CN106326938A - SAR image target discrimination method based on weakly supervised learning - Google Patents

Abstract

The invention discloses an SAR image target discrimination method based on weakly supervised learning, which is mainly used to solve problems of a prior art such as low discrimination performance and high sample marking costs. The SAR image target discrimination method comprises steps that in a training phase, locality-constrained linear coding LLC characteristics of a positive image sample set and locality-constrained linear coding LLC characteristics of a negative image sample set are respectively extracted, and the negative sample set is used to train a potential latent dirichlet allocation LDA model, which is used to select an initial positive sample set from the positive image sample set to iteratively train a second-class SVM discriminator, and an optimal discriminator is acquired; in a testing phase, LLC characteristics of a testing sample set are extracted, and the acquired optimal discriminator is used to discriminate the testing sample set. The SAR image target discrimination method based on the weakly supervised learning is advantageous in that the discrimination performance is close to the fully-supervised SVM discriminator, and at the same time, costs of manual marking are reduced, and therefore practicability is provided; by comparing with a first-class SVDD discriminator of clutter training, the discrimination performance in a complicated scene is better, and the discrimination method provided by the invention is suitable for the SAR image target discrimination.

Description

SAR image target discrimination method based on Weakly supervised study

Technical field

The invention belongs to Radar Technology field, particularly to a kind of target discrimination method, can be used in synthetic aperture radar SAR image differentiates target effectively.

Background technology

Radar imaging technology is to grow up the 1950's, and obtained in 60 years afterwards advancing by leaps and bounds sends out Exhibition, at present, in military affairs, agricultural, geology, ocean, disaster, paint all many-sides such as survey and be widely used.Synthetic aperture Radar SAR has that round-the-clock, round-the-clock, resolution is high and the feature such as penetration power is strong, becomes current earth observation and military affairs are detectd The important means examined.

SAR image automatic target detection is the advanced subject of current SAR application, has important Research Significance and widely Application prospect.U.S.'s Lincoln laboratory proposes the tertiary treatment flow chart of SAR image automatic target detection and is widely used. This flow process uses one layering attention mechanism, and it realizes process and is: first, view picture SAR image carries out detection process, removes figure It is clearly not mesh target area in Xiang, obtains potential target region；Then, potential target area is carried out at target discriminating Reason, to reject natural clutter false-alarm therein, or rejects the region big or less than target；By detection and the mirror of target In the other stage, obtain target region of interest ROI；Finally, then to target ROI Classification and Identification is carried out.

Proposing a lot of SAR image target discrimination method in existing document, conventional two classes having full supervision support vector Machine SVM descriminator and a class Support Vector data description SVDD descriminator.Two class SVM descriminators of the full supervision of training need greatly The training sample of amount labelling, if first step detection is based on unsupervised, then carries out handmarking to the testing result obtained, Not only time-consuming but also effort, simultaneously for sample that is fuzzy or that block, may cause the labelling of mistake, affect the study of descriminator；One Class SVDD descriminator has only to class data and can be carried out study, it is generally the case that clutter data is easier to obtain, although it Decreasing the loaded down with trivial details of handmarking, but for relative target, clutter data is more complicated, kind is more, the model that training obtains SAR image target and the discriminating of clutter under complex scene can not be well adapted for, and then affect last discriminating performance.

Summary of the invention

Present invention aims to above-mentioned the deficiencies in the prior art, propose a kind of SAR based on Weakly supervised study figure As target discrimination method, it is differentiating that performance reduces handmarking's sample while close with the complete two class SVM descriminators supervised Cost, and to the target of complex scene SAR image, a class SVDD descriminator of the clutter training that compares differentiates that performance is more excellent.

The technical scheme is that and be achieved in that:

One, technical thought

In actual applications, it is loaded down with trivial details and time-consuming for being marked the sample after target detection, simultaneously for blocking or The labelling of fuzzy sample is also difficult, and the labelling for whether there being target in the figure before detection is the easiest. So-called Weakly supervised information refers to the labelling whether having target in the figure before detecting, and the image containing target is designated as positive image；Do not contain The image of target is designated as negative image.Present invention introduces the discriminating for SAR image target of the Weakly supervised thought.With traditional Radix Triplostegiae Grandiflorae SAR image is detected by number CFAR, and the testing result composition negative image sample set that negative image obtains, is all negative sample, it is not necessary to Handmarking, and the testing result that positive image obtains forms positive image pattern collection, existing positive sample also has negative sample, utilizes negative figure It is all that the information of negative sample is concentrated from positive image pattern and picked out initial positive sample set as sample set, then utilizes negative image to bear sample This collection and the initial just sample set complete two class SVM descriminators supervised of training selected, the descriminator that recycling obtains is from the most just Sample set is selected the two class SVM descriminators that positive sample set repetitive exercise is supervised entirely, until obtaining the descriminator of optimum.Finally, With the optimum descriminator obtained, test sample collection is carried out target discriminating.

