CN103714148A - SAR image search method based on sparse coding classification - Google Patents

Abstract

The invention provides an SAR image search method based on sparse coding classification. The SAR image search method aims at the defects of an existing image search system and method. Through extraction of characteristics and construction of an over-complete dictionary, solution is carried out through sparse representation based on a dual Memetic algorithm, a sparse representation classifier is trained, the classification process with supervision can be achieved in the classification process, the sparse solution with global optimum can be solved fast, and then search results are output from high to low according to similarity. When the problem of image classification is solved, the method achieves the good improvement effect on classification accuracy, search content similarity, calculating complexity and result robustness.

Description

SAR image search method based on sparse coding classification

Technical field

The invention belongs to SAR image processing field, relate to a kind of SAR image search method based on sparse coding, can to SAR image, classify accurately and realize retrieval, effectively reduced the impact of coherent speckle noise on SAR Images Classification result.

Background technology

Synthetic-aperture radar-SyntheticApertureRadar, is a kind of effective means from earth observation from space, has been applied to widely military surveillance, landforms observation, the aspects such as city planning.Along with the raising of SAR imaging technique in recent years, the quantity of SAR image presents explosive growth, the feature many for SAR image data amount, self-noise is large, how from the SAR image library of magnanimity, to retrieve efficiently and accurately needed image has become problem demanding prompt solution.

Along with the development of infotech, the search method of image has been transformed into content-based retrieval method from initial text based search method, and the method can directly be analyzed image, feature extraction, and similarity analysis is also realized retrieval.At present, main image indexing system comprises: the QBIC system of IBM exploitation, Virage system, the SIMPLIcity system of stanford university research exploitation, the Mires image retrieval prototype system of the still image searching system of the Internet platform of Tsing-Hua University's research and Chinese Academy of Sciences's exploitation.These systems database used has comprised natural image, biometric image, and multispectral image etc., but for the singularity of SAR image, these systems are also inapplicable.

Images Classification is the key link that realizes CBIR.At present, conventional image classification method is mainly divided into two classes: the method and the unsupervised method that have supervision.There is the sorting technique of supervision to comprise: arest neighbors and k-neighbour, Bayes classifier, support vector machine and neural network.These methods can realize the classification of image fast and accurately, but because need foundation and the study of model in processing procedure, the time complexity of method is higher; Unsupervised image classification method has cluster analysis and fuzzy cluster analysis.The implementation procedure of these two kinds of methods is more quick, but the accuracy of classification is lower.Therefore, how realizing fast and accurately the classification of SAR image, is the major issue that solves SAR image retrieval.

Summary of the invention

The object of the invention is to the shortcoming for above-mentioned existing system and method, proposed a kind of based on sparse coding classification have a supervision SAR image search method, it can solve the sparse solution of global optimum fast.The method not only can reduce the computation complexity that SAR image is processed, and can effectively reduce the impact that coherent speckle noise causes SAR image searching result.

Technical scheme of the present invention is that the SAR image search method based on sparse coding classification, is characterized in that: at least comprise the steps:

Step 101: start the SAR image retrieval based on sparse coding classification;

Step 102: choose training image in SAR image library, read in these images, every width image size of selecting in this patent is 256 * 256, and utilize the method for exquisite Lee filtering to carry out pre-service to it, reduce the impact that coherent speckle noise causes Images Classification result, the window size of wave filter is set as 7 * 7;

Step 103: utilize the method for gray level co-occurrence matrixes to carry out feature extraction to pretreated training image, choose 0 °, 45 °, 90 °, energy in 135 ° of directions, entropy, contrast, local similarity, relevant each five features, it is 20 column vector that each width training image correspondence obtains a dimension;

If p _ijwhen (d, θ) is illustrated in given space length d and direction θ, take gray scale i as initial point, occur gray level j probability (i=1,2 ... G; J=1,2 ... G), G is the maximal value of gray level in institute's image under consideration region, and gray level co-occurrence matrixes is the square formation of a G * G, note

$p_x\left(i\right)=\underset{j=1}{\overset{G}{\mathrm{\Σ}}}{p}_{\mathrm{ij}}$ Formula (1)

$p_y\left(j\right)=\underset{i=1}{\overset{G}{\mathrm{\Σ}}}{p}_{\mathrm{ij}}$ Formula (2)

The value of setting θ is 0 °, 45 °, 90 °, 135 °, in order to obtain the textural characteristics of image in all directions, above-mentioned 4 directions is constructed respectively to gray level co-occurrence matrixes and extract corresponding proper vector; The feature of selecting comprises following five kinds of features:

A) energy, claims again angle second moment:

$\mathrm{ASM}=\underset{i=1}{\overset{G}{\mathrm{\Σ}}}\underset{j=i}{\overset{G}{\mathrm{\Σ}}}{p_{\mathrm{ij}}}^2$ Formula (3)

