no longer updated or needs to be retrained across the new data set. Therefore, such methods are not environment-adaptive and can be time-consuming to retrain. In addition, most of the existing methods are single-view classification algorithms, namely, only one type of features are used or a plurality of features are simply connected in series to form a vector feature, and the relationship between different attributes of feature data of each view is ignored, so that the classification precision is influenced.
Aiming at the problems, the invention provides an online multi-view learning method which is used for an online classification task of polarized SAR data. Unlike offline learning, online learning can efficiently update classifiers and does not reuse all previous data. For airborne or space-borne polarimetric SAR systems, the data is often large-scale and is acquired continuously in a continuous sequence. By introducing online learning, the system can incrementally learn a model from the data stream, can efficiently update the classifier for newly added samples, has strong adaptivity to dynamic environments, and has good expansibility for large-scale data. Therefore, the research of the online classification technology is very important for the practical application of the polarized SAR.
In recent years, many online learning methods have been proposed, such as a perceptron (perceptiron) algorithm, an Online Gradient Descent (OGD) algorithm, and a passive-aggressive (PA) algorithm. Among them, the PA algorithm minimizes the distance between the new classifier and the previous classifier and simultaneously minimizes the loss of the new classifier on the current sample (see reference 1: k. crammer, o.dekel, j. keshet, s.shalev-shuwartz, and y.singer, "Online passive-analytical algorithms," j.mach.leann.res., vol.7, no.mar, pp.551-585,2006.), is widely used due to its superior effect and low computational complexity. However, this method is only applicable to the single view angle classification problem. Nguyen et al propose a Two-view PA algorithm (see reference 2: t. Nguyen, k. chang, and s.hui, "Two-view online learning," in proc.pacific-Asia conf.knowl. discov.data min. springer,2012, pp.74-85.) and an Adaptive Two-view PA algorithm, denoted as AdaPA (see reference 3: t.nguyen, k.chang, and s.hui, "Adaptive Two-view online learning for use in the matching of the top view classification," in proc.joint European conf.mach. mach. left. To solve these problems, we propose an online two-view PA algorithm (see reference 4: x.nie, s.ding, h.qiao, b.zhang, and x.y.huang, "polari data association based on multi-view learning," in proc.int.conf.image Process (ICIP). IEEE,2017.), and the algorithm effect is significantly improved. However, these three methods can only be used for two-view and two-category problems, and are not applicable to any number of multi-view and multi-category problems. Wu et al propose an Online multi-modal distance metric learning algorithm for image extraction (see reference 5: p.wu, s.c.hoi, p.zhao, c.miao, and z. -y.liu, "Online multi-modal distance measurement with application to image retrieval," IEEE trans. knowl.data en. vol., vol.28, No.2, pp.454-467,2016.), which can be used in any number of multi-view problems, however, its classifier for each view is updated individually, not exploiting the consistency and complementarity relationships between views, thus affecting the classification accuracy. Aiming at the problems, the online classification method provided by the invention considers the relation before multiple visual angles during modeling, and is suitable for the two-classification and multiple-classification problems of any number of visual angles.
Disclosure of Invention
In order to solve the above problems in the prior art, that is, to solve the problem that the classifier cannot be updated in real time or can only be updated independently, but the improvement of classification accuracy is affected by the consistency and complementarity between views, the invention provides a polarized SAR data classifier real-time updating method based on multi-view learning, which comprises the following steps:
step S1, based on the polarized SAR image at the time t, extracting sample polarization characteristics
Figure BDA0001534028680000031
Color characteristics
Figure BDA0001534028680000032
Texture features
Figure BDA0001534028680000033
Three-view data;
step S2, based on
Figure BDA0001534028680000034
Estimating surface feature class labels of samples through an online multi-view classification model
Figure BDA0001534028680000035
Step S3, according to the real ground feature type label ytCalculating the loss ltJudging whether the sample is correctly represented or not by comparing with a set loss threshold; if the sample is represented by errors, a method for solving a closed solution of the online multi-view classification model by a Lagrange multiplier method is used for updating a classifier in the online multi-view classification model;
and step S4, after the polarized SAR image at the time of t +1 is acquired, repeating the steps S1 to S3 until all the polarized SAR images are processed.
