CN107220659B - High Resolution SAR image classification method based on total sparse model - Google Patents
High Resolution SAR image classification method based on total sparse model Download PDFInfo
- Publication number
- CN107220659B CN107220659B CN201710329004.8A CN201710329004A CN107220659B CN 107220659 B CN107220659 B CN 107220659B CN 201710329004 A CN201710329004 A CN 201710329004A CN 107220659 B CN107220659 B CN 107220659B
- Authority
- CN
- China
- Prior art keywords
- matrix
- pixel
- vector
- classification
- initial
- 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
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
Abstract
The invention discloses a kind of High Resolution SAR image classification method based on total sparse model, the present invention, which solves, to be confined to indicate that image leads to time complexity high technical problem of classifying with comprehensive sparse model when SAR image classification.Its process are as follows: initial pixel value matrix X is chosen in SAR image to be sorted；It chooses analytic operator and learns initial sample；Projection subgradient algorithm and unified professional etiquette model tight frame method are combined into study analytic operator Ω；Sparse coefficient Z altogether is solved with augmentation Lagrangian method；It combines the pixel value vector of the total sparse coefficient vector of each pixel respective pixel block and the block of pixels to obtain feature vector；Classified based on SVM classifier, obtains the prediction label of each pixel feature vector of full figure；Prediction label result is shown with gray level image.The present invention can quickly acquire the rarefaction representation of image, ensure that the timeliness and classification accuracy of SAR image classification, the classification for High Resolution SAR image.
Description
Technical field
The invention belongs to technical field of image processing, it is accurate for be classification for High Resolution SAR image, specifically one
High Resolution SAR image classification method of the kind based on total sparse model is applied to SAR image classification field.
Background technique
As synthetic aperture radar (Synthetic Aperture Radar, SAR) imaging technique is gradually promoted, SAR figure
As using more and more, having promoted further research and development SAR image sorting technique in every field.Due to original SAR
There are coherent speckle noises for image, so that image classification method traditional in the past is not suitable in SAR image classification.SAR image point
Class can be applied to resource detection, military surveillance, medical domain, monitoring of crop growth, disaster hazard evaluation etc., SAR
The value and importance of image application, SAR image classification method also need to be further increased.
Compare classical SAR image classification method step and is generally divided into four steps: feature extraction, dictionary learning, feature coding
Classify with classifier, wherein dictionary learning and coding study are the two parts for focusing on research in recent years.
The emphasis of general Study dictionary learning is based on the collective model in sparse representation model.In fact early in 2011
SangnamNam et al. just proposes another sparse representation model ignored by everybody: analytic modell analytical model, also referred to as sparse solution altogether
It analyses model (Cosparse Analysis Model), and elaborated its principle and algorithm in 2013, be applied to people
Face image denoises direction, but and is not used for image classification aspect.
2014, Sumit Shekhar et al. proposed the classification that total sparse model is applied to facial image, was based on soft threshold
Value method solves optimization problem method and obtains total sparse coefficient, and LOST algorithm is utilized and is extracted the change of each pixel point in image
Characteristic of field is changed, and has done facial image classification experiments and has demonstrated the feasibility of method, and is real with comprehensive sparse model was previously based on
The method of existing facial image classification is compared, and discovery realizes that the rate of facial image classification is improved.
Although sparse analytic modell analytical model has been applied to image classification field altogether, and the method based on total sparse model is
Advantage is shown in terms of rate, but such methods are not applied in High Resolution SAR image classification field also, and image point
Class method only simply uses soft threshold method when solving sparse coefficient altogether, does not make full use of the analytic operator of study,
Image classification accuracy rate is limited.
Summary of the invention
In view of the above-mentioned deficiencies in the prior art, it is an object of the present invention to propose that a kind of classification accuracy is improved and speed is fast
The High Resolution SAR image classification method based on total sparse model.
The present invention is a kind of High Resolution SAR image classification method based on total sparse model, is included the following steps:
Step 1, input picture obtain initial pixel value matrix: a High Resolution SAR image to be sorted are chosen, with all
M is intercepted centered on pixel1×M1The block of pixels of size, each block of pixels transform into a column, traverse all pixels point in image and obtain
Experiment sample X ∈ RM×LL, M is the pixel number in a block of pixels, and LL is the pixel in High Resolution SAR image to be sorted
Number.
Step 2, the initial sample for choosing analytic operator study: L column are randomly selected in initial pixel value matrix X and are constituted greatly
The small matrix Y for M × L, the sample matrix Y=[y as operator study1,y2,...,yI,…,yL], yIIndicate imago in i-th
M where vegetarian refreshments1×M1The column vector of the block of pixels expansion of size.
Step 3, study analytic operator: by projection subgradient and unified professional etiquette model tight frame (Uniform Normalized
Tight Frame, UNTF) both combine, obtain analytic operator Ω using initial sample matrix Y iterative learning.
Step 4 solves sparse coefficient altogether based on Augmented Lagrange method: by analytic operator Ω and initial pixel value matrix
X is added antithesis parametric solution convex programming problem and obtains according to sparse analytic modell analytical model Z=Ω X altogether using Augmented Lagrange method
To total sparse coefficient Z.