It is the tolerance of gradation of image distributing homogeneity or flatness.In gray level co-occurrence matrixes, element distributes while concentrating near principal diagonal, and the interior gradation of image distribution uniform of regional area be described, and image presents thinner texture, and the value of angle second moment is larger;

B) entropy:

$\mathrm{ENT}=-\underset{i=1}{\overset{G}{\mathrm{\Σ}}}\underset{j=1}{\overset{G}{\mathrm{\Σ}}}{p}_{\mathrm{ij}}\log p_{\mathrm{ij}}$ Formula (4)

Entropy is the tolerance of quantity of information that image has, and is the characteristic parameter of measuring grey level distribution randomness, has characterized the complexity of texture in image; The gray scale of image is more even, and entropy is less, and the texture of image is more complicated, and entropy is larger; On the other hand, entropy also can be measured the randomness of image texture, and entropy is larger, and in representative image, intensity profile randomness is large;

C) contrast, claims again moment of inertia:

$\mathrm{CON}=\underset{i=1}{\overset{G}{\mathrm{\Σ}}}\underset{j=1}{\overset{G}{\mathrm{\Σ}}}{(i-j)}^2\·p_{\mathrm{ij}}$ Formula (5)

Local gray level in its token image changes total amount, has reflected the sharpness of image and the rill depth of texture; The rill of texture is dark, and contrast is large, and effect is clear; Otherwise contrast is little, rill is shallow, and effect is fuzzy; That is to say, if depart from cornerwise unit, have higher value, i.e. gradation of image value variation is very fast, and contrast value has larger value;

D) local similarity:

$\mathrm{HOM}=\underset{i=1}{\overset{G}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{\mathrm{<figure-callout\; id="1"\; label="G"\; filenames="HDA0000445116450000022.png"\; state="\{\{state\}\}">G</figure-callout>}}{\mathrm{\&Sigma;}}}\frac{1}{1+{(i-j)}^2}\&CenterDot;p_{\mathrm{ij}}$ Formula (6)

Local similar performance is portrayed the textural characteristics of regional area, is the important tolerance of distinguishing different target;

E) relevant

$\mathrm{COR}=\frac{1}{{\mathrm{\&sigma;}}_x{\mathrm{\&sigma;}}_y}\underset{i=1}{\overset{G}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{G}{\mathrm{\&Sigma;}}}(i-{\mathrm{\&mu;}}_x)(j-{\mathrm{\&mu;}}_y)\&CenterDot;p_{\mathrm{ij}}$ Formula (7) wherein, ${\mathrm{\&mu;}}_x=\underset{i}{\mathrm{\&Sigma;}}i\underset{j}{\mathrm{\&Sigma;}}p(i,j),{\mathrm{\&mu;}}_y=\underset{j}{\mathrm{\&Sigma;}}j\underset{i}{\mathrm{\&Sigma;}}p(i,j)$

Therefore, after having calculated gray level co-occurrence matrixes, every width image can obtain 20 features, and dimensionality reduction to dimension is 20 column vector;

Step 104: utilize the proper vector obtaining in step 103 to form complete dictionary, train sparse sorter;

Step 105: utilize the sparse sorter training to classify to whole SAR picture library;

Step 106: class mark and the corresponding sparse solution of every width image in storage SAR image library;

Step 107: import testing SA R image, require the SAR image equal and opposite in direction in its size and image library, according to the method in step 102, test pattern is carried out to filtering processing;

Step 108: according to the method for the gray level co-occurrence matrixes in step 103, test pattern is carried out to feature extraction, obtain corresponding 20 features;

Step 109: judge in the SAR image library whether test pattern obtain in step 106, if so, directly carry out step 110, otherwise enter step 111;

Step 110: directly extract the SAR image identical with test pattern in having class target SAR image library;

Step 111: the sparse sorter training before utilizing is classified to test pattern;

Step 112: the class label of storage test pattern and corresponding sparse solution thereof;

Step 113: propose the image of identical category according to the classification of test pattern in having class target SAR image library, carry out image similarity coupling; First calculate the Euclidean distance of the sparse solution of test pattern and generic image, then find the position at greatest coefficient place in sparse solution, calculate the difference of two positions, the module of similarity is set as the inverse of absolute value of the product of Euclidean distance and position difference, be worth greatlyr, represent that similarity value is higher.Similarity expression formula is:

$\mathrm{Sim}(x,u)=1/\left({\left(\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{(x_i-u_i)}^2\right)}^{\frac{1}{2}}\right|m_x-n_u\left|\right)$ Formula (8) wherein

for Euclidean distance expression formula, x _irepresent i locational sparse coefficient in the sparse solution that test pattern is corresponding, m _xposition for greatest coefficient place; u _ii locational sparse coefficient in the sparse solution of the training image that representative and test sample book are generic, n _uposition for its greatest coefficient place.

