CN103714148B - SAR image search method based on sparse coding classification - Google Patents
SAR image search method based on sparse coding classification Download PDFInfo
- Publication number
- CN103714148B CN103714148B CN201310733522.8A CN201310733522A CN103714148B CN 103714148 B CN103714148 B CN 103714148B CN 201310733522 A CN201310733522 A CN 201310733522A CN 103714148 B CN103714148 B CN 103714148B
- Authority
- CN
- China
- Prior art keywords
- image
- sparse
- classification
- value
- training
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/58—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
- G06F16/583—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content
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
Technical field
The invention belongs to SAR image process field, is related to a kind of SAR image search method based on sparse coding, can be with
Accurately classified and realized retrieval to SAR image, significantly reduced shadow of the coherent speckle noise to SAR image classification results
Ring.
Background technology
Synthetic aperture radar-SyntheticApertureRadar, is a kind of effective means from earth observation from space,
It has been widely applied to the aspects such as military surveillance, landform observation, urban planning.With carrying for SAR imaging techniques in recent years
Height, the quantity of SAR image is presented explosive growth, many for SAR image data volume, the characteristics of self-noise is big, how from
Image required for efficiently and accurately retrieving in the SAR image storehouse of magnanimity has become problem demanding prompt solution.
With the development of information technology, the search method of image is transformed into from initial text based retrieval method
Content-based retrieval method, the method can directly be analyzed to image, feature extraction, and similarity analysis simultaneously realize inspection
Rope.At present, main image indexing system includes：The QBIC systems of IBM exploitations, Virage systems, stanford university research
The SIMPLIcity systems of exploitation, the still image searching system of the Internet platforms of Tsing-Hua University's research and Chinese Academy of Sciences's exploitation
Mires image retrieval prototype systems.Data base used by these systems contains natural image, biometric image, multispectral figure
As etc., but for the particularity of SAR image, these systems are not simultaneously applied to.
Image classification is to realize the key link of CBIR.At present, conventional image classification method master
It is divided into two classes：There are the method and unsupervised method of supervision.The sorting technique for having supervision includes：Arest neighbors and k- neighbours, shellfish
Leaf this grader, support vector machine and neutral net.These methods can fast and accurately realize the classification of image, but be because
The foundation and study of model are needed in processing procedure, the time complexity of method is higher；Unsupervised image classification method has
Cluster analyses and fuzzy cluster analysis.Both approaches realize that process is more quick, but the accuracy of classification is but than relatively low.
Therefore, the classification of SAR image how is fast and accurately realized, is the major issue for solving SAR image retrieval.
The content of the invention
Present invention aims to the shortcoming of above-mentioned existing system and method, it is proposed that one kind is based on sparse coding
That what is classified has supervision SAR image search method, and it can quickly solve the sparse solution of global optimum.The method can not only
The computation complexity that SAR image is processed is reduced, and coherent speckle noise can be effectively reduced and SAR image retrieval result is caused
Impact.
The technical scheme is that, based on the SAR image search method that sparse coding is classified, it is characterized in that：At least wrap
Include following steps：
Step 101：Start based on the SAR image retrieval of sparse coding classification；
Step 102：Training image is chosen in SAR image storehouse, these images are read in, each image selected in this patent
Size is 256 × 256, and the method filtered using exquisite Lee carries out pretreatment to it, reduces coherent speckle noise to image classification
As a result the impact for causing, the window size of wave filter is set as 7 × 7；
Step 103：Feature extraction is carried out using the method for gray level co-occurrence matrixes to pretreated training image, is chosen
0 °, 45 °, 90 °, the energy, entropy, contrast, local similarity, related each five features on 135 ° of directions, each width training figure
As correspondence obtains the column vector that dimension is 20；
If pij(d, θ) is represented in given space length d and direction θ, with gray scale i as initial point, the probability of gray level j is occurred
(i=1,2 ... G;J=1,2 ... G), G is the maximum of gray level in institute's image under consideration region, then gray level co-occurrence matrixes are one
The square formation of individual G × G, note
Formula（1）
Formula（2）
The value of θ is set as 0 °, 45 °, 90 °, 135 °, in order to obtain image textural characteristics in all directions, to upper
State 4 directions and construct the corresponding characteristic vector of gray level co-occurrence matrixes extraction respectively；From feature include following five kinds of features：
A) energy, also known as angular second moment：
Formula（3）
It is the tolerance of gradation of image distributing homogeneity or flatness.When Elemental redistribution is relatively concentrated in gray level co-occurrence matrixes