Further, the online multi-view classification model is an online multi-view classification model of a binary task;
the online multi-view classification model of the two classification tasks is as follows:
the estimated sample class label is
Figure BDA0001534028680000041
The prediction function is
Figure BDA0001534028680000042
Is a weight parameter and satisfies
Figure BDA0001534028680000043
Loss function of lt＝max{0，1-ytft}，ytLabeling the real sample;
Figure BDA0001534028680000044
wherein the content of the first and second substances,
Figure BDA0001534028680000045
is the weight vector of the classifier at time t, m is the number of view angles from which the samples were taken, λiEqualization parameters for distance variations at different viewing angles, diIs a coupling parameter between the views, c is a positive penalty parameter, ξ is a relaxation variable,
Figure BDA0001534028680000046
for the weight vector of the classifier at time t +1 to be solved, ft+1For the prediction function at time t +1,
Figure BDA0001534028680000047
samples at view angle i at time t.
Further, the method for solving the closed solution of the online multi-view classification model by the lagrangian multiplier method to update the classifier in the online multi-view classification model comprises the following steps:
Figure BDA0001534028680000048
wherein the content of the first and second substances,
Figure BDA0001534028680000049
Figure BDA0001534028680000051
Figure BDA0001534028680000052
Figure BDA0001534028680000053
Figure BDA0001534028680000054
further, the initialized classifier weights of the online multi-view classification model
Figure BDA0001534028680000055
Is a random niColumn vectors of dimensions, i.e.
Figure BDA0001534028680000056
i is the extracted ith view sample.
Further, the online multi-view classification model is an online multi-view classification model of a multi-classification task;
the online multi-view classification model of the multi-classification task is as follows:
the estimated sample class label is
Figure BDA0001534028680000057
The prediction function is
Figure BDA0001534028680000058
Is a weight parameter and satisfies
Figure BDA0001534028680000059
A loss function of
Figure BDA00015340286800000510
ytIn order to be a true sample label,
Figure BDA00015340286800000511
Figure BDA00015340286800000512
wherein the content of the first and second substances,
Figure BDA00015340286800000513
is the weight matrix of the classifier at time t, m is the number of view angles from which the samples were taken, λiEqualization parameters for distance variations at different viewing angles, diIs a coupling parameter between the views, c is a positive penalty parameter, ξ is a relaxation variable,
Figure BDA00015340286800000514
is the weight matrix of the classifier at time t +1, ft+1Is the prediction vector at time t +1,
Figure BDA00015340286800000515
samples at time t for the ith view; f is the Forbenius norm of the matrix.
Further, the method for solving the closed solution of the online multi-view classification model by the lagrangian multiplier method to update the classifier in the online multi-view classification model comprises the following steps:
Figure BDA0001534028680000061
Figure BDA0001534028680000062
wherein the content of the first and second substances,
Figure BDA0001534028680000063
Figure BDA0001534028680000064
Figure BDA0001534028680000065
Figure BDA0001534028680000066
Figure BDA0001534028680000067
Figure BDA0001534028680000068
Figure BDA0001534028680000069
further, the initialized classifier weights of the online multi-view classification model
Figure BDA00015340286800000610
Initialisation to a random K niThe matrix of (a) is,
Figure BDA00015340286800000611
i is the extracted ith view sample.
Further, selecting parameters through cross validation, and selecting a group of the online multi-view classification models with the smallest estimated class label error rate;
parameters selected by cross-validation include:
equalization parameter lambda for distance variation at different viewing anglesiCoupling parameter d between viewing anglesiPenalty parameter c, weight parameter r1，r2。
Further, the polarization features comprise original features directly extracted from the acquired polarized SAR data and the transformation thereof and features based on polarization decomposition;
the color features comprise pseudo color image elements, dominant color weights, HSV images and histograms thereof;
the texture features comprise local binary pattern histograms, gray level co-occurrence matrixes, Gabor and wavelet transform coefficients.
Further, the value range of the input parameter is as follows:
equalization parameter lambda for distance variation at different viewing anglesiIncluding lambda1、λ2、λ3；λ1＝1，λ2，λ3∈{1，1.5}：
Coupling parameter d between viewing anglesiIncluding d1、d2、d3；d1＝d2＝d3{0.001，0.01，0.1}；
The penalty parameter c belongs to {0.05, 0.1, 0.15 };
weight parameter r1，r2Is belonged to {0.3, 0.4}, and satisfies
Figure BDA0001534028680000071
Compared with the prior art, the invention has the following advantages:
(1) realizing online classification of polarized SAR data