Two, technical scheme

According to above-mentioned principle, technical scheme includes the following:

A, training step

(A1) image pattern collection X is aligned⁺Negative sample collection X with negative image^-In each training sample extract M dimension local limit Property processed coding LLC feature, M=1024, wherein, positive image pattern collection X⁺In there is no handmarking, both comprised positive sample and also comprised Negative sample:

(A11) each training sample is extracted intensive Scale invariant features transform SIFT feature, and to all training samples Intensive Scale invariant features transform SIFT feature Kmeans be polymerized to M class；

(A12) with M cluster centre structure code book CB, i.e. M cluster centre is by arranging the matrix rearranged as code book CB；

(A13) with the code book CB of structure, the intensive Scale invariant features transform SIFT feature of each training sample is carried out office The restricted coding in portion, obtains the local limit coding LLC feature after training sample coding；

(A2) with the negative sample collection X of negative image^-Train potential Di Li Cray distribution LDA topic model, and potential with this Di Li Cray distribution LDA topic model from positive image pattern collection X⁺In select initial positive sample set

(A3) with negative sample collection X^-With initial positive sample setThe two class SVM descriminators that iteration is supervised entirely, obtain optimum Descriminator；

B, testing procedure

(B1) each test sample extraction M dimension local limit coding LLC feature test sample concentrated:

(B11) each test sample is extracted intensive Scale invariant features transform SIFT feature；

(B12) with the code book CB of structure in training step (A12), the intensive scale invariant feature of each test sample is become Change SIFT feature and extract local limit coding, obtain the local limit coding LLC feature of each test sample；

(B2) the optimum descriminator obtained with training encodes LLC feature to test specimens according to the local limit of test sample Originally differentiate, obtain identification result.

The present invention compared with prior art, has the advantage that

1, the cost of handmarking in full supervision is reduced

Existing full supervision SAR image target discrimination method needs to be marked, all samples in training set for rear The training of continuous full supervision two class descriminators provides enough training datas, but training sample is carried out handmarking the most time-consuming but also Effort, simultaneously for sample that is fuzzy or that block, will also result in labelling difficulty, affects the training of descriminator.

The present invention utilizes the positive image pattern of existing negative sample collection never handmarking to concentrate and constantly selects positive sample Collection, is used for the training of two class SVM descriminators of full supervision and SAR image is carried out target discriminating, decreasing the one-tenth of handmarking This, in actual applications, have more practicality.

The discriminating performance of the class SVDD descriminator 2, comparing clutter training is more excellent

An existing class descriminator negative sample collection is trained, and with the class descriminator obtained, SAR image is carried out mesh Mark differentiates, the negative sample concentrated due to negative sample is more complicated, and kind is more, and the model that training obtains can not well adapt to complexity Target and the discriminating of clutter under scene.And the present invention utilizes Weakly supervised study, from the sample set of positive image, constantly select positive sample This collection, repetitive exercise is entirely supervised two class SVM descriminators, is obtained optimum descriminator, compare a class SVDD descriminator of clutter training Discriminating performance more excellent.

Below in conjunction with accompanying drawing and example, the present invention is described in further detail.

Accompanying drawing explanation

Fig. 1 is the flowchart of the present invention；

Fig. 2 is the negative sample exemplary plot that the present invention tests the negative image of use；

Fig. 3 is the sample instantiation figure that the present invention tests the positive image of use；

Fig. 4 is the test set sample instantiation figure that the present invention tests use.

Detailed description of the invention

With reference to Fig. 1, the realization of the present invention is divided into training and two stages of test, and its step is as follows:

One, the training stage

Step 1, extracts training sample and concentrates the intensive Scale invariant features transform SIFT feature of each sample.

(1.1) input training sample set X={X⁺,X^-, wherein, X⁺Being the sample set of positive image, existing positive sample also has negative Sample, X^-Being the negative sample collection of negative image, all samples are negative sample；

(1.2) the intensive Scale invariant features transform SIFT feature of extraction training sample concentration training sample x:

(1.2a) training sample x is first carried out two norm normalizeds, obtains the training sample after normalization:Reset the Gaussian template M that size is 5 × 5:

$M=\left[\begin{array}{ccccc}0.0030& 0.0133& 0.0219& 0.0133& 0.0030\\ 0.0133& 0.0596& 0.0983& 0.0596& 0.0133\\ 0.0219& 0.0983& 0.1621& 0.0983& 0.0219\\ 0.0133& 0.0596& 0.0983& 0.0596& 0.0133\\ 0.0030& 0.0133& 0.0219& 0.0133& 0.0030\end{array}\right];$

Wherein norm represents two norms seeking sample；

(1.2b) by the sample x after Gaussian template M and normalization_NConvolution, the training sample after being processed: I=x_N* M, wherein, * represents and seeks convolution；