When near leading diagonal, gradation of image distribution uniform in regional area is illustrated, image is presented thinner texture, angular second moment
Value is larger；
B) entropy：
Formula（4）
Entropy is the tolerance that image has quantity of information, is the characteristic parameter for measuring grey level distribution randomness, is characterized
The complexity of texture in image；The gray scale of image is more uniform, and entropy is less, and the texture of image is more complicated, and entropy is then bigger；Separately
On the one hand, entropy can also measure the randomness of image texture, and entropy is bigger, and gray scale distribution randomness is big in representative image；
C) contrast, also known as the moment of inertia：
Formula（5）
Local gray level change total amount in its phenogram picture, reflects the definition of image and the rill depth of texture；Stricture of vagina
The rill depth of reason, contrast is big, and effect is clear；Conversely, contrast is little, then rill is shallow, and effect is obscured；If that is, partially
Higher value, i.e. gradation of image value changes are have from cornerwise unit quickly, then contrast value has larger value；
D) local similarity：
Formula（6）
Local similar performance portrays the textural characteristics of regional area, is to discriminate between the important measure of different target；
E) it is related
Formula（7）
Wherein,
Therefore, after gray level co-occurrence matrixes have been calculated, each image can obtain 20 features, i.e. dimensionality reduction to a dimension
It is 20 column vector；
Step 104：Complete dictionary was constituted using the characteristic vector obtained in step 103, sparse grader was trained；
Step 105：Whole SAR picture libraries are classified using the sparse grader for training；
Step 106：The category and corresponding sparse solution of each image in storage SAR image storehouse；
Step 107：Import test SAR image, it is desirable to which its size is equal in magnitude with the SAR image in image library, according to step
Method in rapid 102 is filtered process to test image；
Step 108：Method according to the gray level co-occurrence matrixes in step 103 carries out feature extraction to test image, obtains
Corresponding 20 features；
Step 109：Judge that test image whether in the SAR image storehouse that step 106 is obtained, is if so, then directly walked
Rapid 110, otherwise into step 111；
Step 110：Directly extract and test image identical SAR image in the SAR image storehouse for having category；
Step 111：Test image is classified using the sparse grader for training before；
Step 112：The class label and its corresponding sparse solution of storage test image；
Step 113：The image of identical category is proposed in the SAR image storehouse for having category according to the classification of test image, is entered
Row image similarity is matched；Test image and the Euclidean distance of the sparse solution of generic image are calculated first, are then found sparse
The position that greatest coefficient is located in solution, calculates the difference of two positions, and the module of similarity is set as Euclidean distance and position
The inverse of the absolute value of the product of difference, value is bigger, represents Similarity value higher.Similarity expression formula is：
Formula（8）
WhereinFor Euclidean distance expression formula, xiRepresent i-th in the corresponding sparse solution of test image
Sparse coefficient on individual position, mxFor the position that greatest coefficient is located；uiRepresent the dilute of the training image generic with test sample
Discongest the sparse coefficient on i-th position, nuFor the position that its greatest coefficient is located.
Step 114：By the Similarity value for obtaining, arranged according to order from big to small, returned retrieval result；
Step 115：Terminate the SAR image search method of sparse coding classification.
Described step 104, comprises the steps：
Step 201：Start to build complete dictionary and train sparse grader；
Step 202：Constructed complete dictionary：The corresponding characteristic vector of training image is arranged according to classification, it is identical
The feature column vector of classification is discharged successively together, constructs the excessively complete dictionary A=[χ required for rarefaction representation1,χ2,...
χn], χiRepresent one type training sample, χi=[α1,α2,...αk], the total classification number of training sample is n；
Step 203：Obtain the excessively complete dictionary in new local：By each the training sample in sample to be sorted and excessively complete dictionary
This subtraction calculations residual values, set threshold value T, and residual values are proposed more than the training sample corresponding to threshold value, and composition is new
The excessively complete dictionary A in local1=[χ1,χ2,...χk]；
Step 204：Using based on the dilute optimization method y=Ax of dual Local Search Memetic Algorithm for Solving, sparse solution is obtained
X, y is original test image in optimization method, and A is the excessively complete dictionary of training image composition, and x is the corresponding sparse solutions of y；
Step 205：Design classification function δi,（δi∈Rm×n）, coefficients all kinds of in sparse solution x are proposed respectively, construct new
Sparse solution δi(x).The coefficient value for there was only a class in new sparse solution is not zero, and the value of remainder is zero；Therefore, press
Test sample is reconstructed according to below equation
Formula（9）
The process simplification of final image classification is the following problem of solution：
Formula（10）
Wherein, | | | |1Represent L1- norm, | | | |2Represent L2- norm, A is the excessively complete dictionary of training image composition.