The updating steps of the online learning method based on PA provided by the invention are all analyzed and expressed, so that the classifier can be efficiently updated, and the problem that the existing polarized SAR offline classification method does not update the classifier or needs to retrain the classifier by using all data is solved, so that the online learning method based on PA can realize real-time classification, and has strong self-adaptability to dynamic environment and strong expansibility to large-scale data.
(2) Lower online classification error rate
The invention extracts polarization, color and texture characteristics from the polarization SAR data and takes the polarization, color and texture characteristics as different visual angles, the invention fully utilizes the consistency and complementarity relationship between the polarization, the color and texture characteristics to carry out modeling, and overcomes the defect that the prior art only uses a certain single characteristic to cause the information to be underdeveloped or uses a plurality of characteristics to be connected in series into a high-dimensional vector to cause the calculation complexity to be too high, so that the invention utilizes more complete information, thereby the error rate of online classification is lower.
Drawings
FIG. 1 shows an embodiment of the present invention
FIG. 2(a) is a Pauli decomposition pseudo-color image of single view polarization SAR data obtained by ESAR in the Oppe-Fahrenheit Hofen region of Germany;
FIG. 2(b) is a true terrain category label map corresponding to FIG. 2(a) for the Oopper faffin Hofme region of Germany obtained by ESAR;
FIG. 3 is a visual comparison of the overall classification map after the online two-classification is completed;
FIG. 4 is a visual comparison of the overall classification map after the online multi-classification is completed.
Detailed Description
Preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood by those skilled in the art that these embodiments are only for explaining the technical principle of the present invention, and are not intended to limit the scope of the present invention.
The invention discloses a polarized SAR data classifier real-time updating method based on multi-view learning, which comprises the following steps of:
step S1, based on the polarized SAR image at the time t, extracting sample polarization characteristics
Figure BDA0001534028680000081
Color characteristics
Figure BDA0001534028680000082
Texture features
Figure BDA0001534028680000083
Three-view data;
step S2, based on
Figure BDA0001534028680000084
Estimating surface feature class labels of samples through an online multi-view classification model
Figure BDA0001534028680000085
Step S3, according to the real ground feature type label ytCalculating the loss ltJudging whether the sample is correctly represented or not by comparing with a set loss threshold; if the sample is represented by errors, a method for solving a closed solution of the online multi-view classification model by a Lagrange multiplier method is used for updating a classifier in the online multi-view classification model;
and step S4, after the polarized SAR image at the time of t +1 is acquired, repeating the steps S1 to S3 until all the polarized SAR images are processed.
The following describes the technical solution of the present invention in detail according to a more detailed process from input parameter selection, parameter initialization to classifier in the specific implementation process, as shown in fig. 1, including the following steps:
step 1: by cross-validation selecting parameters, we consider the characteristics of polarization, color and texture, and thus three views, i.e., m-3, the specific ranges of input parameters are as follows: the penalty parameter c is belonged to {0.05, 0.1 and 0.15}, and the balance parameter lambda is1＝1，λ2，λ3E {1, 1.5}, coupling parameter d1＝d2＝d3E {0.001, 0.01, 0.1}, weight parameter r1，r2Is belonged to {0.3, 0.4}, and satisfies
Figure BDA0001534028680000091
The parameters given hereAnd selecting a range, wherein an optimal parameter value, namely a group of parameters with the minimum error rate, can be selected in a cross validation mode.
Step 2: initializing the weight of the classifier, classifying the problem,
Figure BDA0001534028680000092
Figure BDA0001534028680000093
namely, it is
Figure BDA0001534028680000094
Is a random niA column vector of dimensions; for the problem of multi-classification,
Figure BDA0001534028680000095
namely, it is
Figure BDA0001534028680000096
Initialisation to a random K niOf the matrix of (a).
And step 3: acquiring polarimetric SAR covariance data, and extracting polarimetric, color and texture characteristics: the polarization characteristics comprise original characteristics and characteristics based on polarization decomposition, which are directly extracted from the obtained polarization SAR data and the transformation thereof; the color features comprise pseudo color image elements, dominant color weights, HSV images and histograms thereof; the texture features comprise local binary pattern histograms, gray level co-occurrence matrixes, Gabor, wavelet transform coefficients and the like. The polarization, color and texture features used in the present invention are shown in Table 1, and it can be seen that the dimensions of these three features are n1＝45，n234 and n3The present invention uses them as three view data for subsequent classification 86.
