CN103984746A - Semi-supervised classification and regional distance measurement based SAR (Synthetic Aperture Radar) image identification method - Google Patents

Abstract

The invention discloses a semi-supervised classification and regional distance measurement based SAR (Synthetic Aperture Radar) image identification method. The implementation steps of the semi-supervised classification and regional distance measurement based SAR image identification method comprise establishing an image library through original SAR image segmentation and selecting SAR image blocks which are single in target from the image library; extracting feature vectors of the image blocks in the image library; dividing the selected SAR images into a plurality of classes, utilizing the corresponding feature vectors to serve as a training sample, training a semi-supervised classifier and performing classification on the image library through the semi-supervised classifier; obtaining the classes of the query image blocks through the trained classifier, wherein the query image blocks are input by users; calculating a class set of the query image blocks according to a confusion matrix, calculating regional similarity distances between the query image blocks and image blocks and returning the user required number of image blocks according to the order that the distances are from small to large, wherein the image blocks belong to the class set in the image library. According to the semi-supervised classification and regional distance measurement based SAR image identification method, the classification error can be corrected, the information identification accuracy is high, and interpretation can be performed on a plurality of SAR images simultaneously.

Description

The SAR image-recognizing method of estimating based on semi-supervised classification and region distance

Technical field

The invention belongs to technical field of image processing, relate to a kind of recognition methods of SAR image, can be applicable to several SAR images to carry out decipher simultaneously.

Background technology

SAR image is because have round-the-clock, a round-the-clock detectivity, and optical imagery is to the complete ind feature of weather conditions especially relatively, and the expansion progressively of its application, comprises military affairs, agricultural, navigation, geographical monitor etc.Cutting apart of SAR image, denoising, changing detection etc. is all study hotspot field, and an important foundation of these research fields is exactly SAR image recognition.Some traditional recognition technologies are mainly for the problem of accuracy of identification, and are mostly applied to the identification problem of region among a small circle of individual SAR image.But these technology have not obviously met the applied environment that SAR amount of images magnanimity increases instantly.In order to overcome above-mentioned technical disadvantages, the present invention estimates and under the framework of CBIR technology, has proposed a kind of SAR image recognition technology in conjunction with graph theory semi-supervised learning method and improved region distance.This technology realizes simple, not only ensures accuracy of identification but also can meet the application scenarios that a large amount of SAR images are identified simultaneously.

In practical problems, what easily obtain has label data often than without label data much less.In order to solve problems, semi-supervised learning method is arisen at the historic moment.Semi-supervised learning method is the One class learning method between supervised learning and unsupervised learning, and the method has been utilized exemplar simultaneously and without exemplar, utilized whole geometry structure to complete class mark and propagate.In semi-supervised learning field, the semi-supervised learning method based on graph theory was most active research direction in recent years.The most famous and the widely used semi-supervised learning method based on graph theory comprises: minimum blanking method, gaussian random field method etc., respectively referring to A.Blum and S.Chawla.Learning from labeled and unlabeled data using graph mincuts.In Prnceedings of the18th International Conference on Machine Learning.Morgan Kaufmann, San Francisco, CA, 2001.19-26; Zhu, Xiaojin, Zoubin Ghahramani, and John Lafferty. " Semi-supervised learning using gaussian fields and harmonic functions. " ICML.Vol.3.2003.These methods using have label and without label data the summit as figure, the similarity between data is as limit and the weights of connect Vertex, and utilizes the geometrical property of figure to complete class mark from having the data of label to the propagation of the data without label, thereby reaches the object of classification.Because gaussian random field method computation complexity is low, consider the ageing of recognition system, the present invention adopts the method.

CBIR CBIR is that visual signature based on image low level completes in image data base retrieval and query image consistent or similar image collection process in terms of content.This technology comprises a series of image processing method, comprises feature extraction, similarity measurement, user feedback etc.So far, existing a lot of ripe, famous searching system is suggested, as SIMPLIcity searching system, referring to James Z.Wang, Jia Li, Gio Wiederhold.SIMPLIcity:Semantics-Sensitive Integrated Matching for Picture Llbraries.IEEE Trans.on Pattern Analysis and Machine Intelligence, 2001, 23 (9): 947-963, this SIMPLIcity searching system is successfully applied to the natural image search problem of magnanimity, but due to technical limitation and SAR image own characteristic, it is directly applied in SAR image recognition to effect also desirable.The SAR image indexing system of the combination gauss hybrid models classification and for example proposing for 2009, it is GMM searching system, referring to Hou, B., Tang, X., Jiao, L., & Wang, S. (2009, October) .SAR image retrieval based on Gaussian Mixture Model classification.In Synthetic Aperture Radar, 2009.APSAR2009.2nd Asian-Pacific Conference on (pp.796-799) .IEEE, the method is towards SAR image, in retrieving, effectively use textural characteristics, but owing to utilizing supervised classification method to make its popularization ability in realistic problem lower, simultaneously because the similarity matching technique of the method is not considered the feature of SAR image, make retrieval effectiveness unsatisfactory.Although provided outstanding experimental result in this article, but these results depend on the original SAR image of overlapping cutting and set up picture library, the image block that this strategy obtains has the cluster characteristic of height, be that sample separation in same class is very little, inhomogeneous sample separation is very large, such data distribution distributes with the data in practical application, and often difference is large, and the experimental result obtaining can not be verified the validity of its method fully.