Calculate original sample y to be sorted and reconstructed sampleTwo poor norms, the classification that the minima for obtaining is located is
The classification that test sample belongs to；
Step 206：Terminate the process of the sparse grader of training.
Described step 204, comprises the steps：
Step 301：Start with based on the dilute optimization method of Memetic Algorithm for Solving of dual Local Search, obtain sparse
Solution；
Step 302：The selection of individuality and coding, by the excessively complete dictionary A in local1In training sample position as individuality
Encoded, coded system adopts decimal coded mode, if including five sample positions in each individuality；
Step 303：The corresponding training sample of five coding sites is proposed into the new dictionary A of composition2, calculated using match tracing
Method-MP, solving-optimizing problem y=Ax obtains the sparse coefficient under the dictionary；
Step 304：The selection of fitness function：By sample y to be sorted and reconstructed sampleDifference two norms be set to adapt to
Degree function, the value of two norms is less to represent that fitness is higher；
Step 305：Judge whether fitness value meets first end condition-residual values less than setting value or reach most
Big iterationses, if meeting, jump directly to step 310, otherwise continue executing with step 306；
Step 306：Select, sample is selected according to the height of fitness value, retain excellent individual, i.e. fitness compared with
High individuality；
Step 307：Intersect, the mode of intersection is handed over the part behind cross point to randomly generate a cross point
Change, adjacent individuality intersects two-by-two；
Step 308：Variation, the mode of variation enters row variation to randomly select single-point.
Step 309：First stage Local Search, when obtaining preferably individual after each iteration, in local dictionary A1In
Using the neighborhood position of left and right n of each position as Local Search candidate constituency, again according to the height of fitness in candidate regions
It is low to be selected, update existing more excellent individuality；After first stage Local Search is completed, global search process steps are returned to
304, individual fitness value is calculated, carry out the judgement of next execute instruction；
Step 310：The Local Search of second stage, after all iterative process are completed, according to the search in step 309
Method, in whole excessively complete dictionary A a Local Search is carried out again；
Step 311：Judge whether to meet end condition-residual values less than setting value or reach maximum iteration time, if full
Foot carries out step 312, otherwise return to step 310；
Step 312：The final global optimum for meeting condition of output is individual, and the position in individuality is required for us
The position that sparse coefficient is located；
Step 313：Terminate the process using the Memetic Algorithm for Solving sparse solutions based on dual Local Search.
There is advantages below compared with prior art in the present invention：
1. this method is first classified to SAR image, then retrieves similar image according to category, which reduces coherent spot
The impact that noise is caused to SAR image retrieval result.
2. in categorizing process, by encoding position of the training sample in excessively complete dictionary, we can build one
New dictionary A2, so as to represent original test sample using less sparse coefficient, operand is efficiently reduced, be conducive to
Quickly realize the classification of image.
3. in categorizing process, this paper presents the method for dual Local Search.Selecting every time, intersecting, after variation,
Carry out first stage Local Search.According to preferably individuality in dictionary A1Middle selection candidate region, candidate region be set as this five
The neighborhood of left and right n of individual position, the method can simply, efficiently obtain optimal solution.After iterative process is completed, is carried out
The Local Search of two-stage, now we candidate spatial is amplified in entirely excessively complete dictionary A, which ensures that optimal solution
It is of overall importance.Also, compared with other evolution algorithms, the method solution procedure is more quick.
4. in similarity mode, the absolute value of Euclidean distance and position difference product is chosen as module, with reference to
The detailed information of image, with principal component information, makes retrieval result more accurate.
The specific implementation step of the present invention is further described with reference to flow process Fig. 1 and other accompanying drawings.