(1.2c) by one-dimensional horizontal gradient template A=[-0.5,0,0.5] and vertical gradient template A^T(1.2b) is obtained Training sample I carry out convolution, obtain the horizontal gradient figure G of this training sample_x=I*A and vertical gradient map G_y=I* Α^T, structure Become the gradient amplitude figure G of this sample_mWith gradient direction figure G_o:

$G_m=\sqrt{{\left(G_x\right)}^2+{\left(G_y\right)}^2},$

$G_o=arctan\left(\frac{G_y}{G_x}\right)$

Wherein, arctan is arctan function, and T represents that asking transposition, * to represent seeks convolution；

(1.2d) in training sample x, take the section P that size is 16 × 16, and respectively at gradient amplitude figure G_mAnd ladder Degree directional diagram G_oIn correspondence position take region that size is 16 × 16 gradient amplitude figure g as section P_mWith gradient direction figure g_o；

(1.2e) by section P with 4 pixels for step-length grid division without overlapping, obtaining 16 sizes is the community of 4 × 4 Territory, by these 16 zonules by from top to bottom, order from left to right sorts, and obtains the zonule of jth 4 × 4 in section P Center: O_j=(O_jx,O_jy), wherein O_jxAnd O_jyBe respectively this zonule center section P in width and height:

$O_{jx}=2.5+4\×\mathrm{int}\left(\frac{j}{4}\right),$

$O_{jy}=2.5+4\×\left(\mathrm{mod}\right(\frac{j}{4})-1)$

Wherein, int () represents and takes business's computing, and mod () represents complementation, j ∈ [1,16] and be integer；

(1.2f) position (l in section P is calculated_x,l_y) gradient amplitude in the region of jth 4 × 4 weights by the pixel at place Value w:

W=(1-| l_x-O_jx|)·I(|l_x-O_jx| ＜ 4) × (1-| l_y-O_jy|)·I(|l_y-O_jy| ＜ 4),

Each pixel in section P is calculated its gradient amplitude weighted value to the zonule of jth 4 × 4, and will be every The position that the gradient amplitude weighted value of the zonule of jth 4 × 4 is according to pixels put in section P by one pixel arranges, Obtain the P amplitude weighting matrix w for the zonule of jth 4 × 4 that cuts into slices_j, wherein, (l_x,l_y) it is that pixel is in section P Position coordinates, i.e. l_xAnd l_yIt is respectively pixel width in section P and height, l_x,l_y∈ [1,16] and be integer, I are instruction letter Number；

(1.2g) according to amplitude weighting matrix w_j, calculate the section P weighted gradient amplitude to the zonule of jth 4 × 4 Figure: g_mj=g_m·w_j, wherein, represent the dot product of corresponding element, g_mRepresent the gradient amplitude figure of section P, w_jRepresent section P pin Amplitude weighting matrix to the zonule of jth 4 × 4；

(1.2h) respectively in weighted gradient map of magnitudes g_mjWith gradient amplitude directional diagram g_oIn with point (O_jx,O_jyPosition centered by) Put and take the region that size is 4 × 4, as the weighted gradient map of magnitudes g ' of jth 4 × 4 zonule in section P_mjAnd gradient direction Figure g '_oj；

(1.2i) [0,2 π] being divided into 8 directions, i-th Direction interval isI ∈ [1,8] and For integer, calculate the gradient direction figure g ' of the zonule of jth 4 × 4_ojWeighted gradient amplitude corresponding in 8 directions and, Weighted gradient amplitude vector H to 8 dimensions_j；

(1.2j) (1.2f)-(1.2i) is repeated in other zonules of 4 × 4 in section P, obtain the weighting ladder of its 8 dimension Degree amplitude vector, these 16 weighted gradient amplitude vector composition column vector d=[H₁,...,H_j,...,H₁₆]^T, wherein T represent turn Putting, d is 128 corresponding for section P dimension Scale invariant features transform SIFT feature；

(1.2k) training sample x is carried out from left to right, from top to bottom with Gridding length L=16, sliding step step=6 Slip obtains the grid that K size is 16 × 16, and the grid to each 16 × 16 extracts corresponding 128 by (1.2d) to (1.2j) Dimension Scale invariant features transform SIFT feature, constitutes intensive Scale invariant features transform SIFT feature D=of this sample x [d₁,...,d_n,...,d_K], wherein, d_nIt is the 128 dimension scale invariant feature that in training sample x, the n-th 16 × 16 grid obtains Conversion SIFT feature, n ∈ [1, K] and be integer,

Wherein,Represent that the size obtained is 16 × 16 Meshes number, W and H is respectively width and the height of sample x, and floor (a) represents the maximum integer taking no more than a；

(1.3) all samples in training set X being repeated step (1.2), the intensive Scale invariant obtaining all samples is special Levy conversion SIFT feature.

Step 2, according to the intensive Scale invariant features transform SIFT feature of all samples of training sample set, builds code book.