Summary of the invention

The object of the invention is the defect existing for above-mentioned prior art, according to the special imaging characteristics of SAR image, under the framework of traditional CBIR, a kind of SAR image-recognizing method of estimating based on semi-supervised classification and improvement region distance is proposed, to reduce the difference that in experimental data and practical application, data distribute, improve the information accuracy of identification in practical application.

The technical scheme that realizes the object of the invention is: use discrete wavelet, resolve the texture information in SAR image, use texture information, utilize Gaussian random field semi-supervised learning method to complete the classification work of SAR image library, and adopt improved regional complex feature similarity matching algorithm to complete the similarity coupling of SAR image.Its specific implementation step comprises as follows:

1) original SAR image is carried out to zero lap cutting, to set up SAR image library { p ₁, p ₂..., p _n, from this image library, select image block { p according to the single principle of target ₁, p ₂..., p _l, wherein l < < N, N represents the SAR image block number in picture library, and l represents the SAR image block number of picking out, and the single principle of described target refers to that in image block, certain target accounts for the over half of total image area;

2) extract the sub belt energy of three layers of conversion of discrete wavelet of all image blocks, as the proper vector of image block $\{f_1,f_2,\&CenterDot;\&CenterDot;\&CenterDot;,f_{\stackrel{\&OverBar;}{n}}\},$ Wherein, $\stackrel{\&OverBar;}{n}=10;$

3) by the SAR image block { p picking out ₁, p ₂..., p _lbe divided into { c according to semantic content _i, 1≤i≤k} class, wherein k represents the number of semantic classes, and with characteristic of correspondence vector as training sample, training Gaussian random field semi-supervised classifier, utilizes this sorter to whole SAR image library { p ₁, p ₂..., p _nclassification, obtain having class target SAR image library;

4) the query image piece p ' to user's input, adopting and step 2) identical method extracts its proper vector f ', and use and step 3) identical training sample and the Gaussian random field semi-supervised classifier training, the classification that obtains query image piece is counted c _i;

5) according to step 4) classification that obtains counts c _iand experience confusion matrix, calculate query image piece classification set c}:

5a) in classified SAR image library, pick out the front K width image block of each class, form new image pattern collection, from then on random choose training sample training Gaussian random field semi-supervised classifier in sample set, carry out 100 random assortments with this sorter and test, obtain experience confusion matrix Con ∈ R ^{k × k}, confusion matrix Con is square formation, wherein i capable j row Con (i, j) represent to belong to c _ithe sample of class is divided into c _jthe number of class, 1≤i≤k, 1≤j≤k;

5b) experience confusion matrix is carried out to row normalization, the each element in being listed as by, divided by the summation of this column element, obtains the posterior probability Matrix C onP ∈ R of experience ^{k × k};

5c) threshold value T is set, by the capable j row of the i ConP (i in posterior probability matrix, j) compare with threshold value T, in the time of ConP (i, j)≤T, by ConP (i, j) be set to 0, otherwise ConP (i, j) remains unchanged, the size of threshold value T is set according to the nonzero element number of each row of expecting;

5d) count c according to the classification of query image piece _i, the position of searching nonzero element in the i row in posterior probability Matrix C onP, finally obtains classification set { c};

6) calculate in query image piece p ' and picture library and belong to the classification set { region distance of all image blocks in c};

7) according to step 6) region distance that obtains, return to order from small to large the image that user needs quantity, complete image recognition.

The present invention has the following advantages compared with prior art:

1, the present invention, owing to creating SAR image library with non-overlapping cutting method, reduces the gap of test sample book and real data, has increased the confidence level of recognition result;

2, the present invention, owing to adopting Gaussian random field semi-supervised learning method to classify to picture library, has reduced the workload of hand picking training sample, has reduced the impact of artificial subjective factor on classification results;

3, the present invention, owing to adopting the strategy of correcting classification error, has effectively reduced the impact of classification error on similarity coupling;

4, the present invention is directed to SAR image singularity, improved the similarity measure of regional complex feature, make SAR image similarity matching result more accurate, improved accuracy of identification.