Table 1: polarization, color and texture features for online classification
Figure BDA0001534028680000097
Figure BDA0001534028680000101
And 4, step 4: and establishing an online multi-view classification model, and predicting the label of the sample according to a classification function.
Specifically, the online multi-view classification model for the two classification task is as follows:
Figure BDA0001534028680000102
s.t.l(w；(xt，yt) Xi is less than or equal to xi; ξ ≧ 0, which represents the relaxation constraint on the binary loss function, where the relaxation variable ξ must be nonnegative.
Wherein the content of the first and second substances,
Figure BDA0001534028680000103
is the weight vector of the classifier at time t, m is the number of view angles from which the samples were taken, λiIs an equalization parameter of the distance variations at different viewing angles, diIs the coupling parameter between views, c is a positive penalty parameter, ξ is the relaxation variable,
Figure BDA0001534028680000104
for the weight vector of the classifier at time t +1 to be solved, ft+1For the prediction function at time t +1,
Figure BDA0001534028680000105
samples at view angle i at time t.
The loss function is defined as the change-loss lt＝max{0，1-ytft}. The prediction function is defined as
Figure BDA0001534028680000106
Wherein r isiEpsilon (0, 1) is a weight parameter and satisfies
Figure BDA0001534028680000107
Class label of the estimated sample
Figure BDA0001534028680000108
The online multi-view classification model for the multi-classification task is as follows:
Figure BDA0001534028680000109
s.t.lMC(W；(xt，yt) Xi is less than or equal to xi; xi is more than or equal to 0. Representing a relaxation constraint on a multi-class classification loss function where the relaxation variable ξ must be greater than or equal to 0.
Wherein λ isi，diAnd c is a positive parameter,
Figure BDA0001534028680000111
is the weight matrix of the classifier at time t,
Figure BDA0001534028680000112
is the weight matrix of the classifier at time t +1, ft+1The prediction vector at time t +1, and F is the Forbenius norm of the matrix.
Prediction function
Figure BDA0001534028680000113
It can be known that
Figure BDA0001534028680000114
lMCA loss function representing a classification of classes, defined as
Figure BDA0001534028680000115
ytIn order to be a true sample label,
Figure BDA0001534028680000116
in most cases, the class labels of the estimated samples
Figure BDA0001534028680000117
And 5: calculating the loss l according to the real label corresponding to the sampletIf l ist＝0，Indicating that the sample is correctly classified, and not updating the classifier; if l ist> 0, indicating that the sample was misclassified and that an update to the current classifier is required.
For the two classification problem, loss lt＝max{0，1-ytftIs, if ltIf the value is more than 0, solving the two-classification optimization problem in the previous step by a Lagrange multiplier method to obtain the following closed solution, and updating the classifier by the closed solution:
Figure BDA0001534028680000118
wherein the content of the first and second substances,
Figure BDA0001534028680000119
Figure BDA00015340286800001110
Figure BDA00015340286800001111
Figure BDA00015340286800001112
Figure BDA00015340286800001113
for multi-classification problems, losses
Figure BDA0001534028680000121
If l istIf the number is more than 0, solving the multi-classification optimization problem in the previous step by a Lagrange multiplier method to obtain the following closed solution, and updating the classifier by the closed solution:
Figure BDA0001534028680000122
Figure BDA0001534028680000123
wherein the content of the first and second substances,
Figure BDA0001534028680000124
Figure BDA0001534028680000125
Figure BDA0001534028680000126
Figure BDA0001534028680000127
Figure BDA0001534028680000128
Figure BDA0001534028680000129
Figure BDA00015340286800001210
step 6: if a new sample is input, returning to the step 1; if all samples are processed, calculating the classification error rate of the whole online learning process, wherein the error rate of each subclass is the ratio of the wrongly-classified samples to the total number of the subclasses, and the total error rate is the ratio of the wrongly-classified samples to the total number of the subclasses; and drawing a final classification map, polarizing pixels with the same ground object type in the SAR image, and expressing the pixels with the same color to further obtain the classification map.
The following further description is made with reference to the subdivision steps in the case of the two-classification task and the multi-classification task, respectively.
In one embodiment of the invention, the online multi-view learning algorithm for the two-classification task comprises the following steps:
(1) parameters were selected by cross validation: penalty parameter c>0, balance parameter λi> 0, coupling parameter di> 0 and a weight parameter riIs epsilon (0, 1) and