Step 114: by the similarity value obtaining, arrange according to order from big to small, return to result for retrieval;

Step 115: the SAR image search method that finishes sparse coding classification.

Described step 104, comprises the steps:

Step 201: start to build complete dictionary and train sparse sorter;

Step 202: constructed complete dictionary: training image characteristic of correspondence vector is arranged according to classification, and the characteristic series vector of identical category is emitted on together successively, constructs the complete dictionary A=[of the needed mistake of rarefaction representation χ ₁, χ ₂... χ _n], χ _irepresentative is a class training sample wherein, χ _i=[α ₁, α ₂... α _k], the total classification number of training sample is n;

Step 203: obtain new part and cross complete dictionary: by sample to be sorted and each training sample subtraction calculations residual values of crossing in complete dictionary, set a threshold value T, the corresponding training sample that residual values is greater than to threshold value proposes, and forms new part and crosses complete dictionary A ₁=[χ ₁, χ ₂... χ _k];

Step 204: utilize based on the rare optimization method y=Ax of dual Local Search Memetic Algorithm for Solving, obtain sparse solution x, in optimization method, y is original test pattern, A is the complete dictionary of mistake that training image forms, x is the sparse solution that y is corresponding;

Step 205: design category function δ _i, (δ _i∈ R ^{m * n}), coefficients all kinds of in sparse solution x is proposed respectively, build the sparse solution δ making new advances _i(x).In new sparse solution, only have the coefficient value of a class non-vanishing, the value of remainder is zero; Therefore, according to following expression re-formation, go out test sample book

$\hat{y}=A{\mathrm{\&delta;}}_i\left(x\right)$ The process simplification of formula (9) final image classification is for solving following problem:

$\hat{x}=\left\{\begin{array}{c}\min \left|\right|x{\left|\right|}_1\\ s.t.\left|\right|y-A{\mathrm{\&delta;}}_i\left(x\right){\left|\right|}_2\&le;\mathrm{\&epsiv;}\end{array}\right.$ Formula (10) wherein, || || ₁represent L ₁-norm, || || ₂represent L ₂-norm, A is the complete dictionary of mistake that training image forms.

Calculate original sample y to be sorted and reconstructed sample

two poor norms, the classification at the minimum value place obtaining is the classification that test sample book belongs to;

Step 206: the process that finishes the sparse sorter of training.

Described step 204, comprises the steps:

Step 301: start to utilize the rare optimization method of Memetic Algorithm for Solving based on dual Local Search, obtain sparse solution;

Step 302: individual chooses and encode, and complete dictionary A is crossed in part ₁in the position of training sample as individuality, encode, coded system adopts decimal coded mode, establishes and in each individuality, comprises five sample position;

Step 303: five training samples corresponding to coding site are proposed to form new dictionary A ₂, utilizing matching pursuit algorithm-MP, solving-optimizing problem y=Ax, obtains the sparse coefficient under this dictionary;

Step 304: the selection of fitness function: by sample y to be sorted and reconstructed sample two norms of difference be made as fitness function, the less expression fitness of value of two norms is higher;

Step 305: judge whether fitness value meets first end condition-residual values and be less than setting value or reach maximum iteration time, if meet, directly jump to step 310, otherwise continue execution step 306;

Step 306: select, according to the height of fitness value, sample is selected, retain excellent individual, the i.e. higher individuality of fitness;

Step 307: intersect, the mode of intersection is the random point of crossing that produces, and the part after point of crossing is exchanged, and adjacent individuality intersects between two;

Step 308: variation, the mode of variation makes a variation for choosing at random single-point.

Step 309: first stage Local Search obtains when preferably individual, at local dictionary A after each iteration ₁middle using the left and right n of each a position neighborhood position as Local Search candidate constituency, in candidate regions, the height according to fitness is selected again, upgrades existing more excellent individuality; After completing first stage Local Search, get back to global search process steps 304, calculate individual fitness value, carry out the judgement that next carries out instruction;

Step 310: the Local Search of subordinate phase after completing all iterative process, according to the method for the search in step 309, again carries out Local Search one time in the complete dictionary A of whole mistake;

Step 311: judge whether to meet end condition-residual values and be less than setting value or reach maximum iteration time, carry out step 312 if meet, otherwise return to step 310;

Step 312: export the final global optimum satisfying condition individuality, the position in individuality is the position at our needed sparse coefficient place;

Step 313: the process that finishes to utilize the Memetic Algorithm for Solving sparse solution based on dual Local Search.

Compared with prior art there is following advantage in the present invention:

1. this method is first classified to SAR image, then, according to class mark retrieval similar image, has reduced like this impact that coherent speckle noise causes SAR image searching result.