(2.1) training sample is concentrated the intensive Scale invariant features transform SIFT feature of all samples by row combination, obtains To the Scale invariant features transform SIFT feature collection S that size is 128 × (K × N), training sample during wherein N is training set X Number；

(2.2) with Kmeans feature set S is polymerized to M class, and by M cluster centre obtaining by row arrangement, obtains one Code book CB=[the cb of 128 × M₁,...,cb_m...,cb_M], wherein cb_mIt it is m-th cluster centre.

Step 3, carries out local to the intensive Scale invariant features transform SIFT feature of each sample that training sample is concentrated Restricted coding, obtains local limit coding LLC feature.

(3.1) with the code book CB obtained in step 2, the intensive Scale invariant features transform SIFT feature of training sample x is entered Row local limit encode, obtain local limit coding LLC feature:

(3.1a) the intensive Scale invariant features transform SIFT feature D=[d to training sample x₁,...,d_n,...,d_KIn] The n-th 128 dimension Scale invariant features transform SIFT feature d_n, calculate d_nWith the Euclidean distance of codebook element each in code book CB, Find 5 codebook element that Euclidean distance is minimum, and be arranged to make up local loop LB by row_n；

(3.1b) respectively by Scale invariant features transform SIFT feature d_nCorresponding local loop LB_nIn 5 elements Euclidean distance, divided by itself and the Euclidean distance sum of 5 elements, obtains local loop LB_nIn code coefficient corresponding to 5 elements

(3.1c) by code coefficient composition column vector corresponding for codebook element all of in code book CB, feature d is obtained_nVolume Code coefficient c_n, wherein except local loop LB in code book CB_nIn codebook element outside, the coding system that other codebook element is corresponding Number is 0；

(3.1d) the intensive Scale invariant features transform SIFT feature D=[d to training sample x₁,...,d_n,...,d_KIn] K 128 dimension Scale invariant features transform SIFT feature encode by step (3.1a)-(3.1c), obtain corresponding coding Coefficient matrix c=[c₁,...,c_n,...,c_K], each row element in code coefficient matrix c is sought its maximum, is trained The M dimension local limit coding LLC feature of sample x；

(3.2) to training sample set X={X⁺,X^-The intensive Scale invariant features transform SIFT of all training samples in } Feature is encoded by step (3.1), obtains training sample and concentrates the M dimension local limit coding LLC feature of all samples.

Step 4, with the negative sample collection X of negative image^-The local limit coding potential Di Li Cray of LLC feature learning divide Join LDA topic model.

(4.1) according to the negative sample collection X of negative image^-Local limit coding LLC feature construction " document-word " matrix:

With the negative sample collection X of negative image^-Making a collection of document, each negative sample is as a document；

With local limit coding LLC feature as M the word being likely to occur in collection of document；

With the local limit coding probability that occurs as m-th word of LLC eigenvalue of m dimension, wherein m ∈ [1, M] and For integer；

By said process, the M of each negative sample is tieed up local limit coding LLC feature and is converted into M dimension term vector table Show, and by with the M dimension negative sample that represents of term vector by rows, form " document-word " matrix；

(4.2) " document-word " matrix built in (4.1) is inserted in potential Di Li Cray distribution LDA model, set Theme number T=9, utilizes variation Bayes's VB algorithm to obtain potential Di Li Cray distribution LDA topic model:

$p\left(W\right|\mathrm{\α},\mathrm{\β})=\∫p\left(\mathrm{\θ}\right|\mathrm{\α})\underset{i=1}{\overset{M}{\Π}}\underset{z_i}{\Σ}p(z_i=k|\mathrm{\θ})p\left(w_i\right|z_i=k)p\left({\mathrm{\φ}}_{z_i=k}\right|\mathrm{\β})d\mathrm{\θ}$

Wherein, W represents the term vector of a document, and M represents the word number of a document；

θ represents the theme distribution of a document, and θ obeys the Di Li Cray distribution that parameter is α, α be a T dimension row to Amount；

z_iRepresent the theme of i-th word, z_iObeying parameter is the multinomial distribution of θ, and k represents kth theme；

w_iRepresent the i-th word in document, w_iObedience parameter is z_iMultinomial distribution；

φ_kRepresent the word distribution of kth theme, φ_kObey the Di Li Cray distribution that parameter is β, β be M dimension row to Amount.

Step 5, the potential Di Li Cray acquired by step 4 distributes LDA topic model from positive image pattern collection X⁺In choose Select initial positive sample set

(5.1) according to the potential Di Li Cray distribution LDA model of step 4 study, all samples in training sample are calculated Likelihood probability；

(5.2) by the negative sample collection X of negative image^-In all negative samples likelihood probability by from big to small order sequence, The likelihood probability of the negative sample taking 0.8 × Q selects thresholding Tr as select initial positive sample set, and wherein Q is negative sample collection X^-The number of middle negative sample；

(5.3) by the sample set X of positive image⁺The likelihood probability of middle sample is selected less than the sample selecting thresholding Tr, composition New sample set, is the initial positive sample set selected

Step 6, with the negative sample collection X of negative image^-With initial positive sample setThe two class SVM mirror that repetitive exercise is supervised entirely Other device, obtains optimum descriminator.