Figure BDA00015340286800001211
(2) initializing weight vectors for classifiers
Figure BDA0001534028680000131
Figure BDA0001534028680000132
Where m is the number of viewing angles, niIs the dimension of the ith viewing angle
Figure BDA0001534028680000133
(3) At time t, receiving polarized SAR data, and extracting polarization
Figure BDA0001534028680000134
Colour(s)
Figure BDA0001534028680000135
And texture features
Figure BDA0001534028680000136
Respectively as different viewing angles, so that the received samples
Figure BDA0001534028680000137
Figure BDA0001534028680000138
(4) Computing a prediction function
Figure BDA0001534028680000139
(5) Estimating class labels of samples
Figure BDA00015340286800001310
(6) Receive the correct tag yt∈{+1，-1}；
(7) Calculating the loss lt＝max{0，1-ytft}；
(8) If l istWhen the sample is classified correctly, the classifier is not updated, and the next iteration is directly carried out; if l istIf > 0, the current classifier is updated as follows:
(9) the auxiliary variables are calculated according to the following formula:
Figure BDA00015340286800001311
Figure BDA00015340286800001312
Figure BDA00015340286800001313
Figure BDA00015340286800001314
Figure BDA00015340286800001315
(10) updating the classifier:
Figure BDA00015340286800001316
(11) if a new sample arrives, t is t +1, the step (3) is executed in a returning way, otherwise, the algorithm is terminated.
The online multi-view learning algorithm for the multi-classification task in one embodiment of the invention comprises the following steps:
(1) parameters were selected by cross validation: c. C>0，λi＞0，di> 0 and riIs epsilon (0, 1) and
Figure BDA00015340286800001317
wherein m is the number of viewing angles;
(2) initializing weight vectors for classifiers
Figure BDA0001534028680000141
Figure BDA0001534028680000142
Where K is the number of classes, niIs the dimension of the ith viewing angle
Figure BDA0001534028680000143
(3) At time t, receiving polarized SAR data, and extracting polarization
Figure BDA0001534028680000144
Colour(s)
Figure BDA0001534028680000145
And texture features
Figure BDA0001534028680000146
Respectively as different viewing angles, so that the received samples
Figure BDA0001534028680000147
Figure BDA0001534028680000148
(4) Computing a prediction function
Figure BDA0001534028680000149
It can be known that
Figure BDA00015340286800001410
(5) Estimating class labels of samples
Figure BDA00015340286800001411
(6) Receive the correct tag yt∈Y＝[1，2，...，K}；
(7) Computing
Figure BDA00015340286800001412
(8) Calculating loss
Figure BDA00015340286800001413
(9) If l istIf the sample is classified correctly, the classifier is not updated, and the next iteration is directly carried out, otherwise, if l is not updatedtIf > 0, the current classifier is updated as follows:
(10) the auxiliary variables are calculated according to the following formula:
Figure BDA00015340286800001414
Figure BDA00015340286800001415
Figure BDA00015340286800001416
Figure BDA00015340286800001417
Figure BDA00015340286800001418
Figure BDA00015340286800001419
Figure BDA00015340286800001420
(11) updating the classifier:
Figure BDA00015340286800001421
Figure BDA00015340286800001422
(12) if a new sample arrives, t is t +1, the step (3) is executed in a returning way, otherwise, the algorithm is terminated.
The effects of the present invention are further described below with reference to fig. 2-4:
experimental data and conditions:
the invention uses real polarized SAR data to make test experiment, which is the single-view L wave band data of German Ohbo Fahrenheit Hofen area obtained by E-SAR sensor, and can be downloaded from European space bureau website. The Pauli decomposition pseudo color image of the data is shown in FIG. 2(a), the size of which is 1300 x 1200, and FIG. 2(b) is a corresponding real ground object classification diagram, wherein color blocks respectively represent urban areas, forest lands, roads, farmlands and other scenes.
In the simulation experiment, the software used: MATLAB R2015b, processor: intel (R) core (TM) i7-6700HQ, memory: 20.0GB, operating System: 64-bit Windows 10.
And (3) analyzing the experimental content and the result:
two-classification tasks and multi-classification tasks are considered in the experiment and respectively correspond to the problems of urban scene extraction and ground feature classification. To better evaluate the effect of the proposed method, the results compared to the PA on polarization, color, texture and their combined features are denoted PA _ Pol, PA _ Col, PA _ Tex, PA _ Cat, respectively. In addition, also in comparison with the results of the AdaPA method proposed in reference 3 and the OMDML method proposed in reference 5, it is noted that the AdaPA method is only suitable forFor binary problems, so it is not compared to in the multi-classification task. For better comparison of these methods, the parameters contained in them were selected by cross-validation, and the selection range of the parameters was set as follows: attack parameter c of PA method belongs to [0.05, 0.15 ]](ii) a The coupling parameter d of the AdaPA method belongs to {0.001, 0.01, 0.1}, the weight parameter r belongs to (0, 1), and the penalty parameter c belongs to [0.01, 0.15}](ii) a Penalty parameter c of OMDML method belongs to [0.01, 0.15 ]]The discount parameter beta ∈ [0.8, 1 ]](ii) a Method of the invention, lambda1＝1，λ2，λ3∈{1，1.5}，d1＝d2＝d3∈{0.001，0.01，0.1}，r1，r2∈{0.3，0.4}，c∈{0.05，0.1，0.15}。