2. in assorting process, by the position of training sample in crossing complete dictionary of encoding, we can build a new dictionary A ₂thereby, with sparse coefficient still less, represent original test sample book, effectively reduced operand, be conducive to realize fast the classification of image.

3. in assorting process, the method for dual Local Search has been proposed herein.In each selection, intersect, after variation, carry out first stage Local Search.According to preferably individual at dictionary A ₁in choose candidate region, candidate region is set as a left and right n neighborhood of these five positions, the method can obtain optimum solution simply, efficiently.After iterative process completes, carry out the Local Search of subordinate phase, now we are amplified to candidate space in the complete dictionary A of whole mistake, have guaranteed like this of overall importance of optimum solution.And, to compare with other evolution algorithms, the method solution procedure is more quick.

4. when similarity is mated, choose the absolute value of Euclidean distance and position difference product as module, combine the detailed information of image, with principal component information, make result for retrieval more accurate.

Below in conjunction with process flow diagram 1 and other accompanying drawings, specific embodiment of the invention step is further described.

Accompanying drawing explanation

Fig. 1 is the SAR image retrieval process flow diagram that the present invention is based on sparse classification;

Fig. 2 is the process flow diagram that the present invention realizes Images Classification;

Fig. 3 (a) is rarefaction representation process, is (b) example of sparse solution x;

Fig. 4 is the process flow diagram of the Memetic algorithm based on dual Local Search in this paper;

Fig. 5 is the mode of individual coding when producing initial population and the process diagram that produces new dictionary;

Fig. 6 chooses the method figure of Local Search candidate region after calculating more excellent solution;

Fig. 7 is that the 5 required class SAR images of classifying in experimentation are herein followed successively by city, farmland, bridge, mountains and rivers, waters;

Fig. 8 is by method in this paper and uses quadrature matching algorithm, least square method, and base tracing algorithm, step is moved orthogonal matching pursuit algorithm, the comparing result of the classify accuracy that Memetic algorithm obtains;

Fig. 9 is the result for retrieval of every class SAR image.

Embodiment

As shown in Figure 1.SAR image search method based on sparse coding classification, at least comprises process step:

Step 101: start the SAR image search method based on sparse coding classification;

Step 102: choose training image in SAR image library, read in these images, every width image size of selecting in this patent is 256 * 256, and utilize the method for exquisite Lee filtering to carry out pre-service to it, reduce the impact that coherent speckle noise causes Images Classification result, the window size of wave filter is set as 7 * 7;

Step 103: utilize the method for gray level co-occurrence matrixes to carry out feature extraction to pretreated training image, the value of set direction θ is 0 °, 45 °, 90 °, 135 °, i.e. east-west, northeast-southwest, south-north, 4 of the southeast-northwests direction; Each direction is chosen 5 features, is respectively energy, entropy, contrast, local similarity and relevant.Therefore, every width image can obtain 20 features, and dimensionality reduction to dimension is 20 column vector.

In an embodiment of the present invention, using the processing of classifying of 5 class SAR images, is respectively cities and towns, farmland, bridge, mountains and rivers and waters, and every class amount of images is 200, and image size is 256 * 256, as shown in Figure 7;

Step 104: utilize in step 103 to proper vector formed complete dictionary, solving-optimizing equation y=Ax, trains sparse sorter.In optimization method, y is original test pattern, and A is the complete dictionary of mistake that training image forms, and x is the sparse solution that y is corresponding;

Step 105: utilize the sparse sorter training to classify to whole SAR picture library;

Step 106: class mark and the corresponding sparse solution of every width image in storage SAR image library;

Step 107: import testing SA R image, require the SAR image equal and opposite in direction in its size and image library, according to the method in step 102, test pattern is carried out to filtering processing.

In an embodiment of the present invention, the number of every class testing sample is respectively, and 360,433,167,400,400;

Step 108: according to the method for the gray level co-occurrence matrixes in step 103, test pattern is carried out to feature extraction, obtain corresponding 20 features;

Step 109: judge in the SAR image library whether test pattern obtain in step 106, if so, directly carry out step 110, otherwise enter step 111;

Step 110: directly extract the SAR image identical with test pattern in having class target SAR image library;

Step 111: the sparse sorter training before utilizing is classified to test pattern;

Step 112: the class label of storage test pattern and corresponding sparse solution thereof;

Step 113: propose the image of identical category according to the classification of test pattern in having class target SAR image library, carry out image similarity coupling.First calculate the Euclidean distance of the sparse solution of test pattern and generic image, then find the position at greatest coefficient place in sparse solution, calculate the difference of two positions, the module of similarity is set as the inverse of absolute value of the product of Euclidean distance and position difference, be worth greatlyr, represent that similarity value is higher.Similarity expression formula is:

$\mathrm{Sim}(x,u)=1/\left({\left(\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{(x_i-u_i)}^2\right)}^{\frac{1}{2}}\right|m_x-n_u\left|\right)$ Formula (8) wherein for Euclidean distance expression formula, x _irepresent i locational sparse coefficient in the sparse solution that test pattern is corresponding, m _xposition for greatest coefficient place; u _ii locational sparse coefficient in the sparse solution of the training image that representative and test sample book are generic, n _uposition for its greatest coefficient place.