Brief description of the drawings

Fig. 1 is realization flow schematic diagram of the present invention;

Fig. 2 is for setting up the original SAR image of SAR image library in the present invention;

Fig. 3 is the SAR image block sample figure that the present invention picks out in SAR picture library;

Fig. 4 is the present invention and GMM searching system overall performance comparison diagram;

Fig. 5 is the present invention and the performance comparison diagram of GMM searching system in each semantic classes.

Embodiment

With reference to Fig. 1, specific implementation step of the present invention is as follows:

Step 1, sets up SAR image library { p ₁, p ₂..., p _n, and select SAR image block according to the single principle of target.

Being implemented as follows of this step:

1a) select pixel size to be followed successively by 19035 × 7330,7082 × 7327 2 width large scale SAR images, as the original SAR image of setting up picture library, respectively as Fig. 2 (a), shown in Fig. 2 (b);

1b) 2 selected original SAR images are carried out to non-overlapping cutting, after cutting, obtain size and be 256 × 256 2828 width SAR image blocks, set up SAR image library { p with this ₁, p ₂..., p _n, wherein N=2828;

1c) in image library, select SAR image block { p according to the single principle of target ₁, p ₂..., p _l, wherein l < < N, the SAR image block number in N presentation video storehouse, l represents the SAR image block number of picking out, the single principle of described target refers to that in image block, certain target accounts for the over half of total image area.The present invention, in the time selecting, has selected 28 width SAR image blocks, i.e. l=28 altogether.

Step 2, carries out feature extraction to all image blocks in picture library.

Select the sub belt energy ξ of three layers of conversion of discrete wavelet as the proper vector of image block wherein, the dimension of representation feature vector, this example is selected but be not limited to 10.To a certain subband, energy definition is:

$\mathrm{\&xi;}={\left(\underset{n_1=1}{\overset{m_1}{\mathrm{\&Sigma;}}}\underset{n_2=1}{\overset{m_2}{\mathrm{\&Sigma;}}}b{(n_1,n_2)}^2\right)}^{\frac{1}{2}}/(m_1\&times;m_2)---<1>$

Wherein, m ₁× m ₂for subband size, (n ₁, n ₂) represent the index of this sub-band coefficients, b (n ₁, n ₂) represent n in this subband ₁row n ₂the coefficient value of row, the energy of all the other 9 subbands calculates according to <1> formula.

Step 3, to whole SAR image library { p ₁, p ₂..., p _nclassification, obtain having class target SAR image library.

Being implemented as follows of this step:

3a) in the l width SAR image block of having picked out, carry out artificial semantic classification, this example adopts the strategy of area percentage to judge image category, even in a sub-picture piece p, and c _ithe total area size of class object exceedes 50% of this total image area, just specifies that this is c as piece p _iclass, l width SAR image block people is for being divided into mountain region, ocean, city, suburb totally 4 classes the most at last, each class 7 width image block.As shown in Figure 3, wherein Fig. 3 (a) is mountain region to master drawing, and Fig. 3 (b) is ocean, and Fig. 3 (c) is city, Fig. 3 (d) suburb;

3b) according to each class image block, the sub belt energy feature of three layers of conversion of corresponding one group of discrete wavelet, obtain 4 stack features vectors, and using this 4 stack features vector as training sample, utilize Gaussian random field semi-supervised learning sorter to carry out the classification of SAR image library:

The energy feature of the 28 width image blocks of 3b1) picking out by step 1 is as there being exemplar { (x ₁, y ₁) ... (x _l, y _l), with the energy feature that remains 2800 width SAR image blocks in SAR image library as without exemplar { x _l+1... x _l+u; with having exemplar and without exemplar foundation figure G=(V, E), wherein l < < u; u indicates the number without exemplar; l+u=N, N represents the number of total sample, the summit of V presentation graphs G; the limit of E presentation graphs G; and l=28, u=2800, N=2828;

3b2) try to achieve the similar matrix W of figure G according to formula <2>,

$W={\left({\widetilde{\mathrm{\&omega;}}}_{\tilde{i},\tilde{j}}\right)}_{N\&times;N},{\widetilde{\mathrm{\&omega;}}}_{\tilde{i},\tilde{j}}=\exp (-\underset{\tilde{d}=1}{\overset{\stackrel{\&OverBar;}{n}}{\mathrm{\&Sigma;}}}\frac{{(x_{\tilde{i}\tilde{d}}-x_{\tilde{j}\tilde{d}})}^2}{{\mathrm{\&sigma;}}_{\tilde{d}}^2})---<2>$