Fig. 3 shows a visual comparison result of the overall classification chart after the online two-classification is completed, fig. 3(a) is a real classification label chart extracted from an urban area scene, and a non-urban area is labeled as white; FIGS. 3(b) - (h) are graphs of the classification results of PA _ Pol, PA _ Col, PA _ Tex, PA _ Cat, AdaPA, OMDML and the present invention, respectively. Table 2 shows the classification error rate comparison results of these methods in the case of two classifications. As can be seen from fig. 3, the classification result of PA _ Cat is significantly better than the results of PA _ Pol, PA _ Col and PA _ Tex, since more characteristic information is used in PA _ Cat, which can also be confirmed from table 2. In addition, from table 2, it can be seen that the ratios of the urban areas to be misclassified into non-urban areas are 43.95% and 51.08%, respectively, which are significantly higher than the ratio of PA _ Cat to urban areas of 29.72%, and it can also be seen from fig. 3(e) - (g) that there are many places in (f) and (g) to be misclassified into white areas. From table 2 and fig. 3, it can be concluded that the method proposed by the present invention results in the lowest positive sample (i.e. urban) classification error rate of 22.89% and the lowest overall classification error rate of 7.05%.
Table 2: comparison result of classification error rate under two classification conditions
Method of producing a composite material PA_Pol PA_Col PA_Tex PA_Cat AdaPA OMDML Method for producing a composite material
Urban area 0.4169 0.5449 0.4442 0.2972 0.4395 0.5108 0.2289
Non-urban area 0.0777 0.1039 0.0847 0.0571 0.0346 0.0199 0.0396
Total error rate 0.1333 0.1758 0.1435 0.0965 0.1007 0.1012 0.0705
Fig. 4 is a visual comparison result of the overall classification chart after the online multi-classification is completed, and the corresponding real class label chart is shown in fig. 2 (b). FIGS. 4(a) - (f) are graphs of the classification results of PA _ Pol, PA _ Col, PA _ Tex, PA _ Cat, OMDML, and the present invention, respectively. Table 3 shows the classification error rate comparison results of these methods for this multi-classification case. It can be seen that the polarization feature provides better discrimination information than the color and texture features because the overall error rate of PA _ Pol is 10% lower than that of PA _ Col and PA _ Tex, and in addition, more than half of the pixels of PA _ Col and PA _ Tex are misclassified in road and farmland areas. The results for PA _ Cat are significantly better than PA _ Pol, PA _ Col, and PA _ Tex, and PA _ Cat improves the discrimination of the boundary region, as shown in fig. 4 (d). The overall classification accuracy of OMDML is improved by 6% over PA _ Cat and the boundaries of the different regions become clearer according to fig. 4 (e). From the numerical results in table 3 and the visual results in fig. 4, it can be seen that the proposed method allows a large portion of the samples to be correctly classified, reaching the lowest overall error rate compared to other methods.
Table 3: classification error rate comparison under multi-classification condition
Figure BDA0001534028680000171
Those of skill in the art will appreciate that the method steps of the examples described in connection with the embodiments disclosed herein may be embodied in electronic hardware, computer software, or combinations of both, and that the components and steps of the examples have been described above generally in terms of their functionality in order to clearly illustrate the interchangeability of electronic hardware and software. Whether such functionality is implemented as electronic hardware or software depends upon the particular application and design constraints imposed on the solution. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present invention.
The steps of a method or algorithm described in connection with the embodiments disclosed herein may be embodied in hardware, a software module executed by a processor, or a combination of the two. A software module may reside in Random Access Memory (RAM), memory, Read Only Memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.
The terms "comprises," "comprising," or any other similar term are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus.
So far, the technical solutions of the present invention have been described in connection with the preferred embodiments shown in the drawings, but it is easily understood by those skilled in the art that the scope of the present invention is obviously not limited to these specific embodiments. Equivalent changes or substitutions of related technical features can be made by those skilled in the art without departing from the principle of the invention, and the technical scheme after the changes or substitutions can fall into the protection scope of the invention.