(6.1) initial false alarm rate FR is set₀=1, and set the positive sample set of trainingWhereinFor select Initial positive sample set；

(6.2) from negative sample collection X^-In randomly choose and initial positive sample setThe negative sample composition instruction that number of samples is identical The negative sample collection practiced

(6.3) with the positive sample set of trainingNegative sample collection with trainingOne is trained to encode based on local limit Two class SVM descriminators of the full supervision of LLC feature, obtain the classification model W of descriminator₀With deviation b₀；

(6.4) descriminator obtained according to (6.3), the false alarm rate of calculating descriminator:

Wherein, N_fIt it is the negative sample collection of trainingMiddle misjudgement is the erroneous judgement set of positive sampleIn number of samples, It it is the negative sample collection of trainingIn q-th negative sample,It it is negative sampleCorresponding Local limit coding LLC feature, T represents and seeks transposition, N_TIt it is the negative sample collection of trainingThe number of middle negative sample；

(6.5) the descriminator false alarm rate FR that (6.4) are obtained and the false alarm rate FR of setting₀Compare；

(6.5a) for FR ＜ FR₀Situation, the false alarm rate FR of the most more new settings₀=FR, and the positive sample set to trainingUpdate as follows, return step (A32):

First, the descriminator obtained according to training calculates initial positive sample setIn the score of each sample:

$Score\left(x_p^{+}\right)={W_0}^T\·f_p^{+}+b_0$

Wherein,It it is initial positive sample setIn pth sample,It it is sampleCorresponding M ties up local limit Coding LLC feature, W₀And b₀It is classification model and the deviation factor of descriminator respectively,It it is sampleScore；

Then, according to initial positive sample setIn the score of all samples calculate the thresholding T of more new samples_s:

$T_s=\frac{\underset{p=1}{\overset{R}{\Σ}}{y}_p^{+}\·Score\left(x_p^{+}\right)}{\underset{p=1}{\overset{R}{\Σ}}{y}_p^{+}},$

Wherein, R is initial positive sample setIn number of samples,It it is sampleSorted labelling, if sampleScoreMore than or equal to 0, thenOtherwise,

Finally, by initial positive sample setMiddle sample score is more than or equal to thresholding T_sThe sample new sample set of composition, It is the positive sample set of the training after renewal；

(6.5b) for FR >=FR₀Situation, then the grader (A33) obtained as optimum descriminator, i.e. optimum mirror Other device template W^*=W₀, deviation factor b^*=b₀。

Two, test phase

Step 7, each test sample concentrated test sample according to the method for step 1 extracts intensive scale invariant feature Conversion SIFT feature.

Step 8, concentrates the intensive Scale invariant features transform SIFT feature of each test sample according to step test sample The coded method of rapid 3 carries out local limit coding, obtains the local limit coding LLC feature of test sample.

Step 9, the optimum classifier obtained with the training stage encodes LLC feature according to the local limit of test sample and enters Row taxonomic history, obtains the identification result of test sample collection.

(9.1) test sample collection X is calculated with the optimum descriminator obtained_testMiddle test sample x_testScore:

Score(x_test)=(W^*)^T·f_test+b^*

Wherein, x_testIt is test sample collection X_testIn a test sample, f_testIt it is test sample x_testCorresponding M dimension Local limit coding LLC feature, W^*And b^*It is classification model and deviation factor, the Score (x of optimum descriminator respectively_test) be Test sample x_testScore；

(9.2) according to test sample x_testScore test sample is carried out taxonomic history: if test sample x_test? Divide Score (x_test) more than or equal to 0, then this test sample being differentiated is positive sample, otherwise, this test sample is differentiated for negative sample This；

(9.3) by step (9.1)-(9.2) to test sample collection X_testIn each test sample carry out taxonomic history, To identification result.

The effect of the present invention can be further illustrated by following experiment:

1. experiment condition

Experiment operation platform: MATLAB R2012a, Intel (R) Core (TM) i5-4590CPU@3.30GHZ, Windows7 Ultimate.

Experiment data used are U.S.'s Sandia MiniSAR public data collection, and it is little for needing the target differentiated in data set Car.Sample containing dolly is positive sample, and the sample without dolly is negative sample.The training sample set used in experiment includes bearing The negative sample collection of image and the sample set of positive image.Wherein:

The negative sample collection of negative image has 333 sizes to be the negative sample of 100 × 100, and Fig. 2 gives the negative of 12 negative images Sample instantiation；

The sample set of positive image has 469 sizes to be the sample of 100 × 100, wherein have 156 positive samples and 313 bear Sample, Fig. 3 gives the sample instantiation of 12 positive images；

It is the test sample of 100 × 100 that test sample is concentrated with 426 sizes, wherein has 153 positive samples and 273 Negative sample, Fig. 4 gives 12 test sample collection examples.