Step 5, the input sample for determining support vector machines: by the total sparse coefficient of i-th pixel respective pixel block to
Measure zIWith the pixel value vector x of the block of pixelsIThe feature vector s of pixel when being combined to obtain final classificationI, obtain all pictures
Eigenmatrix S=[the s of vegetarian refreshments1,s1,…,sI,…,sLL]。
Step 6, the label that all pixels point is predicted with support vector machines (Support Vector Machine, SVM):
A part of column vector in sample characteristics matrix S is randomly selected as training sample, determines the class label of these column vectors, benefit
It is predicted to obtain the class label of all column vectors with libsvm classifier, the class label of column vector is each picture in image
The classification of vegetarian refreshments belongs to.
Step 7, display classification results: the class label of prediction is shown with gray level image, it is high that input can be obtained
Differentiate the final classification results figure of SAR image.
The present invention will be total to sparse model applied to High Resolution SAR image classification field, propose to combine using transform domain feature
Pixel value tag realizes the classification of High Resolution SAR image based on support vector machines, and obtains the fast high-efficient classification results of speed.
Compared with the prior art, the present invention has the following advantages:
1, the rarefaction representation of image, the prior art are mainly based upon the methodology of collective model in rarefaction representation in order to obtain
Handwriting practicing allusion quotation obtains the rarefaction representation coefficient of image with the linear expression of atom a small amount of in dictionary, to carry out using sparse coefficient
Image classification because dictionary learning during used a large amount of matrix inversion and decomposition computation so that dictionary learning need compared with
More times and calculating storage capacity；And the present invention is that High Resolution SAR image classification is realized based on sparse model altogether, the sparse table of image
Show it is based on analytic operator, and the learning process of operator avoids matrix inversion calculating, learning method speed faster, classification effect
Rate greatly improves.
2, different from the total sparse coefficient method for solving in the prior art applied in image classification, simply by dilute altogether
Thin model formation adds the method for soft-threshold just to obtain total sparse coefficient, and the present invention is the sparse mould altogether used in image denoising
Type algorithm, slightly modification adapt it to image classification process, are solved altogether using the method that augmentation Lagrange multiplier seeks optimization
Sparse coefficient makes full use of analytic operator.And the total sparse coefficient and pixel value of pixel are combined on this basis, then
The textural characteristics that image is utilized make detailed information when classification more acurrate, improve High Resolution SAR image classification result
Accuracy.
Detailed description of the invention
Fig. 1 is flow chart of the invention；
Fig. 2 is classification results figure of the present invention on the High Resolution SAR image that a width includes more detailed information, wherein
Fig. 2 (a) is the part forest image of 256 × 256 sizes, and Fig. 2 (b) is to be based on original pixel value as sample characteristics
The classification results of SVM.Fig. 2 (c) is the classification results that the soft-threshold that Sumit Shekhar is proposed is total to sparse coding, and Fig. 2 (d) is
The result figure of classification method proposed by the present invention；
Fig. 3 is classification results of the present invention in the true SAR image that a width size is 480 × 512, and wherein Fig. 3 (a) is
Original image comprising forest, airfield runway and common soil three classes, Fig. 3 (b) are directly to make the gray value of original pixels block
The classification effect encoded altogether for sample classification feature based on the simple soft-threshold that SVM, Fig. 3 (c) are Sumit Shekhar proposition
Fruit；Fig. 3 (d) is classifying quality figure of the invention.
Specific embodiment
In the following with reference to the drawings and specific embodiments to the detailed description of the invention.
The prior art is to learn dictionary using collective model, needs plenty of time and calculation amount.Although sparse parsing mould altogether
Type is also applied to facial image classification field, but image classification method solves when being total to sparse coefficient only simply using soft
Threshold method, classification accuracy are limited.And High Resolution SAR image classification field it is not applied to also.For these technical problem sheets
Invention expands research and innovation, proposes a kind of High Resolution SAR image classification method based on total sparse model.
Embodiment 1
It referring to Fig. 1 includes as follows that the present invention, which is a kind of High Resolution SAR image classification method based on total sparse model,
Step:
Step 1, input picture obtain initial pixel value matrix: choosing a High Resolution SAR image to be sorted, such as Fig. 2
(a) image Matlab software is read, is needed if input picture is the RGB image in three channels by it by image shown in
It is converted into gray-value image, the gray scale value matrix of image, each member in matrix are then directly obtained if single pass gray level image
A pixel in plain correspondence image.
M is intercepted centered on all pixels point1×M1The block of pixels of size, for convenience later data processing, will be each
Block of pixels transforms into a column.All pixels point obtains experiment sample X ∈ R in traversal imageM×LL, M is the pixel in a block of pixels
Point number, LL arrive the initial pixel value matrix X of input picture for the pixel number in High Resolution SAR image to be sorted.
Step 2, the initial sample for choosing analytic operator study: L column are randomly selected in initial pixel value matrix X and are constituted greatly
The small matrix Y for M × L, the initial sample matrix Y=[y as operator study1,y2,...,yI,...,yL], yIIndicate i-th
M where central pixel point1×M1Size block of pixels expansion column vector to get to learn analytic operator when initial sample moment
Battle array Y.