Claims (9)

1. A polarized SAR data classifier real-time updating method based on multi-view learning is characterized by comprising the following steps:
step S1, based on the polarized SAR image at the time t, extracting sample polarization characteristics
Figure FDA0002733540270000011
Color characteristics
Figure FDA0002733540270000012
Texture features
Figure FDA0002733540270000013
Three-view data;
step S2, based on
Figure FDA0002733540270000014
Estimating surface feature class labels of samples through an online multi-view classification model
Figure FDA0002733540270000015
Step S3, according to the real ground feature type label ytCalculating the loss ltJudging whether the sample is correctly represented or not by comparing with a set loss threshold; if the sample is represented by errors, a method for solving a closed solution of the online multi-view classification model by a Lagrange multiplier method is used for updating a classifier in the online multi-view classification model;
step S4, after acquiring the polarized SAR image at the time of t +1, repeating the steps S1 to S3 until all the polarized SAR images are processed;
the online multi-view classification model comprises an online multi-view classification model of a two-classification task and an online multi-view classification model of a multi-classification task;
the online multi-view classification model of the multi-classification task is as follows:
the estimated sample class label is
Figure FDA0002733540270000016
The prediction function is
Figure FDA0002733540270000017
riEpsilon (0, 1) is a weight parameter and satisfies
Figure FDA0002733540270000018
A loss function of
Figure FDA0002733540270000019
ytIs a true sample class label, ytE.y ═ {1, 2., K }, K denotes the total number of category labels,
Figure FDA00027335402700000110
Figure FDA00027335402700000111
Figure FDA0002733540270000021
Figure FDA0002733540270000022
is the weight matrix of the classifier at time t, m is the number of view angles from which the samples were taken, λiEqualization parameters for distance variations at different viewing angles, diIs a coupling parameter between the views, c is a positive penalty parameter, ξ is a relaxation variable,
Figure FDA0002733540270000023
is the weight matrix of the classifier at time t +1, niIs the dimension of the ith view angle and i ═ 1t+1For the prediction function at time t +1,
Figure FDA0002733540270000024
samples at time t for the ith view; f is the Forbenius norm of the matrix.
2. The multi-view learning-based polarized SAR data classifier real-time updating method according to claim 1, characterized in that the online multi-view classification model is an online multi-view classification model of a binary task;
the online multi-view classification model of the two classification tasks is as follows:
the estimated sample class label is
Figure FDA0002733540270000025
The prediction function is
Figure FDA0002733540270000026
riEpsilon (0, 1) is a weight parameter and satisfies
Figure FDA0002733540270000027
Loss function of lt＝max{0，1-ytft}，ytLabeling the real sample;
Figure FDA0002733540270000028
s.t.lt＝max{0，1-ytft}≤ξ；ξ≥0
wherein the content of the first and second substances,
Figure FDA0002733540270000029
is the weight vector of the classifier at time t, m is the number of view angles from which the samples were taken, λiEqualization parameters for distance variations at different viewing angles, diIs a coupling parameter between the views, c is a positive penalty parameter, ξ is a relaxation variable,
Figure FDA00027335402700000210
for the weight vector of the classifier at time t +1 to be solved, ft+1For the prediction function at time t +1,
Figure FDA00027335402700000211
samples at view angle i at time t.
3. The method for updating the polarized SAR data classifier in real time based on multi-view learning of claim 2 is characterized in that the classifier in the linear multi-view classification model of the two classification tasks is updated by a method of solving the closed solution of the online multi-view classification model by a Lagrange multiplier method, and the method comprises the following steps:
Figure FDA0002733540270000031
wherein the content of the first and second substances,
Figure FDA0002733540270000032
Figure FDA0002733540270000033
Figure FDA0002733540270000034
Figure FDA0002733540270000035
Figure FDA0002733540270000036
4. the multi-view learning-based polarized SAR data classifier real-time updating method as claimed in claim 3, wherein the initialized classifier weights of the online multi-view classification model
Figure FDA0002733540270000037
Is a random niA column vector of the dimension(s),
Figure FDA0002733540270000038
i is the extracted ith view sample.
5. The method for updating the polarized SAR data classifier in real time based on the multi-view learning of claim 1 is characterized in that the classifier in the line multi-view classification model of the multi-classification task is updated by a method of solving the closed solution of the online multi-view classification model by a Lagrangian multiplier method, and the method comprises the following steps:
Figure FDA0002733540270000039
Figure FDA00027335402700000310
wherein the content of the first and second substances,
Figure FDA00027335402700000311
Figure FDA00027335402700000312
Figure FDA0002733540270000041
Figure FDA0002733540270000042
Figure FDA0002733540270000043
Figure FDA0002733540270000044
Figure FDA0002733540270000045
6. the multi-view learning-based polarized SAR data classifier real-time updating method as claimed in claim 5, wherein the initialized classifier weights of the online multi-view classification model
Figure FDA0002733540270000046
Initialisation to a random K niThe matrix of (a) is,
Figure FDA0002733540270000047
i is the extracted ith view sample.
7. The method for updating the polarized SAR data classifier based on multi-view learning in real time as claimed in any one of claims 1-6, characterized in that a group with the smallest class label error rate estimated by the online multi-view classification model is selected by selecting parameters through cross validation;
parameters selected by cross-validation include:
equalization parameter lambda for distance variation at different viewing anglesiCoupling parameter d between viewing anglesiPenalty parameter c, weight parameter r1，r2。
8. The multi-view learning-based polarized SAR data classifier real-time updating method according to claim 7,
the polarization characteristics comprise original characteristics and characteristics based on polarization decomposition, which are directly extracted from the obtained polarization SAR data and the transformation thereof;
the color features comprise pseudo color image elements, dominant color weights, HSV images and histograms thereof;
the texture features comprise local binary pattern histograms, gray level co-occurrence matrixes, Gabor and wavelet transform coefficients.
9. The multi-view learning-based real-time updating method for the polarized SAR data classifier according to claim 7, wherein the value range of the input parameters is as follows:
equalization parameter lambda for distance variation at different viewing anglesiIncluding lambda1、λ2、λ3；λ1＝1，λ2，λ3∈{1，1.5}：
Coupling parameter d between viewing anglesiIncluding d1、d2、d3；d1＝d2＝d3∈{0.001，0.01，0.1}；
The penalty parameter c belongs to {0.05, 0.1, 0.15 };
weight parameter r1，r2Is belonged to {0.3, 0.4}, and satisfies
Figure FDA0002733540270000051
CN201711482208.1A 2017-12-29 2017-12-29 Real-time updating method of polarized SAR data classifier based on multi-view learning Active CN108388907B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201711482208.1A CN108388907B (en) 2017-12-29 2017-12-29 Real-time updating method of polarized SAR data classifier based on multi-view learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201711482208.1A CN108388907B (en) 2017-12-29 2017-12-29 Real-time updating method of polarized SAR data classifier based on multi-view learning