Step 114: by the similarity value obtaining, arrange according to order from big to small, return to result for retrieval;

Step 115: the SAR image search method that finishes sparse coding classification.

As shown in Figure 2,

Described step 104, comprises the steps:

Step 201: start to build complete dictionary and train sparse sorter;

Step 202: constructed complete dictionary: training image characteristic of correspondence vector is arranged according to classification, and the characteristic series vector of identical category is emitted on together successively, constructs the complete dictionary A=[of the needed mistake of rarefaction representation χ ₁, χ ₂... χ _n], χ _irepresentative is a class training sample wherein, χ _i=[α ₁, α ₂... α _k], the total classification number of training sample is n;

Step 203: obtain new part and cross complete dictionary: by sample to be sorted and each training sample subtraction calculations residual values of crossing in complete dictionary, set a threshold value T, the corresponding training sample that residual values is greater than to threshold value proposes, and forms new part and crosses complete dictionary A ₁=[χ ₁, χ ₂... χ _k];

Step 204: utilize based on the rare optimization method y=Ax of dual Local Search Memetic Algorithm for Solving, obtain sparse solution x, in optimization method, y is original test pattern, A is the complete dictionary of mistake that training image forms, x is the sparse solution that y is corresponding;

In an embodiment of the present invention, the sparse solution arriving of solving-optimizing equation as shown in Figure 3;

Step 205: design category function δ _i, (δ _i∈ R ^{m * n}), coefficients all kinds of in sparse solution x is proposed respectively, build the sparse solution δ making new advances _i(x).In new sparse solution, only have the coefficient value of a class non-vanishing, the value of remainder is zero; Therefore, according to following expression re-formation, go out test sample book

$\hat{y}=A{\mathrm{\&delta;}}_i\left(x\right)$ The process simplification of formula (9) final image classification is for solving following problem:

$\hat{x}=\left\{\begin{array}{c}\min \left|\right|x{\left|\right|}_1\\ s.t.\left|\right|y-A{\mathrm{\&delta;}}_i\left(x\right){\left|\right|}_2\&le;\mathrm{\&epsiv;}\end{array}\right.$ Formula (10) wherein, || || ₁represent L ₁-norm, || || ₂represent L ₂-norm, A is the complete dictionary of mistake that training image forms.

Calculate original sample y to be sorted and reconstructed sample

two poor norms, the classification at the minimum value place obtaining is the classification that test sample book belongs to;

Step 206: the process that finishes the sparse sorter of training;

As shown in Figure 4,

Described step 204, comprises the steps:

Step 301: start to utilize the rare optimization method of Memetic Algorithm for Solving based on dual Local Search, obtain sparse solution;

Step 302: individual chooses and encode, and complete dictionary A is crossed in part ₁in the position of training sample as individuality, encode, coded system adopts decimal coded mode, establishes and in each individuality, comprises five sample position;

In an embodiment of the present invention, each individuality of every class sample to be sorted five positions that are encoded, every generation comprises 50 individualities, calculates altogether for 200 generations, and its coded system is as shown in Figure 5.

Step 303: five training samples corresponding to coding site are proposed to form new dictionary A ₂, utilizing matching pursuit algorithm-MP, solving-optimizing problem y=Ax, obtains the sparse coefficient under this dictionary;

Step 304: the selection of fitness function: by sample y to be sorted and reconstructed sample

two norms of difference be made as fitness function, the less expression fitness of value of two norms is higher;

Step 305: judge whether fitness value meets first end condition-residual values and be less than setting value or reach maximum iteration time, if meet, directly jump to step 310, otherwise continue execution step 306;

Step 306: select, according to the height of fitness value, sample is selected, retain excellent individual, the i.e. higher individuality of fitness;

Step 307: intersect, the mode of intersection is the random point of crossing that produces, and the part after point of crossing is exchanged, and two adjacent individualities intersect between two, and in an embodiment of the present invention, crossover probability is made as 0.6;

Step 308: variation, the mode of variation makes a variation for choosing at random single-point, and in an embodiment of the present invention, variation probability is made as 0.01;

Step 309: first stage Local Search obtains when preferably individual, at local dictionary A after each iteration ₁middle using the left and right n of each a position neighborhood position as Local Search candidate constituency, in candidate regions, the height according to fitness is selected again, upgrades existing more excellent individuality; After completing first stage Local Search, get back to global search process steps 304, calculate individual fitness value, carry out the judgement that next carries out instruction;

In an embodiment of the present invention, using the left and right n of each position in an individuality neighborhood position as Local Search candidate constituency, selection mode as shown in Figure 6;

Step 310: the Local Search of subordinate phase after completing all iterative process, according to the method for the search in step 309, again carries out Local Search one time in the complete dictionary A of whole mistake;

Step 311: judge whether to meet end condition-residual values and be less than setting value or reach maximum iteration time, carry out step 312 if meet, otherwise return to step 310;

Step 312: export the final global optimum satisfying condition individuality, the position in individuality is the position at our needed sparse coefficient place;

Step 313: the process that finishes to utilize the Memetic Algorithm for Solving sparse solution based on dual Local Search.