Wherein, represent similar matrix W the row the element of row, the dimension of presentation video block eigenvector, represent sample dimension component, represent sample dimension component, the super parameter of dimension, $1\&le;\tilde{i}\&le;N,1\&le;\tilde{j}\&le;N,1\&le;\tilde{d}\&le;\stackrel{\&OverBar;}{n},\stackrel{\&OverBar;}{n}=10;$

3b3) according to the energy function of formula <3> structural map G:

$E\left(f^s\right)=\&infin;\underset{\tilde{i}}{\mathrm{\&Sigma;}}{(f^s\left(\tilde{i}\right)-y_{\tilde{i}})}^2+\frac{1}{2}\underset{\tilde{i},\tilde{j}}{\mathrm{\&Sigma;}}{\widetilde{\mathrm{\&omega;}}}_{\tilde{i},\tilde{j}}{(f^s\left(\tilde{i}\right)-f^s\left(\tilde{j}\right))}^2---<3>$

Wherein, E (f ^s) energy function of presentation graphs G, indicate the real function of exemplar on figure G, indicate the real function on figure G without exemplar, represent functional value, the similar matrix W of presentation graphs G row the element of row, $1\&le;\tilde{i}\&le;N,1\&le;\tilde{j}\&le;N;$

3b4) solve energy function E (f ^s) optimum solution argminE (f ^s), obtain classification results, the final classification results that is 2828 width image blocks in image library is mountain region 1143 width successively, ocean 748 width, city 344 width, suburb 593 width, amount to 2828 width SAR image blocks, the target of this step is that SAR image library is become and has class target SAR image library, to reduce the workload of similarity coupling, improves recognition speed.

Step 4, classifies to the query image piece p ' of user's input.

Being implemented as follows of this step:

4a) adopt the method identical with step 2 to extract the proper vector f ' of query image piece p ';

4b) use the training sample identical with step 3 and Gaussian random field semi-supervised classifier, the classification that obtains query image piece p ' is counted c _i.

Step 5, in order to reduce the impact of classification error on recognition result, the classification obtaining according to step 4 is counted c _iand experience confusion matrix, the classification set { c} of calculating query image piece p '.

Being implemented as follows of this step:

5a) in classified SAR image library, pick out the front 100 width image blocks of each class, form new image pattern collection, from then on random choose training sample training Gaussian random field semi-supervised classifier in sample set, carrying out 100 random assortments with this sorter tests, wherein the ratio of training sample and test sample book is 1:99, thereby obtains experience confusion matrix Con ∈ R ^{k × k}, confusion matrix Con is square formation, wherein i capable j row Con (i, j) represent to belong to c _ithe sample of class is divided into c _jthe number of class, 1≤i≤k, 1≤j≤k, k represents semantic classes number, and R represents real number field, and table 1 has provided experience confusion matrix Con ∈ R ^{k × k}sample, wherein semantic classes number k=4, is respectively mountain region, ocean, city, suburb, each numeric representation belongs to c _ithe sample of class is divided into c _jthe number of class, for example Con (1,1) represents that the number that the sample that belongs to mountain region is divided into mountain region is 41;

100 random assortments of table 1 are tested the confusion matrix obtaining

Con∈R k×k	Mountain region	Ocean	City	Suburb
					Mountain region	41	20	0	38
Ocean	6	93	0	0
					City	0	0	97	2
Suburb	26	0	13	60

5b) to experience confusion matrix Con ∈ R ^{k × k}carry out row normalization, be about to a wherein each element being listed as and, divided by the summation of this column element, obtain the posterior probability Matrix C onP ∈ R of experience ^{k × k}.Table 2 has provided posterior probability Matrix C onP ∈ R ^{k × k}sample, wherein semantic classes number k=4, is respectively mountain region, ocean, city, suburb, each numerical value is the experience confusion matrix Con ∈ R that his-and-hers watches 1 provide ^{k × k}carry out that row normalization obtains;

The posterior probability matrix of table 2 experience

ConP∈R k×k	Mountain region	Ocean	City	Suburb
					Mountain region	0.5616	0.1770	0	0.3800
Ocean	0.0822	0.8230	0	0
					City	0	0	0.8818	0.0200
Suburb	0.3562	0	0.1182	0.6000