In experiment, relaxation factor C=1 of linear classifier SVM.

2. experiment content:

Experiment 1, constantly chooses from the sample set of the known positive image including positive sample and negative sample by the inventive method Select positive sample set, be used for training optimum descriminator, then test sample collection is carried out taxonomic history with obtaining optimum descriminator.

Experiment 2, aligns the positive sample set in image and the negative sample of negative image with traditional full supervision two class SVM descriminators Collection is trained, and obtains descriminator, then test sample collection is carried out taxonomic history.This experiment needs to align the sample of image Collection carries out the labelling of positive negative sample.

Experiment 3, is trained the negative sample collection of negative image with a traditional linear class SVDD descriminator, obtains a class mirror Other device, then carries out taxonomic history to test sample collection.One class SVDD descriminator is with reference to calendar year 2001 Delft University of The thesis for the doctorate one-class classification of the D.Tax of Technology.

Experiment 1, experiment 2 and the identification result such as table 1 of experiment 3:

The identification result contrast of table 1 the inventive method and control methods

3. interpretation

As can be seen from Table 1, for the SAR image data used by experiment, the present invention applies based on Weakly supervised study SAR image target discrimination method achieves the discriminating of target and clutter, shows that the discrimination method of the present invention has good performance.

By table 1 it can also be seen that the discriminating performance of two class SVM descriminators of full supervision is slightly better than the inventive method, but Full supervised classification needs to provide enough tape label training samples, and for mass data, sample is carried out handmarking and both consumed Time effort again, had a strong impact on the process of system real time；And the present invention is compared with the identification result of a class SVDD descriminator, bright Show to have and preferably differentiate performance.Although a class SVDD descriminator of clutter training only make use of clutter data, decrease artificial The cost of marker samples, but due to the complicated variety of clutter data, the model that training obtains can not well adapt to complexity Target and the taxonomic history of clutter under scene.

To sum up, the present invention utilizes Weakly supervised information, constantly selects positive sample set, the two class SVM mirror that repetitive exercise is supervised entirely Other device, obtains optimum descriminator and realizes, to target and the taxonomic history of clutter, comparing two class SVM descriminators of full supervision, at mirror Decrease the cost of handmarking while other similar nature, in actual applications, have more practicality, and compare and instruct with clutter The class SVDD descriminator practiced, has in the discriminating of complex scene target and clutter and preferably differentiates performance, have good Application prospect.

Claims (8)

1. SAR image target discrimination method based on Weakly supervised study, including:

A, training step

(A1) image pattern collection X is aligned⁺Negative sample collection X with negative image^-In each training sample extract M tie up local limit Coding LLC feature, M=1024, wherein, positive image pattern collection X⁺In there is no handmarking, both comprised positive sample and also comprised negative sample This:

(A11) each training sample is extracted intensive Scale invariant features transform SIFT feature, and close to all training samples Collection Scale invariant features transform SIFT feature Kmeans is polymerized to M class；

(A12) with M cluster centre structure code book CB, i.e. M cluster centre is by arranging the matrix rearranged as code book CB；

(A13) with the code book CB of structure, the intensive Scale invariant features transform SIFT feature of each training sample is carried out local to limit Property processed encodes, and obtains the local limit coding LLC feature after training sample coding；

(A2) with the negative sample collection X of negative image^-Train potential Di Li Cray distribution LDA topic model, and with this potential Di Li Cray distribution LDA topic model is from positive image pattern collection X⁺In select initial positive sample set

(A3) with negative sample collection X^-With initial positive sample setThe two class SVM descriminators that repetitive exercise is supervised entirely, obtain optimum Descriminator；

B, testing procedure

(B1) each test sample extraction M dimension local limit coding LLC feature test sample concentrated:

(B11) each test sample is extracted intensive Scale invariant features transform SIFT feature；

(B12) with the code book CB of the structure intensive Scale invariant features transform to each test sample in training step (A12) SIFT feature extracts local limit coding, obtains the local limit coding LLC feature of each test sample；

(B2) test sample is entered by the optimum descriminator obtained with training according to the local limit coding LLC feature of test sample Row taxonomic history, obtains identification result.