Publications (2)

Publication Number Publication Date
CN108388907A CN108388907A (en) 2018-08-10
CN108388907B true CN108388907B (en) 2021-03-02

Family

ID=63076666

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201711482208.1A Active CN108388907B (en) 2017-12-29 2017-12-29 Real-time updating method of polarized SAR data classifier based on multi-view learning

Country Status (1)

Country Link
CN (1) CN108388907B (en)

Families Citing this family (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110222778B (en) * 2019-06-11 2021-04-13 中国科学院自动化研究所 Online multi-view classification method, system and device based on deep forest
CN110222793B (en) * 2019-06-20 2021-06-22 中国科学院自动化研究所 Online semi-supervised classification method and system based on multi-view active learning
CN110414623B (en) * 2019-08-06 2022-06-24 广东工业大学 Classification model construction method and device based on multi-view learning
CN110766084B (en) * 2019-10-28 2021-06-15 北京理工大学 Small sample SAR target identification method based on CAE and HL-CNN
CN111949794A (en) * 2020-08-14 2020-11-17 扬州大学 Online active machine learning method for text multi-classification task

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103514456A (en) * 2013-06-30 2014-01-15 安科智慧城市技术（中国）有限公司 Image classification method and device based on compressed sensing multi-core learning
CN104463193A (en) * 2014-11-04 2015-03-25 西安电子科技大学 Polarization SAR image classifying method based on depth sparsity ICA
CN105303198A (en) * 2015-11-17 2016-02-03 福州大学 Remote-sensing image semi-supervision classification method based on customized step-size learning
CN105512680A (en) * 2015-12-02 2016-04-20 北京航空航天大学 Multi-view SAR image target recognition method based on depth neural network
CN107133649A (en) * 2017-05-09 2017-09-05 西安电子科技大学 Increment type polarization SAR terrain classification method based on a domain distance

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20150316646A1 (en) * 2014-05-01 2015-11-05 Utah State University Synthetic aperture radar target modeling
US9846822B2 (en) * 2015-12-31 2017-12-19 Dropbox, Inc. Generating and utilizing normalized scores for classifying digital objects

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103514456A (en) * 2013-06-30 2014-01-15 安科智慧城市技术（中国）有限公司 Image classification method and device based on compressed sensing multi-core learning
CN104463193A (en) * 2014-11-04 2015-03-25 西安电子科技大学 Polarization SAR image classifying method based on depth sparsity ICA
CN105303198A (en) * 2015-11-17 2016-02-03 福州大学 Remote-sensing image semi-supervision classification method based on customized step-size learning
CN105512680A (en) * 2015-12-02 2016-04-20 北京航空航天大学 Multi-view SAR image target recognition method based on depth neural network
CN107133649A (en) * 2017-05-09 2017-09-05 西安电子科技大学 Increment type polarization SAR terrain classification method based on a domain distance

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
POLSAR DATA ONLINE CLASSIFICATION BASED ON MULTI-VIEW LEARNING;Xiangli Nie等;《IEEE Signal Processing Society》;20170915;第1页 *
基于极化SAR技术的舰船检测与识别算法研究;杨慧;《中国优秀硕士学位论文全文数据库》;20140915;第I136-469页 *

Also Published As

Publication number Publication date
CN108388907A (en) 2018-08-10

Similar Documents

Publication Publication Date Title
CN108388907B (en) Real-time updating method of polarized SAR data classifier based on multi-view learning
Qu et al. Hyperspectral anomaly detection through spectral unmixing and dictionary-based low-rank decomposition
Montazer et al. An improved radial basis function neural network for object image retrieval
CN108537102B (en) High-resolution SAR image classification method based on sparse features and conditional random field
Mobahi et al. Segmentation of natural images by texture and boundary compression
McIver et al. Estimating pixel-scale land cover classification confidence using nonparametric machine learning methods
Krylov et al. Supervised high-resolution dual-polarization SAR image classification by finite mixtures and copulas
Awad An Unsupervised Artificial Neural Network Method for Satellite Image Segmentation.
Ke et al. Adaptive change detection with significance test
CN107491734B (en) Semi-supervised polarimetric SAR image classification method based on multi-core fusion and space Wishart LapSVM
CN109359525B (en) Polarized SAR image classification method based on sparse low-rank discrimination spectral clustering
CN110427835B (en) Electromagnetic signal identification method and device for graph convolution network and transfer learning
Fu et al. Contextual online dictionary learning for hyperspectral image classification
Qin et al. Learning discriminative embedding for hyperspectral image clustering based on set-to-set and sample-to-sample distances
Yang et al. Unsupervised images segmentation via incremental dictionary learning based sparse representation
Rusyn et al. Segmentation of atmospheric clouds images obtained by remote sensing
CN113723492A (en) Hyperspectral image semi-supervised classification method and device for improving active deep learning
Alhichri et al. Clustering of hyperspectral images with an ensemble method based on fuzzy c-means and markov random fields
Li et al. Alter-cnn: An approach to learning from label proportions with application to ice-water classification
Du et al. Training SegNet for cropland classification of high resolution remote sensing images
Bolton et al. Random set framework for context-based classification with hyperspectral imagery
Teke et al. Texture classification and retrieval using the random neural network model
Rekik et al. A k-means clustering algorithm initialization for unsupervised statistical satellite image segmentation
Cao et al. Hyperspectral image classification based on three-dimensional adaptive sampling and improved iterative shrinkage-threshold algorithm
Yang et al. A probabilistic spectral graph matching algorithm for robust correspondence between lunar surface images

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