5c) threshold value T is set, by posterior probability Matrix C onP ∈ R ^{k × k}in the capable j row of i ConP (i, j) compare with threshold value T, in the time of ConP (i, j)≤T, by ConP (i, j) be set to 0, otherwise ConP (i, j) remains unchanged, the size of threshold value T is set according to the nonzero element number of each row of expecting, this example is established T=0.1, and table 3 has provided the posterior probability Matrix C onP ∈ R after threshold process ^{k × k}sample, wherein semantic classes number k=4, is respectively mountain region, ocean, city, suburb, each numerical value is the posterior probability Matrix C onP ∈ R that his-and-hers watches 2 provide ^{k × k}carry out that threshold process obtains;

Posterior probability matrix after table 3 threshold process

ConP∈R k×k	Mountain region	Ocean	City	Suburb
					Mountain region	0.5616	0.1770	0	0.3800
Ocean	0	0.8230	0	0
					City	0	0	0.8818	0
Suburb	0.3562	0	0.1182	0.6000

5d) count c according to the classification of query image piece _i, the position of searching nonzero element in the i row in posterior probability Matrix C onP, finally obtains classification set { c}.

Step 6, calculates in query image piece p ' and picture library and belongs to the classification set { region distance of all image blocks in c}.

Being implemented as follows of this step:

6a) for image block p and query image piece p ', respectively taking 4 × 4 area size as unit, calculate the high-frequency sub-band energy of discrete wavelet one deck conversion and gray feature and use as cutting apart feature, utilize adaptive k-means algorithm to carry out cluster to cutting apart feature, obtain the texture region collection R1={r of image block p ₁, r ₂..., r _h... r _mand the texture region collection R of query image piece p ' ₂=r ' ₁, r ' ₂..., r ' _o... r ' _n, r _h, r ' _oeach region after presentation video piece p and query image piece p ' utilize textural characteristics to cut apart respectively, wherein 1≤h≤m, 1≤o≤n, the number of the texture region of m presentation video piece p, n represents the number of the texture region of query image piece p ';

In order to make the adaptive work of k-means algorithm, this example adopts following two kinds of strategies to complete adaptive k-means to cut apart:

6a1) to divergence D (num) setting threshold T _num=0.035, wherein D (num) is defined as follows:

$D\left(\mathrm{num}\right)=\underset{\stackrel{\&OverBar;}{i}=1}{\overset{L}{\mathrm{\&Sigma;}}}\underset{1\&le;\stackrel{\&OverBar;}{j}\&le;\mathrm{num}}{\min }{(X_{\stackrel{\&OverBar;}{i}}-{\hat{X}}_{\stackrel{\&OverBar;}{j}})}^2,1\&le;\stackrel{\&OverBar;}{i}\&le;L,1\&le;\stackrel{\&OverBar;}{j}\&le;\mathrm{num}---<4>$

In formula, num represents cluster number, and L represents to cut apart Characteristic Number, represent a certain proper vector of cutting apart, represent the cluster centre of a certain class, the value of num increases progressively since 2, as D (num) < T _numtime, num stops increasing progressively, and cluster number num determines, on the contrary the value of num increases by 1;

6a2) cluster number num is arranged to upper limit num _max=5, as num > num _maxtime, num stops increasing progressively, and cluster number num determines, on the contrary the value of num increases by 1;

6b) according to the distance d (r between formula <5> computed image piece p and query image piece p ' texture region _h, r ' _o),

Wherein, for texture region r in image block p _haveraged feature vector, for the middle texture region r ' of query image piece p ' _oaveraged feature vector, for the weight coefficient of each vector, wherein and

6c) for image block p and query image piece p ', utilize Prewitt operator and binary segmentation method to obtain the fringe region collection RE of image block p ₁={ re _vand the fringe region collection RE of query image piece p ' ₂=re ' _z, re _v, re ' _zeach region after presentation video piece p and query image piece p ' utilize edge feature to cut apart respectively, wherein 1≤v≤2,1≤z≤2;

6c1) set Prewitt operator:

Wherein λ=4, g ^yrepresent the edge detection operator of vertical direction, g ^xrepresent respectively the edge detection operator of horizontal direction;

6c2) according to Prewitt operator, the marginal information of computed image piece p vertical direction and the marginal information of horizontal direction respectively:

$G^y=p\&CircleTimes;g^y,G^x=p\&CircleTimes;g^x---<7>$

Wherein, G ^ythe marginal information of presentation video piece p vertical direction, G ^xthe marginal information of presentation video piece p horizontal direction, represent linear convolution;

6c3) according to the marginal information G of image block p vertical direction ^ymarginal information G with horizontal direction ^x, the edge feature f of computed image piece p ^e:

$f^E=\sqrt{{\left(G^y\right)}^2+{\left(G^x\right)}^2}---<8>$

Utilize binary segmentation method edge feature f ^ecut apart and obtain fringe region collection;

6d) according to the distance d (re between formula <9> computed image piece p and query image piece p ' fringe region _v, re ' _z):

$d({\mathrm{re}}_v,{\mathrm{re}}_z^{\&prime;})=\underset{\mathrm{kk}=1}{\overset{2}{\mathrm{\&Sigma;}}}{\mathrm{\&omega;e}}_{\mathrm{kk}}{({\mathrm{fe}}_{\mathrm{kk}}-{\mathrm{fe}}_{\mathrm{kk}}^{\&prime;})}^2---<9>$

Wherein, kk=1,2, in the time of kk=1, fe _kkfor fringe region re in image block p _vcharacteristics of mean, fe ' _kkfor the middle fringe region re ' of query image piece p ' _zcharacteristics of mean; In the time of kk=2, fe _kkfor fringe region re in image block p _vvariance feature, fe ' _kkfor the middle fringe region re ' of query image piece p ' _zvariance feature, ω e _kkfor the weight coefficient of each vector, and $\underset{\mathrm{kk}=1}{\overset{2}{\mathrm{\&Sigma;}}}{\mathrm{\&omega;e}}_{\mathrm{kk}}=1;$

6e) according to the Significance factors s mating between formula <10> computed image piece p and the each texture region of query image piece p ' _h,o,

$\begin{array}{c}\underset{o=1}{\overset{n}{\mathrm{\&Sigma;}}}{s}_{h,o}=P_h,h=1,\&CenterDot;\&CenterDot;\&CenterDot;,m\\ \underset{h=1}{\overset{m}{\mathrm{\&Sigma;}}}{s}_{h,o}=P_o^{\&prime;},o=1,\&CenterDot;\&CenterDot;\&CenterDot;,n\end{array}---<10>$

Wherein, P _hfor texture region r in image block p _haccount for the area percentage of image block, P ' _ofor the middle texture region r ' of query image piece p ' _oaccount for the area percentage of query image piece, the number of the texture region of m presentation video piece p, n represents the number of the texture region of query image piece p ';

6f) according to the Significance factors s mating between formula <11> computed image piece p and the each fringe region of query image piece p ' _v,z,

$\begin{array}{c}\underset{z=1}{\overset{2}{\mathrm{\&Sigma;}}}{s}_{v,z}=P_v,v=1,2\\ \underset{v=1}{\overset{2}{\mathrm{\&Sigma;}}}{s}_{v,z}=P_z^{\&prime;},z=1,2\end{array}---<11>$

Wherein, P _vfor fringe region re in image block p _vaccount for the area percentage of image block, P ' _zfor the middle fringe region re ' of query image piece p ' _zaccount for the area percentage of query image piece;

6g) according to the Significance factors s mating between texture region _h,o, the Significance factors s mating between fringe region _v,z, the distance d (r between texture region _h, r ' _o), the distance d (re between fringe region _v, re ' _z), obtain the region distance d of query image piece p ' and image block p according to formula <12>,

d＝ω ₁×dT(R ₁,R ₂)+ω ₂×dE(RE ₁,RE ₂),ω ₁+ω ₂＝1 <12>

Wherein, presentation video piece p and the distance of query image piece p ' based on texture region, 1≤h≤m, 1≤o≤n; presentation video piece p and the distance of query image piece p ' based on fringe region, 1≤v≤2,1≤z≤2; ω ₁represent dT (R ₁, R ₂) weights, ω ₂represent dE (RE ₁, RE ₂) weights;

6h) by step 6a) to 6g) method, calculate in query image piece p ' and picture library and belong to the classification set { region distance of all image blocks in c}.

Step 7, the region distance obtaining according to step 6, returns to order from small to large the image that user needs quantity, completes image recognition.

Effect of the present invention can further illustrate by following emulation:

1. simulated conditions

This Case Simulation condition is as follows: CORE i5 3.2GHz PC Windows 7 operating systems, Matlab2012 operation platform.

2. emulation content and result

100 width SAR image blocks are chosen at random in emulation 1. from image library, respectively the SAR image block of choosing are identified to emulation by the present invention and GMM searching system, calculate the ensemble average precision ratio of simulation result, and result as shown in Figure 4.Wherein, precision ratio precision is defined as follows:

precision＝n _c/n _s <13>

In formula, n _cthe image block number satisfying condition in the image block that expression system is returned, n _sthe image block number that expression system is returned.The image block quantity that in Fig. 4, transverse axis method for expressing returns, the longitudinal axis represents average precision.