In order to make analytic operator study is more acurrate can also guarantee the other picture of every type with the initial sample matrix Y of artificial selection
The corresponding initial sample vector number of vegetarian refreshments is uniform, but will increase manpower and time in this way.And composition is randomly selected in the present invention
The method of initial sample matrix also can guarantee that analytic operator correctly learns while simple and convenient.
Step 3 learns analytic operator based on UNTF algorithm: mutually being tied with both subgradient and unified professional etiquette model tight frame is projected
The method of conjunction obtains analytic operator Ω using initial sample matrix Y renewal learning.
Step 4 solves sparse coefficient altogether based on Augmented Lagrange method: by analytic operator Ω and initial pixel value matrix
X is added antithesis parametric solution convex programming problem and obtains according to sparse analytic modell analytical model Z=Ω X altogether using Augmented Lagrange method
To total sparse coefficient Z.
Step 5 determines that the input sample of SVM classifier obtains the feature vector of all pixels point: byI pixel
The total sparse coefficient vector z of point respective pixel blockIWith the pixel value vector x of the block of pixelsIPixel when combining to obtain final classification
The feature vector s of pointI, obtain the eigenmatrix S=[s of all pixels point1,s1,…,sI,…,sLL].SVM classifier owns
Input sample is characterized matrix S.
Pixel value vector xIChoose the block of pixels vector that can be truncated to for original image；It is also possible to solve sparse altogether
The corresponding block of pixels vector of i-th pixel in the matrix that the X changed when coefficient is finally determined.
Step 6, the label that all pixels point is predicted with SVM classifier: one in sample characteristics matrix S is randomly selected
Divide column vector as training sample, determines the class label of these column vectors, predict to obtain all column using libsvm classifier
The class label of vector, the belonging kinds of each pixel as in image.
6.1, the input of training sample (train data) and training label (train label) is determined.
6.2, the nuclear parameter of libsvm, training SVM classifier model are determined.
6.3, input test sample (test data) obtains prediction label (predict into SVM classifier model
Label), test sample is characterized matrix S, and prediction label is classification results.
Step 7, display classification results: the class label of prediction is shown with gray level image, it is high that input can be obtained
Differentiate the final classification results figure of SAR image.
The present invention realizes total sparse model in the application in High Resolution SAR image classification field, and using sparse operator altogether
Learning method avoids matrix inversion process, and faster than collective model dictionary learning method speed, efficiency greatly improves.
Embodiment 2
Based on the High Resolution SAR image classification method of total sparse model with embodiment 1, projection time is wherein used in step 3
The method construct analytic operator that both gradient and unified professional etiquette model tight frame combine includes following steps:
3.1, initial analytic operator Ω is constructed0
Excessively complete matrix D=[d that a size is M × N is randomly generated1,d2,…,dq,…,dN], wherein dqFor column to
Amount, 1≤q≤N, N are the dimension of analytic operator, the artificial selection as needed of the size of N but have to be larger than M, are initially parsed
Operator Ω0, Ω0For transposed matrix, that is, Ω of excessively complete matrix D0=DT.3.2, by subgradient analytic operator projection formulaCalculate subgradient analytic operator Ωg, 0≤i≤Kmax1, Kmax1For the maximum number of iterations for seeking analytic operator, Ωi
For i-th of analytic operator.
Wherein
3.3, calculate Ωi+1
The method for taking drop gradient updates analytic operator, initial step length η is given, according to formula Ωi+1=PTF{PUN(Ωi-η
Ωg) calculate Ωi+1, PUN{ } indicates to project to unified professional etiquette norm space, PTF{ } expression projects to tight frame.
3.4, stopping criterion for iteration
As f (Ωi+1)≤f(Ωi) or i > Kmax1When terminate iterative process, f (Ωi) it is that a definition analytic operator misses
The function of difference, f (Ωi) it is equal to matrix ΩiThe sum of the absolute value of all elements in Y, obtains final analytic operator Ω=Ωi, no
Step-length is then modified, is enabledIt repeats step (3.2) and (3.3), until Ωi+1Meet stopping criterion for iteration and obtains analytic operator Ω
=Ωi。
With the sample matrix Y renewal learning operator, the original that signal data can change when solving sparse coefficient in step 4 is utilized
Reason, recalculates step 3 and obtains coefficient Z1=Ω Y, meets coefficient Z1It must be sparse.
The present invention is the rarefaction representation that High Resolution SAR image is obtained using analytic modell analytical model, avoids image and is based on comprehensive mould
It needs to spend a lot of time the problem of solving optimal dictionary when type rarefaction representation, analytic operator is simple and convenient when study, significantly
Reduce amount of storage and calculate the time, so that classification method speed is improved.
Embodiment 3
Based on the High Resolution SAR image classification method of total sparse model with embodiment 1-2, augmentation is wherein utilized in step 4
Lagrangian method solves sparse coefficient Z problem altogether:S.t.Z=Ω X, includes following steps:
4.1, initial pixel value matrix X is given0=X, analytic operator Ω are initially total to sparse coefficient Z0=Ω X0, initial parameter
Matrix B0For size and Z0Equal null matrix, constant parameter λ < 1, constant coefficient γ < 1.