Method the most according to claim 1, wherein extracts intensive Scale invariant to each training sample in step (A11) special Levy conversion SIFT feature, carry out as follows:

(A11.1) training sample x is normalized and smoothing processing, the training sample I after being processed；

(A11.2) the training sample I after processing is extracted its gradient amplitude figure G_mWith gradient direction figure G_o；

(A11.3) in training sample x, take the section P that size is 16 × 16, and respectively at gradient amplitude figure G_mWith gradient side To figure G_oIn correspondence position take region that size is 16 × 16 gradient amplitude figure g as section P_mWith gradient direction figure g_o；

(A11.4) by section P with 4 pixels for step-length grid division without overlapping, obtaining 16 sizes is the zonule of 4 × 4, By these 16 zonules by from top to bottom, order from left to right sorts, and obtains the zonule of jth 4 × 4 in section P Center: O_j=(O_jx,O_jy), wherein O_jxAnd O_jyBe respectively this zonule center section P in width and height:

Wherein, int () represents and takes business's computing, and mod () represents complementation, j ∈ [1,16] and be integer；

(A11.5) calculate each pixel gradient amplitude weighted value to the zonule of jth 4 × 4 in section P, obtain this and cut Sheet P is for the amplitude weighting matrix w of the zonule of jth 4 × 4_j, wherein position (l in section P_x,l_y) pixel at place is to jth The gradient amplitude weighted value w computing formula of the zonule of individual 4 × 4 is as follows:

W=(1-| l_x-O_jx|)·I(|l_x-O_jx| ＜ 4) × (1-| l_y-O_jy|)·I(|l_y-O_jy| ＜ 4)

Wherein, (l_x,l_y) be the position coordinates of pixel, i.e. l_xAnd l_yIt is respectively pixel width in section P and height, l_x,l_y ∈ [1,16] and be integer, I is indicator function；

(A11.6) according to amplitude weighting matrix w_j, calculate the section P weighted gradient map of magnitudes to the zonule of jth 4 × 4: g_mj =g_m·w_j, wherein, represent the dot product of corresponding element, g_mRepresent the gradient amplitude figure of section P, w_jRepresent that section P is for jth The amplitude weighting matrix of the zonule of individual 4 × 4；

(A11.7) respectively in weighted gradient map of magnitudes g_mjWith gradient amplitude directional diagram g_oIn with point (O_jx,O_jyCentered by), position takes Size is the region of 4 × 4, as the weighted gradient map of magnitudes g ' of jth 4 × 4 zonule in section P_mjWith gradient direction figure g′_oj；

(A11.8) [0,2 π] being divided into 8 directions, i-th Direction interval isI ∈ [1,8] and be whole Number, calculates the gradient direction figure g ' of the zonule of jth 4 × 4_ojWeighted gradient amplitude corresponding in 8 directions and, obtain 8 The weighted gradient amplitude vector H of dimension_j；

(A11.9) step (A11.5)-(A11.8) is repeated in other zonules of 4 × 4 in section P, obtain the weighting of its 8 dimension Gradient amplitude vector, these 16 weighted gradient amplitude vector composition column vector d=[H₁,...,H_j,...,H₁₆]^T, wherein T represents Transposition, d is 128 corresponding for section P dimension Scale invariant features transform SIFT feature；

(A11.10) training sample x is carried out from left to right with Gridding length L=16, sliding step step=6, slides from top to bottom Moving and obtain the grid that K size is 16 × 16, the grid to each 16 × 16 extracts corresponding 128 by (A11.3) to (A11.9) Dimension Scale invariant features transform SIFT feature, constitutes intensive Scale invariant features transform SIFT feature D=of this sample x [d₁,...,d_n,...,d_K], wherein, d_nIt is the 128 dimension Scale invariant features transform that in sample x, the n-th 16 × 16 grid obtains SIFT feature, n ∈ [1, K] and be integer；

Wherein, K is the meshes number that size is 16 × 16 obtained, W and H is respectively width and the height of sample x, and floor (a) represents Take the maximum integer of no more than a.

Method the most according to claim 2, is wherein normalized and smooths place to training sample x in step (A11.1) Reason, is carried out as follows:

First, training sample x is carried out two norm normalizeds, obtains the training sample after normalization: Wherein norm represents and seeks two norms；

Then, be sized be 5 × 5 Gaussian template M:

By the training sample x after Gaussian template M and normalization_NConvolution, the training sample after being processed: I=x_N* M, its In, * represents and seeks convolution.

Method the most according to claim 2, wherein obtains the gradient amplitude figure G of training sample I in step (A11.2)_mAnd ladder Degree directional diagram G_o, carry out as follows:

First, by one-dimensional horizontal gradient template A=[-0.5,0,0.5] and vertical gradient template A^TTo training sample I volume Long-pending, obtain the horizontal gradient figure G of this training sample_x=I*A and vertical gradient map G_y=I* Α^T, wherein, T represents and seeks transposition, * table Show and seek convolution；

Then, with horizontal gradient figure G_xWith vertical gradient map G_yConstitute the gradient amplitude figure G of this training sample_mWith gradient direction figure G_o:

Wherein, arctan is arctan function.