As can be seen from Figure 4, ensemble average accuracy of identification of the present invention is wanted excellent and GMM searching system.

100 width SAR image blocks are chosen at random in emulation 2. from image library, respectively the SAR image block of choosing is identified to emulation by the present invention and GMM searching system, calculate the precision ratio of each semantic classes in simulation result, as shown in Figure 5, wherein require to return to image block quantity is 35 to result.Transverse axis represents each semantic classes, and the longitudinal axis represents precision ratio.

As can be seen from Figure 5, every kind of semantic classes, accuracy of identification of the present invention is wanted excellent and GMM searching system.

Can be illustrated by above experiment, on for SAR problem of image recognition, no matter the present invention is overall precision ratio or the precision ratio of each semantic classes, all be better than existing GMM searching system.

Claims (6)

1. a SAR image-recognizing method of estimating based on semi-supervised classification and region distance, comprises the steps:

1) original SAR image is carried out to zero lap cutting, to set up SAR image library { p ₁, p ₂..., p _n, from this image library, select image block { p according to the single principle of target ₁, p ₂..., p _l, wherein l < < N, N represents the SAR image block number in picture library, and l represents the SAR image block number of picking out, and the single principle of described target refers to that in image block, certain target accounts for the over half of total image area;

2) extract the sub belt energy of three layers of conversion of discrete wavelet of all image blocks, as the proper vector of image block wherein, $\stackrel{\&OverBar;}{n}=10;$

3) by the SAR image block { p picking out ₁, p ₂..., p _lbe divided into { c according to semantic content _i, 1≤i≤k} class, wherein k represents the number of semantic classes, and with characteristic of correspondence vector as training sample, training Gaussian random field semi-supervised classifier, utilizes this sorter to whole SAR image library { p ₁, p ₂..., p _nclassification, obtain having class target SAR image library;

4) the query image piece p ' to user's input, adopting and step 2) identical method extracts its proper vector f ', and use and step 3) identical training sample and the Gaussian random field semi-supervised classifier training, the classification that obtains query image piece is counted c _i;

5) according to step 4) classification that obtains counts c _iand experience confusion matrix, calculate query image piece classification set c}:

5a) in classified SAR image library, pick out the front K width image block of each class, form new image pattern collection, from then on random choose training sample training Gaussian random field semi-supervised classifier in sample set, carry out 100 random assortments with this sorter and test, obtain experience confusion matrix Con ∈ R ^{k × k}, confusion matrix Con is square formation, wherein i capable j row Con (i, j) represent to belong to c _ithe sample of class is divided into c _jthe number of class, 1≤i≤k, 1≤j≤k;

5b) experience confusion matrix is carried out to row normalization, the each element in being listed as by, divided by the summation of this column element, obtains the posterior probability Matrix C onP ∈ R of experience ^{k × k};

5c) threshold value T is set, by the capable j row of the i ConP (i in posterior probability matrix, j) compare with threshold value T, in the time of ConP (i, j)≤T, by ConP (i, j) be set to 0, otherwise ConP (i, j) remains unchanged, the size of threshold value T is set according to the nonzero element number of each row of expecting;

5d) count c according to the classification of query image piece _i, the position of searching nonzero element in the i row in posterior probability Matrix C onP, finally obtains classification set { c};

6) calculate in query image piece p ' and picture library and belong to the classification set { region distance of all image blocks in c};

7) according to step 6) region distance that obtains, return to order from small to large the image that user needs quantity, complete image recognition.

2. SAR image-recognizing method according to claim 1, wherein step 6) belong in described calculating query image piece p ' and picture library classification set the region distance of all image blocks in c}, carry out as follows:

6a) for image block p and query image piece p ', respectively taking 4 × 4 area size as unit, calculate the high-frequency sub-band energy of discrete wavelet one deck conversion and gray feature and use as cutting apart feature, utilize adaptive k-means algorithm to carry out cluster to cutting apart feature, obtain the texture region collection R of image block p ₁={ r ₁, r ₂..., r _h... r _mand the texture region collection R of query image piece p ' ₂=r ' ₁, r ' ₂..., r ' _o... r ' _n, r _h, r ' _oeach region after presentation video piece p and query image piece p ' utilize textural characteristics to cut apart respectively, wherein 1≤h≤m, 1≤o≤n, the number of the texture region of m presentation video piece p, n represents the number of the texture region of query image piece p ';

6b) distance d (the r between computed image piece p and query image piece p ' texture region _h, r ' _o);