4.2, pixel matrix X is calculated according to the following formulai+1, altogether sparse coefficient Zi+1With parameter matrix Bi+1, 0≤i≤
Kmax2, Kmax2For the maximum number of iterations for seeking total sparse coefficient.
Wherein
Bi+1=Bi-(Zi+1-ΩXi+1)
4.3, repeat step 4.2 until | | Zi+1-ΩXi+1||2>=ε i > Kmax2When terminate iterative process, ε is remote small
In 1 error constant coefficient, until obtaining final total sparse coefficient Z=Zi, ZiIt is i-th of total sparse coefficient.
The method of sparse coefficient and soft-threshold simple in facial image classification method altogether are solved in the present invention asks method different,
Sparse coefficient altogether is solved by the method for seeking optimization using augmentation Lagrange multiplier, makes full use of analytic operator, so that figure
The rarefaction representation of picture is more acurrate.
Embodiment 4
Based on the High Resolution SAR image classification method of total sparse model with embodiment 1-3, the present invention in steps of 5 will be every
The total sparse coefficient vector z of a pixelIWith the initial vector x of the pixel respective pixel blockIWhen combining to obtain final classification
The feature vector s of pixelISpecific steps include: the I column vector z taken in total sparse coefficient matrix ZI∈RNFinally to divide
The feature vector s of pixel when classIPreceding N row, which corresponds to the I column vector x of initial pixel value matrixI∈RMFor most
The feature vector s of pixel when classifying eventuallyIRear M row, obtain the feature vector s of final pixelI∈RM+N, all pixels point
Final feature vector composition characteristic vector matrix S=[s1,s2,...,sI,...,sLL]。
The present invention combines the total sparse coefficient and pixel value of pixel, makes image classification using the textural characteristics of image
When detailed information it is more acurrate, improve the accuracy of SAR image classification results.
A more detailed example is given below, the present invention is explained again.
Embodiment 5
Based on the High Resolution SAR image classification method of total sparse model with embodiment 1-4,
It is referring to Fig.1, of the invention that the specific implementation steps are as follows:
Step 1, input picture obtain initial pixel value matrix: choosing a High Resolution SAR image to be sorted, such as Fig. 3
(a) image shown in intercepts M centered on all pixels point1×M1The block of pixels of size, and transform into a column.Traverse the picture of full figure
Vegetarian refreshments obtains experiment sample X.
Specifically, a width size is L1×L2SAR image to be sorted, it is contemplated that picture is constituted centered on pixel
When plain block, edge pixel point will appear the inadequate problem of pixel block size, needs to handle original digital image data, first constructs one
It is a M all bigger than original image length and width1- 1 null matrix, i.e. size are (L1+M1-1)×(L2+M1- 1) null matrix X1, then make matrix X1
Middle L1×L2The element of big fraction is the pixel value of former SAR image.It avoids to intercept picture centered on marginal point in this way
Pixel block size inadequate problem when plain block.Matrix X1Random matrix can be taken when initial construction, because of edge pixel point
Be limited, be also when finally taking classification samples it is random, the probability for getting these marginal points is lower, on classification results influence not
Greatly.
After pretreatment image, takes block of pixels to traverse full figure centered on original image pixel according to capable sequence, LL can be obtained
A size is M1×M1Block of pixels, LL=L1×L2, transformed into a column xI∈RM*1, M=M1×M1, all pixels block composition
The initial pixel value matrix X=[x that one size is M × LL1,x2,...,xI,...,xLL]。
The size M of block of pixels1Value determined that but scale cannot be too according to the thickness of atural object textural characteristics in image
Small, usual smallest dimension is 3 × 3, is typically chosen the block of pixels of the sizes such as 5 × 5,7 × 7,9 × 9.
Step 2, the sample for choosing analytic operator study: the L column in initial pixel value matrix X are randomly selected, i.e., in original image
M is intercepted centered on any L pixel1×M1The vector being unfolded after the block of pixels of size constitutes sample matrix as initial characteristics
Y=[y1,y2,...,yI,...,yL], yIM where indicating some central pixel point1×M1Size block of pixels expansion column to
Amount, the size of matrix Y are M × L.
Pixel number L is randomly selected manually to be determined according to experiment concrete condition.
Step 3, study analytic operator: by projection subgradient and unified professional etiquette model tight frame (UniformNormalized
Tight Frame, UNTF) both combine, based on NAAOL algorithm learn analytic operator Ω ∈ RN×M。
The present invention projects to Ω in unified professional etiquette model tight frame, the specific steps are as follows:
(a) Ω is projected in tight frame (Tight Frame, TF), tight frame is exactly a spatial domain, then Ω projection can
Simply to be replaced with the singular value decomposition of Ω, it may be assumed that ΡTF{ Ω }=UWVT.Wherein PTF{ Ω } indicates projection, and U is orthogonal matrix,
Vector inside U is known as left singular vector, and W is diagonal matrix, and in addition to cornerwise element is all 0, the member on diagonal line is called usually
For singular value, V is orthogonal matrix, and the vector inside V is known as left singular vector, ()TThe transposition of representing matrix.
Singular value decomposition is common matrix decomposition mode, and also belonging in machine learning field fullly visible is method,
Purpose is to extract the most important feature of matrix.