The part that the present embodiment does not describe in detail belongs to the known conventional means of the industry, here not narration one by one.

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

1. experiment condition and content:

Experiment condition:

At CPU, be to use Matlab2010 to carry out emulation in core22.4GHZ, internal memory 2G, WINDOWSXP system.

Experiment content:

The present invention tests SAR image library used and comprises 5 class SAR images, is respectively cities and towns, farmland, bridge, mountains and rivers and waters, and image size is 256 * 256, and sum is respectively 560,633,367,600,600, as shown in Figure 7.In every class image, select at random 200 width as training image, remaining is as test pattern.

(1) comparison of classify accuracy: use respectively algorithm in this paper and orthogonal matching pursuit algorithm, least square method, step is moved orthogonal matching pursuit algorithm, base tracing algorithm, Memetic algorithm etc. carries out Images Classification, more final classification results.

(2) calculate the precision ratio of SAR image retrieval:

Recall ratio and precision ratio are the standard evaluation methods in information retrieval, are used more and more in the middle of CBIR now, use precision ratio as the evaluation criterion of result for retrieval herein, the accuracy of its reflection result for retrieval.Precision ratio is defined as:

$\mathrm{Precision}=p\left(A\right|B)=\frac{p(A\&cup;B)}{p\left(B\right)}$ Wherein, Q represents whole image data base to formula (11), and A represents the set of associated picture, the image collection that B representative retrieves.

2. experimental result:

(1) by above-mentioned carried method, this 5 class SAR image is classified, as shown in Figure 7, wherein green line represents the classification results based on orthogonal matching pursuit algorithm to result; The classification results of blue line representative based on least square method; Black line is the result that base method for tracing obtains; Yellow line moves for walking the result that orthogonal matching pursuit algorithm obtains; Red line is the classification results based on Memetic algorithm; Pink colour line is the Memetic based on dual Local Search that proposes herein and the resulting classification results of method of sparse coding.

From result contrast table 1, evolution algorithm can play good effect in the optimization problem of Images Classification, compares Memetic algorithm, and the method that we propose can obtain higher classify accuracy, and robustness is higher.Concrete classify accuracy is (%):

Table 1

(2) result of the SAR image retrieval obtaining by method in this paper is as shown in table 2, and the precision ratio that every class SAR image calculation obtains is:

Table 2

In summary, method in this paper when solving the problem of SAR image retrieval, in classify accuracy, retrieving similarity, computation complexity, result robustness aspect has all played preferably effect.

Claims (3)

1. the SAR image search method based on sparse coding classification, is characterized in that: at least comprise the steps:

Step 101: start the SAR image retrieval based on sparse coding classification;

Step 102: choose training image in SAR image library, read in these images, every width image size of selecting in this patent is 256 * 256, and utilize the method for exquisite Lee filtering to carry out pre-service to it, reduce the impact that coherent speckle noise causes Images Classification result, the window size of wave filter is set as 7 * 7;

Step 103: utilize the method for gray level co-occurrence matrixes to carry out feature extraction to pretreated training image, choose 0 °, 45 °, 90 °, energy in 135 ° of directions, entropy, contrast, local similarity, relevant each five features, it is 20 column vector that each width training image correspondence obtains a dimension;

If p _ijwhen (d, θ) is illustrated in given space length d and direction θ, take gray scale i as initial point, occur gray level j probability (i=1,2 ... G; J=1,2 ... G), G is the maximal value of gray level in institute's image under consideration region, and gray level co-occurrence matrixes is the square formation of a G * G, note

$p_x\left(i\right)=\underset{j=1}{\overset{G}{\mathrm{\&Sigma;}}}{p}_{\mathrm{ij}}$ Formula (1)

$p_y\left(j\right)=\underset{i=1}{\overset{G}{\mathrm{\&Sigma;}}}{p}_{\mathrm{ij}}$ Formula (2)

The value of setting θ is 0 °, 45 °, 90 °, 135 °, in order to obtain the textural characteristics of image in all directions, above-mentioned 4 directions is constructed respectively to gray level co-occurrence matrixes and extract corresponding proper vector; The feature of selecting comprises following five kinds of features:

1) energy, claims again angle second moment:

$\mathrm{ASM}=\underset{i=1}{\overset{G}{\mathrm{\&Sigma;}}}\underset{j=i}{\overset{G}{\mathrm{\&Sigma;}}}{p_{\mathrm{ij}}}^2$ Formula (3)

It is the tolerance of gradation of image distributing homogeneity or flatness.In gray level co-occurrence matrixes, element distributes while concentrating near principal diagonal, and the interior gradation of image distribution uniform of regional area be described, and image presents thinner texture, and the value of angle second moment is larger;

2) entropy:

$\mathrm{ENT}=-\underset{i=1}{\overset{G}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{G}{\mathrm{\&Sigma;}}}{p}_{\mathrm{ij}}\log p_{\mathrm{ij}}$ Formula (4)

Entropy is the tolerance of quantity of information that image has, and is the characteristic parameter of measuring grey level distribution randomness, has characterized the complexity of texture in image; The gray scale of image is more even, and entropy is less, and the texture of image is more complicated, and entropy is larger; On the other hand, entropy also can be measured the randomness of image texture, and entropy is larger, and in representative image, intensity profile randomness is large;

3) contrast, claims again moment of inertia:

$\mathrm{CON}=\underset{i=1}{\overset{G}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{G}{\mathrm{\&Sigma;}}}{(i-j)}^2\&CenterDot;p_{\mathrm{ij}}$ Formula (5)

Local gray level in its token image changes total amount, has reflected the sharpness of image and the rill depth of texture; The rill of texture is dark, and contrast is large, and effect is clear; Otherwise contrast is little, rill is shallow, and effect is fuzzy; That is to say, if depart from cornerwise unit, have higher value, i.e. gradation of image value variation is very fast, and contrast value has larger value;

4) local similarity:

$\mathrm{HOM}=\underset{i=1}{\overset{G}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{G}{\mathrm{\&Sigma;}}}\frac{1}{1+{(i-j)}^2}\&CenterDot;p_{\mathrm{ij}}$ Formula (6)

Local similar performance is portrayed the textural characteristics of regional area, is the important tolerance of distinguishing different target;

5) relevant

$\mathrm{COR}=\frac{1}{{\mathrm{\&sigma;}}_x{\mathrm{\&sigma;}}_y}\underset{i=1}{\overset{G}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{G}{\mathrm{\&Sigma;}}}(i-{\mathrm{\&mu;}}_x)(j-{\mathrm{\&mu;}}_y)\&CenterDot;p_{\mathrm{ij}}$ Formula (7) wherein, ${\mathrm{\&mu;}}_x=\underset{i}{\mathrm{\&Sigma;}}i\underset{j}{\mathrm{\&Sigma;}}p(i,j),{\mathrm{\&mu;}}_y=\underset{j}{\mathrm{\&Sigma;}}j\underset{i}{\mathrm{\&Sigma;}}p(i,j).$

Therefore, after having calculated gray level co-occurrence matrixes, every width image can obtain 20 features, and dimensionality reduction to dimension is 20 column vector;

Step 104: utilize the proper vector obtaining in step 103 to form complete dictionary, train sparse sorter;

Step 105: utilize the sparse sorter training to classify to whole SAR picture library;

Step 106: class mark and the corresponding sparse solution of every width image in storage SAR image library;

Step 107: import testing SA R image, require the SAR image equal and opposite in direction in its size and image library, according to the method in step 102, test pattern is carried out to filtering processing;

Step 108: according to the method for the gray level co-occurrence matrixes in step 103, test pattern is carried out to feature extraction, obtain corresponding 20 features;

Step 109: judge in the SAR image library whether test pattern obtain in step 106, if so, directly carry out step 110, otherwise enter step 111;

Step 110: directly extract the SAR image identical with test pattern in having class target SAR image library;

Step 111: the sparse sorter training before utilizing is classified to test pattern;

Step 112: the class label of storage test pattern and corresponding sparse solution thereof;

Step 113: propose the image of identical category according to the classification of test pattern in having class target SAR image library, carry out image similarity coupling; First calculate the Euclidean distance of the sparse solution of test pattern and generic image, then find the position at greatest coefficient place in sparse solution, calculate the difference of two positions, the module of similarity is set as the inverse of absolute value of the product of Euclidean distance and position difference, be worth greatlyr, represent that similarity value is higher.Similarity expression formula is:

$\mathrm{Sim}(x,u)=1/\left({\left(\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{(x_i-u_i)}^2\right)}^{\frac{1}{2}}\right|m_x-n_u\left|\right)$ Formula (8) wherein

for Euclidean distance expression formula, x _irepresent i locational sparse coefficient in the sparse solution that test pattern is corresponding, m _xposition for greatest coefficient place; u _ii locational sparse coefficient in the sparse solution of the training image that representative and test sample book are generic, nu is the position at its greatest coefficient place.