6c) for image block p and query image piece p ', utilize Prewitt operator and binary segmentation method to obtain the fringe region collection RE of image block p ₁={ re _vand the fringe region collection RE of query image piece p ' ₂=re ' _z, re _v, re ' _zeach region after presentation video piece p and query image piece p ' utilize edge feature to cut apart respectively, wherein 1≤v≤2,1≤z≤2;

6d) distance d (the re between computed image piece p and query image piece p ' fringe region _v, re ' _z);

Significance factors s 6e) mating between computed image piece p and the each texture region of query image piece p ' _h,o, the Significance factors s mating between image block p and the each fringe region of query image piece p ' _{v, z};

6f) according to the Significance factors s mating between texture region _h,o, the Significance factors s mating between fringe region _v,z, the distance d (r between texture region _h, r ' _o), the distance d (re between fringe region _v, re ' _z), obtain the region distance d of query image piece p ' and image block p:

d＝ω ₁×dT(R ₁,R ₂)+ω ₂×dE(RE ₁,RE ₂),ω ₁+ω ₂＝1，

Wherein, presentation video piece p and the distance of query image piece p ' based on texture region, 1≤h≤m, 1≤o≤n; presentation video piece p and the distance of query image piece p ' based on fringe region, 1≤v≤2,1≤z≤2; ω ₁represent dT (R ₁, R ₂) weights, ω ₂represent dE (RE ₁, RE ₂) weights;

6g) by step 6a) to 6f) method, calculate in query image piece p ' and picture library and belong to the classification set { region distance of all image blocks in c}.

3. SAR image-recognizing method according to claim 2, wherein said step 6b) in distance d (r between computed image piece p and query image piece p ' texture region _h, r ' _o), calculate by following formula:

Wherein, for texture region r in image block p _haveraged feature vector, for the middle texture region r ' of query image piece p ' _oaveraged feature vector, for the weight coefficient of each vector, wherein and

4. SAR image-recognizing method according to claim 2, wherein said step 6d) in distance d (re between computed image piece p and query image piece p ' fringe region _v, re ' _z), calculate by following formula:

$d({\mathrm{re}}_v,{\mathrm{re}}_z^{\&prime;})=\underset{\mathrm{kk}=1}{\overset{2}{\mathrm{\&Sigma;}}}{\mathrm{\&omega;e}}_{\mathrm{kk}}{({\mathrm{fe}}_{\mathrm{kk}}-{\mathrm{fe}}_{\mathrm{kk}}^{\&prime;})}^2,$

Wherein, kk=1,2, in the time of kk=1, fe _kkfor fringe region re in image block p _vcharacteristics of mean, fe ' _kkfor the middle fringe region re ' of query image piece p ' _zcharacteristics of mean; In the time of kk=2, fe _kkfor fringe region re in image block p _vvariance feature, fe ' _kkfor the middle fringe region re ' of query image piece p ' _zvariance feature, ω e _kkfor the weight coefficient of each vector, and

5. SAR image-recognizing method according to claim 2, wherein said step 6e) in the Significance factors s that mates between computed image piece p and the each texture region of query image piece p ' _h,o, calculate by following formula:

$\begin{array}{c}\underset{o=1}{\overset{n}{\mathrm{\&Sigma;}}}{s}_{h,o}=P_h,h=1,\&CenterDot;\&CenterDot;\&CenterDot;,m\\ \underset{h=1}{\overset{m}{\mathrm{\&Sigma;}}}{s}_{h,o}=P_o^{\&prime;},o=1,\&CenterDot;\&CenterDot;\&CenterDot;,n\end{array},$

Wherein, P _hfor texture region r in image block p _haccount for the area percentage of image block, P ' _ofor the middle texture region r ' of query image piece p ' _oaccount for the area percentage of query image piece, the number of the texture region of m presentation video piece p, n represents the number of the texture region of query image piece p '.

6. SAR image-recognizing method according to claim 2, wherein said step 6e) in the Significance factors s that mates between computed image piece p and the each fringe region of query image piece p ' respectively _v,z, calculate by following formula:

$\begin{array}{c}\underset{z=1}{\overset{2}{\mathrm{\&Sigma;}}}{s}_{v,z}=P_v,v=1,2\\ \underset{v=1}{\overset{2}{\mathrm{\&Sigma;}}}{s}_{v,z}=P_z^{\&prime;},z=1,2\end{array},$

Wherein, P _vfor fringe region re in image block p _vaccount for the area percentage of image block, P ' _zfor the middle fringe region re ' of query image piece p ' _zaccount for the area percentage of query image piece.