(b) Ω is projected in row canonical frame (Uniform Normalized Frame, UN), when element is 0, is thrown
A vector is arbitrarily selected to replace when shadow in [- 1,1].When not projecting to fixed row canonical frame for 0 element,
It need to only be replaced with the standardization vector of this element.That is:
PUN{ Ω }=[PUN{wI}]IWherein PUN{ Ω } indicates projection, wIIt is the I row vector of Ω.
Wherein, vector v is any vector in critical field.
3.1, initial operator Ω is constructed0。
Matrix D=[d that a size is M × N is randomly generated1,d2,…,dq,…,dN], wherein dq(1≤q≤N) is row
Vector, N are the dimension for analyzing operator, the artificial selection as needed of the size of N but have to be larger than M.
Repeat the above steps (a) and (b), obtains the transposed matrix D of initial operator0, then Ω0=D0 T。
3.2, by subgradient analytic operator projection formulaCalculate subgradient analytic operator ΩG, 0≤i≤
Kmax1, Kmax1For the maximum number of iterations for seeking analytic operator, ΩiFor i-th of analytic operator；
Wherein
3.3, it updates and calculates analytic operator Ωi+1。
The mode update operator of drop gradient is taken, initial step length η is given, η is the constant coefficient much smaller than 1, can choose 10-5、10-7、10-8Deng according to formula Ωi+1=PTF{PUN(Ωi-ηΩG) update calculating analytic operator Ωi+1。
3.4, stopping criterion for iteration.
As f (Ωi+1)≤f(Ωi) when terminate iterative process, f (Ωi) it is the function for defining analytic operator error, determine f
(Ωi) it is matrix ΩiThe sum of the absolute value of all elements in Y obtains analysis operator Ω=Ωi+1=[ω1 T,ω1 T,…,
ωi T,…,ωL T], ωiIt is the i-th row vector of Ω.Otherwise step-length is modified, is enabledRepeat step 3.2 and 3.3 obtain it is new
Ωi+1, analytic operator Ω=Ω is obtained until meeting stopping criterion for iterationi。
Occur not terminating phenomenon in iterative process in order to prevent, can be set one and seek greatest iteration during analytic operator
Number Kmax1, size is by needing to determine.
There are many kinds of the methods of analytic operator Ω study, and the NAAOL algorithm that the present invention uses not only accurately has learnt parsing
Operator, it is thus also avoided that matrix inversion calculates, so that analytic operator computation rate greatly improves.
Step 4, by above-mentioned operator Ω ∈ RL×MWith initial pixel value matrix X ∈ RM×LL, according to sparse analytic modell analytical model Z=altogether
Ω X solves sparse coefficient Z altogether using based on Augmented Lagrange method.
Nonconvex programming is solved the problems, such as based on Augmented Lagrange method:Wherein Z meets Z=Ω X.It is right with one
Even parameter B ∈ RL×M, and a penalty term < B, Ω is addedi+1X-Z >, then above-mentioned minimization problem can be equivalent to solve one newly
Objective function g (X, Z, B) minimum value, this objective function is convex programming.
Wherein, γ is constant coefficient greater than 0, can choose 0.1,0.2,0.3 etc., specific value according to experiment experience and
It is fixed.
Seeking total sparse coefficient Z, specific step is as follows:
4.1, determine initial pixel value matrix X0=X, analytic operator Ω are initially total to sparse coefficient Z0=Ω X0, initial parameter square
Battle array B0For size and Z0Equal null matrix, constant parameter λ > 0, constant coefficient γ > 0.
4.2.1, X is calculated by following formulai+1。
4.2.2, calculate sparse coefficient
Wherein
4.2.3, penalty term B is calculatedi+1。
Bi+1=Bi-(Zi+1-ΩXi+1)。
4.3, stopping criterion for iteration.
When | | Zi+1-ΩXi+1||2Terminate iterative process when >=ε, ε is the constant coefficient much smaller than 1, and obtains Z=Zi=Ω
Xi.Otherwise, step 4.2 is repeated until meeting stopping criterion for iteration.
Similarly occur not terminating phenomenon in iterative process in order to prevent, can be set one seek total sparse coefficient during
Maximum number of iterations Kmax2, the size of maximum number of iterations determines by experiment experience.
Step 5 takes I column vector z in total sparse coefficient matrix ZI∈RNThe feature vector of pixel when for final classification
sIPreceding N row, which corresponds to the I column vector x of initial pixel value matrixI∈RMThe feature of pixel when for final classification
Vector sIRear M row, obtain the feature vector s of final pixelI∈RM+N, the final feature vector of all pixels point forms special
Levy vector matrix S=[s1,s2,...,sI,...,sLL], 1≤I≤LL.
Step 6 randomly selects a part of column vector in sample characteristics matrix S as training sample, determines training sample
Tag along sort, using libsvm classifier classify, a kind of common classifier when Libsvm is image classification.
6.1, the input of training sample (train data) and training label (train label) is determined.
Randomly select r column vector composing training sample matrix SS=[ss in eigenmatrix S1,ss2,...,
ssq,...,ssr], ssqFor any column vector in matrix S, SS ∈ R(M+N)×r。
The Characteristic Number r randomly selected is determined as needed, to guarantee that correctness and the rapidity of experimental result are generally selected
Take the 10 of whole sample numbers.