Step 114: by the similarity value obtaining, arrange according to order from big to small, return to result for retrieval;

Step 115: the SAR image search method that finishes sparse coding classification.

2. the SAR image search method based on sparse coding classification according to claim 1, is characterized in that: described step 104, comprises the steps:

Step 201: start to build complete dictionary and train sparse sorter;

Step 202: constructed complete dictionary: training image characteristic of correspondence vector is arranged according to classification, and the characteristic series vector of identical category is emitted on together successively, constructs the complete dictionary A=[of the needed mistake of rarefaction representation χ ₁, χ ₂... χ _n], χ _irepresentative is a class training sample wherein, χ _i=[α ₁, α ₂... α _k], the total classification number of training sample is n;

Step 203: obtain new part and cross complete dictionary: by sample to be sorted and each training sample subtraction calculations residual values of crossing in complete dictionary, set a threshold value T, the corresponding training sample that residual values is greater than to threshold value proposes, and forms new part and crosses complete dictionary A ₁=[χ ₁, χ ₂... χ _k];

Step 204: utilize based on the rare optimization method y=Ax of dual Local Search Memetic Algorithm for Solving, obtain sparse solution x, in optimization method, y is original test pattern, A is the complete dictionary of mistake that training image forms, x is the sparse solution that y is corresponding;

Step 205: design category function δ _i, (δ _i∈ R ^{m * n}), coefficients all kinds of in sparse solution x is proposed respectively, build the sparse solution δ making new advances _i(x).In new sparse solution, only have the coefficient value of a class non-vanishing, the value of remainder is zero; Therefore, according to following expression re-formation, go out test sample book

$\hat{y}=A{\mathrm{\&delta;}}_i\left(x\right)$ Formula (9)

The process simplification of final image classification is for solving following problem:

$\hat{x}=\left\{\begin{array}{c}\min \left|\right|x{\left|\right|}_1\\ s.t.\left|\right|y-A{\mathrm{\&delta;}}_i\left(x\right){\left|\right|}_2\&le;\mathrm{\&epsiv;}\end{array}\right.$ Formula (10) wherein, || || ₁represent L ₁-norm, || || ₂represent L ₂-norm, A is the complete dictionary of mistake that training image forms.

Calculate original sample y to be sorted and reconstructed sample two poor norms, the classification at the minimum value place obtaining is the classification that test sample book belongs to;

Step 206: the process that finishes the sparse sorter of training.

3. the SAR image search method based on sparse coding classification according to claim 2, is characterized in that: described step 204, comprises the steps:

Step 301: start to utilize the rare optimization method of Memetic Algorithm for Solving based on dual Local Search, obtain sparse solution;

Step 302: individual chooses and encode, and complete dictionary A is crossed in part ₁in the position of training sample as individuality, encode, coded system adopts decimal coded mode, establishes and in each individuality, comprises five sample position;

Step 303: five training samples corresponding to coding site are proposed to form new dictionary A ₂, utilizing matching pursuit algorithm-MP, solving-optimizing problem y=Ax, obtains the sparse coefficient under this dictionary;

Step 304: the selection of fitness function: by sample y to be sorted and reconstructed sample

two norms of difference be made as fitness function, the less expression fitness of value of two norms is higher;

Step 305: judge whether fitness value meets first end condition-residual values and be less than setting value or reach maximum iteration time, if meet, directly jump to step 310, otherwise continue execution step 306;

Step 306: select, according to the height of fitness value, sample is selected, retain excellent individual, the i.e. higher individuality of fitness;

Step 307: intersect, the mode of intersection is the random point of crossing that produces, and the part after point of crossing is exchanged, and adjacent individuality intersects between two, and in an embodiment of the present invention, crossover probability is made as 0.6;

Step 308: variation, the mode of variation makes a variation for choosing at random single-point, and in an embodiment of the present invention, variation probability is made as 0.01;

Step 309: first stage Local Search obtains when preferably individual, at local dictionary A after each iteration ₁middle using the left and right n of each a position neighborhood position as Local Search candidate constituency, in candidate regions, the height according to fitness is selected again, upgrades existing more excellent individuality; After completing first stage Local Search, get back to global search process steps 304, calculate individual fitness value, carry out the judgement that next carries out instruction;

Step 310: the Local Search of subordinate phase after completing all iterative process, according to the method for the search in step 309, again carries out Local Search one time in the complete dictionary A of whole mistake;

Step 311: judge whether to meet end condition-residual values and be less than setting value or reach maximum iteration time, carry out step 312 if meet, otherwise return to step 310;

Step 312: export the final global optimum satisfying condition individuality, the position in individuality is the position at our needed sparse coefficient place;

Step 313: the process that finishes to utilize the Memetic Algorithm for Solving sparse solution based on dual Local Search.

Images