Method the most according to claim 1, wherein obtains the local limit coding of each training sample in step (A13) LLC feature, is carried out as follows:

(A13.1) the intensive Scale invariant features transform SIFT feature D=[d to training sample x₁,...,d_n,...,d_KIn] The n-th 128 dimension Scale invariant features transform SIFT feature d_n, calculate d_nWith the Euclidean distance of codebook element each in code book CB, look for To 5 codebook element that Euclidean distance is minimum, and obtain local loop LB by row arrangement_n；

(A13.2) respectively by Scale invariant features transform SIFT feature d_nCorresponding local loop LB_nIn Europe of 5 elements Formula distance, divided by itself and the Euclidean distance sum of these 5 elements, obtains local loop LB_nIn code coefficient corresponding to 5 elements

(A13.3) by code coefficient composition column vector corresponding for codebook element all in code book CB, feature d is obtained_nCode coefficient c_n, wherein, except local loop LB in code book CB_nIn codebook element outside, the code coefficient that other codebook element is corresponding is 0；

(A13.4) the intensive Scale invariant features transform SIFT feature D=[d to training sample x₁,...,d_n,...,d_KK in] Individual 128 dimension Scale invariant features transform SIFT feature are encoded by step (A13.1)-(A13.3), obtain the coding system of correspondence Matrix number c=[c₁,...,c_n,...,c_K], each row element in code coefficient matrix c is sought its maximum, obtains training sample M dimension local limit coding LLC feature C of this x.

Method the most according to claim 1, wherein the negative sample collection X of step (A2) middle negative image^-Learn potential Di Li Cray distribution LDA topic model, is carried out as follows:

(A21) according to the negative sample collection X of negative image^-Local limit coding LLC feature construction " document-word " matrix:

Making a collection of document with the negative sample collection X-of negative image, each negative sample is as a document；

With local limit coding LLC feature as M the word being likely to occur in collection of document；

With the local limit coding probability that occurs as m-th word of LLC eigenvalue of m dimension, wherein m ∈ [1, M] and be whole Number；

By said process, the M dimension local limit coding LLC feature of each negative sample is converted into M dimension term vector and represents, and By with the M dimension negative sample that represents of term vector by rows, form " document-word " matrix；

(A22) " document-word " matrix built in (A21) is inserted in potential Di Li Cray distribution LDA model, set theme Number T=9, utilizes variation Bayes's VB algorithm to obtain potential Di Li Cray distribution LDA topic model:

Wherein, W represents the term vector of a document, and M represents the word number of a document；

θ represents the theme distribution of a document, and θ obeys the Di Li Cray distribution that parameter is α, and α is the row vector of a T dimension；

z_iRepresent the theme of i-th word, z_iObeying parameter is the multinomial distribution of θ, and k represents kth theme, k ∈ [1, T] and be Integer；

w_iRepresent the i-th word in document, w_iObedience parameter is z_iMultinomial distribution；

φ_kRepresent the word distribution of kth theme, φ_kObeying the Di Li Cray distribution that parameter is β, β is the row vector of a M dimension.

Method the most according to claim 1, wherein obtains optimum descriminator, carries out as follows in step (A3):

(A31) initial false alarm rate FR is set₀=1, if the positive sample set of trainingWhereinFor select initial the most just Sample set；

(A32) from the negative sample collection X of negative image^-In the positive sample set that randomly chooses and trainThe negative sample that number of samples is identical The negative sample collection of composition training

(A33) with the positive sample set of trainingNegative sample collection X with training₀ ^-Train one based on local limit coding LLC spy Two class SVM descriminators of the full supervision levied, obtain the classification model W of descriminator₀With deviation factor b₀；

(A34) descriminator obtained according to (A33), the false alarm rate of calculating descriminator:

Wherein, N_fIt it is the negative sample collection of trainingMiddle misjudgement is the erroneous judgement set of positive sampleIn number of samples, It it is the negative sample collection of trainingIn q-th negative sample,It it is negative sampleCorresponding Local limit coding LLC feature, T represents and seeks transposition, N_TIt it is the negative sample collection of trainingThe number of middle negative sample；

(A35) the descriminator false alarm rate FR that (A34) is obtained and the false alarm rate FR of setting₀Compare:

If FR is ＜ FR₀, the false alarm rate FR of the most more new settings₀=FR, and update the positive sample set of trainingReturn step (A32)；

If FR >=FR₀, then the descriminator that (A33) obtains is optimum descriminator, optimum descriminator template W^*=W₀, deviation system Number b^*=b₀。

Method the most according to claim 7, wherein updates the positive sample set of training, enters as follows in step (A35) OK:

(A35.1) descriminator obtained according to training calculates initial positive sample setIn the score of each sample, computing formula As follows:

Wherein,It it is initial positive sample setIn pth sample,It it is sampleCorresponding M dimension local limit coding LLC feature, W₀And b₀It is classification model and the deviation factor of descriminator respectively,It it is sampleScore；

(A35.2) according to initial positive sample setIn the score of all samples calculate the thresholding T of more new samples_s:

Wherein, R is initial positive sample setIn number of samples,It it is sampleSorted labelling, if sample's ScoreMore than or equal to 0, thenOtherwise,

(A35.3) by initial positive sample setMiddle sample score is more than or equal to thresholding T_sThe sample new sample set of composition, i.e. For the positive sample set of training after updating.