Training label is the corresponding tag along sort of these training samples, i.e., each column vector ssqCorresponding in former SAR image
The corresponding label of block of pixels is that the pixel position at the block of pixels center corresponds to same position in former SAR image standard drawing
Pixel corresponding to tag along sort.
6.2, libsvm parameter, training sorter model are determined.
Nuclear parameter is selected according to test of many times Comparative result in Libsvm, and the experimental selection in the present invention is RBF core.
6.3, the input of test sample (test data) is determined.
Test sample of the invention is complete eigenmatrix S ∈ R(M+L)×LL, obtained prediction label is that a size is
The column vector of LL × 1.
Step 7, the label matrix that the label vector of prediction is reduced into original image size by function, and each pixel with
Its prediction label is corresponding, and this matrix is shown with gray level image, and final classification results figure can be obtained.
A panel height fast and accurately can be differentiated into SAR image classification according to above-mentioned steps, it is sharp by means of the present invention
The belonging kinds of all samples can be fast and accurately obtained with fraction known label sample.
Technical effect of the invention is verified and illustrated below by emulation experiment.
Embodiment 6
Based on the High Resolution SAR image classification method of total sparse model with embodiment 1-5.
Simulated conditions and content:
Running environment is Windows, software Matlab2016b, is based on SVM with method and original pixel value of the invention
The method that the method for classification, simple soft-threshold seek total sparse coefficient carries out classification experiments to two width SAR images respectively, from classification
As a result region consistency, mistake divides situation, edge holding etc. assessment.
Gradient step parameter η=10 when experiment-7, seek the maximum number of iterations K of analytic operatormax1=50000, constant parameter
The maximum number of iterations K of total sparse coefficient is sought in λ=γ=0.1, error constant coefficient ε=0.001max2=500.
The simulation experiment result:
Fig. 2 be include more detailed information true SAR image classification results.Wherein Fig. 2 (a) is that a width size is
The part forest image of 256 × 256 pixels, the uniform region of texture is road and common soil, the non-uniform area of texture in figure
Domain is trees.Fig. 2 (b) is the classification results for being directly based upon the svm classifier method of pixel point feature and obtaining；Fig. 2 (c) is that application is soft
The classification results that the method that threshold method seeks total sparse coefficient obtains；Fig. 2 (d) is that method of the invention classify to Fig. 2 (a)
The result figure arrived.These three classification results be all 7 × 7 scales under carry out, that is, the pixel block size M intercepted1=7.
Compare this few width figure can be seen that based on original image vegetarian refreshments classification results Fig. 2 (b) apparently without provide correctly point
Class has only separated rough profile information.Compare Fig. 2 (c) and Fig. 2 (d), discovery Fig. 2 (d) is punished at different classes of edge
Class is more acurrate, also higher for the discrimination in the trees region of rough grain；In corresponding original image 2 (a) on leading diagonal direction
Completely black part is road, wherein the common path that the region for having part grey is, hence it is evident that identifying for Fig. 2 (d) is more general
Logical path；The common soil in original image Fig. 2 (a) rightmost middle section rough grain region is corresponded to, the present invention can be accurate
Be identified as common soil, and Fig. 2 (c) this part mistake is divided into equally be rough grain trees classification, this part is in the present invention
The method of proposition identifies that lower result is more accurate.
Embodiment 7
Based on the High Resolution SAR image classification method of total sparse model with embodiment 1-5, simulated conditions and content with implementation
Example 6.
Fig. 3 is the classification results for the true SAR image that a width size is 480 × 512, and Fig. 3 (a) is to include forest, airport
The original image of runway and common soil three classes；Fig. 3 (b) is directly special using the gray value of original pixels block as pixel classification
Levy the result obtained based on SVM classifier classification；Fig. 3 (c) is to ask the classification of total sparse coefficient method to imitate using simple soft-threshold
Fruit；Fig. 3 (d) is classifying quality figure of the invention.These three classification results be all 9 × 9 scales under carry out, that is, intercept
Pixel block size M1=9.
Comparing each figure from Fig. 3 can be seen that method of the Fig. 3 (b) based on original pixel value can only find out rough profile, should
Trees and airfield runway are almost identified as one kind by method.And simple soft-threshold seeks classification results Fig. 3 of total sparse coefficient method
(c) forest part does not identify substantially in；Result figure 3 (d) of the invention by this block region most contents identify for
Trees, recognition accuracy improve.That path of image is crossed in corresponding middle section original image 3 (a), and Soft thresholding mainly will
This road Identification is forest part, and Dark grey is shown as in Fig. 3 (c)；And method of the invention is largely to identify it
It is divided into road, black is shown as in Fig. 3 (d), the present invention is more accurate in the identification of detail section.
In brief, the invention discloses a kind of High Resolution SAR image classification method based on total sparse model, belong to figure
As processing technology field, it is confined to indicate that image causes the classification time multiple with comprehensive sparse model when solving SAR image classification
The high technical problem of miscellaneous degree.Its assorting process are as follows: intercept block of pixels centered on pixel in SAR image to be sorted and constituted
Initial pixel value matrix X；Selected part pixel value vector constitutes the initial learning sample of analytic operator；Utilize projection subgradient and system
The method that a line specification tight frame combines learns analytic operator Ω；Optimization problem is asked to obtain based on Augmented Lagrange method
Sparse coefficient Z altogether；The total sparse coefficient vector of each pixel respective pixel block is mutually tied with the pixel value vector of the block of pixels
Conjunction obtains the feature vector of pixel when final classification；It randomly selects Partial Feature vector and constitutes test sample SS, based on SVM points
The classification of class device, obtains the prediction label of each pixel feature vector of full figure；Prediction label result is shown with gray level image.
The present invention can quickly acquire the rarefaction representation of image, ensure that the timeliness and classification accuracy of SAR image classification.
Claims (3)
1. a kind of High Resolution SAR image classification method based on total sparse model, which is characterized in that comprise the following steps that
Step 1, input picture obtain initial pixel value matrix X: a High Resolution SAR image to be sorted are chosen, with all pictures
M is intercepted centered on vegetarian refreshments1×M1The block of pixels of size, each block of pixels transform into a column, traverse all pixels point in image and obtain reality
Test sample X ∈ RM×LL, M is the pixel number in a block of pixels, and LL is the pixel in High Resolution SAR image to be sorted
Number；
Step 2, the initial sample for choosing analytic operator study: L column composition size is randomly selected in initial pixel value matrix X is
The matrix Y of M × L, the initial sample matrix Y=[y as operator study1,y2,…,yI,...,yL], yIIndicate imago in i-th
M where vegetarian refreshments1×M1The column vector of the block of pixels expansion of size；
Step 3, study analytic operator: it is combined with both projection subgradient and unified professional etiquette model tight frame, utilizes initial sample
Matrix Y iterative learning obtains analytic operator Ω, comprising the following steps:
3.1) initial analytic operator Ω is constructed0；
Excessively complete matrix D=[d that a size is M × N is randomly generated1,d2,…,dq,…,dN], wherein dqFor column vector, 1≤
Q≤N, N are the dimension of analytic operator, the artificial selection as needed of the size of N but have to be larger than M, obtain initial analytic operator
Ω0, Ω0For transposed matrix, that is, Ω of excessively complete matrix D0=DT；
3.2) by subgradient analytic operator projection formulaCalculate subgradient analytic operator Ωg, 0≤i≤Kmax1, Kmax1
For the maximum number of iterations for seeking analytic operator, ΩiFor i-th of analytic operator；
Wherein
3.3) Ω is calculatedi+1；
The method for taking drop gradient updates analytic operator, initial step length η is given, according to formula Ωi+1=PTF{PUN(Ωi-ηΩg)}
Calculate Ωi+1, PUN{ } indicates to project to unified professional etiquette norm space, PTF{ } expression projects to tight frame；
3.4) stopping criterion for iteration；
As f (Ωi+1)≤f(Ωi) or i > Kmax1When terminate iterative process, f (Ωi) it is a definition analytic operator error
Function, f (Ωi) it is equal to matrix ΩiThe sum of the absolute value of all elements in Y, obtains final analytic operator Ω=Ωi, otherwise repair
Change step-length, enablesIt repeats step (3.2) and (3.3), until Ωi+1Meet stopping criterion for iteration obtain analytic operator Ω=
Ωi；
Step 4 solves sparse coefficient altogether based on Augmented Lagrange method: by analytic operator Ω and initial pixel value matrix X, root
Antithesis parametric solution convex programming problem is added and is total to using Augmented Lagrange method according to sparse analytic modell analytical model Z=Ω X altogether
Sparse coefficient Z；
Step 5, the input sample for determining SVM classifier: by the total sparse coefficient vector z of i-th pixel respective pixel blockIWith
The pixel value vector x of the block of pixelsIThe feature vector s of pixel when being combined to obtain final classificationI, obtain all pixels point
Eigenmatrix S=[s1,s2,…,sI,…,sLL]；
Step 6, the label that all pixels point is predicted with SVM classifier: a part randomly selected in sample characteristics matrix S arranges
Vector determines the class label of these column vectors as training sample, predicts to obtain all column vectors using libsvm classifier
Class label be each pixel in image classification ownership；
Step 7, display classification results: the class label of prediction is shown with gray level image, input high-resolution can be obtained
The final classification results figure of SAR image.
2. utilizing augmentation in step 4 as described in claim 1 based on the High Resolution SAR image classification method of total sparse model
Lagrangian method solves sparse coefficient Z problem altogether:S.t.Z=Ω X, includes following steps:
4.1) initial pixel value matrix X is given0=X, analytic operator Ω are initially total to sparse coefficient Z0=Ω X0, initial parameter matrix
B0For size and Z0Equal null matrix, constant parameter λ < 1, constant coefficient γ < 1；
4.2) pixel matrix X is calculated according to the following formulai+1, altogether sparse coefficient Zi+1With parameter matrix Bi+1, 0≤i≤Kmax2,
Kmax2For the maximum number of iterations for seeking total sparse coefficient；
Wherein
Bi+1=Bi-(Zi+1-ΩXi+1)；
4.3) repeat step (4.2) until | | Zi+1-ΩXi+1||2≥ε2Or i > Kmax2When terminate iterative process, ε2It is to be much smaller than
1 constant coefficient, until obtaining final total sparse coefficient Z=Zi。
3. as described in claim 1 based on the High Resolution SAR image classification method of total sparse model, which is characterized in that step 5
The total sparse coefficient vector z by each pixelIWith the initial vector x of the pixel respective pixel blockIIt combines to obtain
The feature vector s of pixel when final classificationISpecific steps include: the I column vector z taken in total sparse coefficient matrix ZI∈
RNThe feature vector s of pixel when for final classificationIPreceding N row, which corresponds to the I column vector of initial pixel value matrix
xI∈RMThe feature vector s of pixel when for final classificationIRear M row, obtain the feature vector s of final pixelI∈RM+N,
The final feature vector composition characteristic vector matrix S=[s of all pixels point1,s2,...,sI,...,sLL]。
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710329004.8A CN107220659B (en) | 2017-05-11 | 2017-05-11 | High Resolution SAR image classification method based on total sparse model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710329004.8A CN107220659B (en) | 2017-05-11 | 2017-05-11 | High Resolution SAR image classification method based on total sparse model |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107220659A CN107220659A (en) | 2017-09-29 |
| CN107220659B true CN107220659B (en) | 2019-10-25 |
Family
ID=59944891
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710329004.8A Active CN107220659B (en) | 2017-05-11 | 2017-05-11 | High Resolution SAR image classification method based on total sparse model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107220659B (en) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109543767A (en) * | 2018-11-29 | 2019-03-29 | 东南大学 | Urban land use automatic classification method under big data environment |
| CN110059705B (en) * | 2019-04-22 | 2021-11-09 | 厦门商集网络科技有限责任公司 | OCR recognition result judgment method and device based on modeling |
| CN110826599B (en) * | 2019-10-16 | 2023-04-18 | 电子科技大学 | Sparse representation sample distribution boundary retention feature extraction method |
Family Cites Families (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8447119B2 (en) * | 2010-03-16 | 2013-05-21 | Nec Laboratories America, Inc. | Method and system for image classification |
| US8718380B2 (en) * | 2011-02-14 | 2014-05-06 | Mitsubishi Electric Research Laboratories, Inc. | Representing object shapes using radial basis function support vector machine classification |
| CN103077503A (en) * | 2013-01-09 | 2013-05-01 | 西安电子科技大学 | Discrete cosine transform (DCT) dictionary synchronous sparse representation-based synthetic aperture radar (SAR) image speckle reduction method |
| CN104215937B (en) * | 2014-09-19 | 2017-01-11 | 西安电子科技大学 | Airborne radar space time adaptation processing method based on environment dynamic perception |
-
2017
- 2017-05-11 CN CN201710329004.8A patent/CN107220659B/en active Active
Also Published As
| Publication number | Publication date |
|---|---|
| CN107220659A (en) | 2017-09-29 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Shen et al. | Detection of stored-grain insects using deep learning | |
| CN110414377B (en) | Remote sensing image scene classification method based on scale attention network | |
| CN109344736B (en) | Static image crowd counting method based on joint learning | |
| CN110348319B (en) | Face anti-counterfeiting method based on face depth information and edge image fusion | |
| Shen et al. | Object skeleton extraction in natural images by fusing scale-associated deep side outputs | |
| CN109359559B (en) | Pedestrian re-identification method based on dynamic shielding sample | |
| Kae et al. | Augmenting CRFs with Boltzmann machine shape priors for image labeling | |
| CN105260738B (en) | High-resolution remote sensing image change detecting method and system based on Active Learning | |
| CN101667245B (en) | Human face detection method by cascading novel detection classifiers based on support vectors | |
| CN108647602B (en) | A kind of aerial remote sensing images scene classification method determined based on image complexity | |
| CN103886342B (en) | Hyperspectral image classification method based on spectrums and neighbourhood information dictionary learning | |
| CN106203523A (en) | The classification hyperspectral imagery of the semi-supervised algorithm fusion of decision tree is promoted based on gradient | |
| CN109784203A (en) | Weakly supervised x-ray image contraband inspection method based on delamination propagation and activation | |
| CN105893925A (en) | Human hand detection method based on complexion and device | |
| CN103824079B (en) | Multi-level mode sub block division-based image classification method | |
| CN106023257A (en) | Target tracking method based on rotor UAV platform | |
| CN107220659B (en) | High Resolution SAR image classification method based on total sparse model | |
| CN105930794A (en) | Indoor scene identification method based on cloud computing | |
| CN112926652B (en) | Fish fine granularity image recognition method based on deep learning | |
| CN109886147A (en) | A kind of more attribute detection methods of vehicle based on the study of single network multiple-task | |
| CN110008899B (en) | Method for extracting and classifying candidate targets of visible light remote sensing image | |
| CN112163450A (en) | Based on S3High-frequency ground wave radar ship target detection method based on D learning algorithm | |
| CN108764302A (en) | A kind of bill images sorting technique based on color characteristic and bag of words feature | |
| CN112560967A (en) | Multi-source remote sensing image classification method, storage medium and computing device | |
| CN111382766A (en) | Equipment fault detection method based on fast R-CNN |
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 |