Description of the drawings
Fig. 1 is SAR image retrieval flow figure of the present invention based on sparse classification；
Fig. 2 is the flow chart that the present invention realizes image classification；
Fig. 3 (a) is rarefaction representation process, (b) is the example of sparse solution x；
Fig. 4 be set forth herein the Memetic algorithms based on dual Local Search flow chart；
Fig. 5 is the process diagram of the mode of individual UVR exposure and the new dictionary of generation when initial population is produced；
Fig. 6 is the method figure that Local Search candidate region is chosen after more excellent solution is calculated；
Fig. 7 is that 5 class SAR images needed for classifying in this paper experimentations are followed successively by city, farmland, bridge, mountains and rivers, water
Domain；
Fig. 8 be use set forth herein method and use orthogonal matching algorithm, method of least square, base tracing algorithm, step move it is orthogonal
Matching pursuit algorithm, the comparing result of the classification accuracy that Memetic algorithms are obtained；
Fig. 9 is the retrieval result of every class SAR image.
Specific embodiment
As shown in Figure 1.Based on the SAR image search method that sparse coding is classified, at least including process step：
Step 101：Start based on the SAR image search method of sparse coding classification；
Step 102：Training image is chosen in SAR image storehouse, these images are read in, each image selected in this patent
Size is 256 × 256, and the method filtered using exquisite Lee carries out pretreatment to it, reduces coherent speckle noise to image classification
As a result the impact for causing, the window size of wave filter is set as 7 × 7；
Step 103：Feature extraction, set direction are carried out using the method for gray level co-occurrence matrixes to pretreated training image
The value of θ is 0 °, 45 °, 90 °, 135 °, i.e. east-west, northeast-southwest, southern-northern, direction of the southeast-northwest 4；Each side
It is respectively energy, entropy, contrast, local similarity and correlation to 5 features are chosen.Therefore, each image can obtain 20
Feature, i.e. dimensionality reduction are to the column vector that dimension is 20.
In an embodiment of the present invention, classification process is carried out using 5 class SAR images, is respectively cities and towns, farmland, bridge, mountain
River and waters, are 200 per class amount of images, and image size is 256 × 256, as shown in Figure 7；
Step 104：Using in step 103 to characteristic vector constituted complete dictionary, solving-optimizing equation y=Ax, instruction
The sparse grader of white silk.Y is original test image in optimization method, and A is the excessively complete dictionary of training image composition, and x is that y is corresponding
Sparse solution；
Step 105：Whole SAR picture libraries are classified using the sparse grader for training；
Step 106：The category and corresponding sparse solution of each image in storage SAR image storehouse；
Step 107：Import test SAR image, it is desirable to which its size is equal in magnitude with the SAR image in image library, according to step
Method in rapid 102 is filtered process to test image.
In an embodiment of the present invention, it is respectively per the number of class testing sample, 360,433,167,400,400；
Step 108：Method according to the gray level co-occurrence matrixes in step 103 carries out feature extraction to test image, obtains
Corresponding 20 features；
Step 109：Judge that test image whether in the SAR image storehouse that step 106 is obtained, is if so, then directly walked
Rapid 110, otherwise into step 111；
Step 110：Directly extract and test image identical SAR image in the SAR image storehouse for having category；
Step 111：Test image is classified using the sparse grader for training before；
Step 112：The class label and its corresponding sparse solution of storage test image；
Step 113：The image of identical category is proposed in the SAR image storehouse for having category according to the classification of test image, is entered
Row image similarity is matched.Test image and the Euclidean distance of the sparse solution of generic image are calculated first, are then found sparse
The position that greatest coefficient is located in solution, calculates the difference of two positions, and the module of similarity is set as Euclidean distance and position
The inverse of the absolute value of the product of difference, value is bigger, represents Similarity value higher.Similarity expression formula is：
Formula（8）
WhereinFor Euclidean distance expression formula, xiRepresent i-th in the corresponding sparse solution of test image
Sparse coefficient on individual position, mxFor the position that greatest coefficient is located；uiRepresent the dilute of the training image generic with test sample
Discongest the sparse coefficient on i-th position, nuFor the position that its greatest coefficient is located.
Step 114：By the Similarity value for obtaining, arranged according to order from big to small, returned retrieval result；
Step 115：Terminate the SAR image search method of sparse coding classification.
As shown in Fig. 2
Described step 104, comprises the steps：
Step 201：Start to build complete dictionary and train sparse grader；
Step 202：Constructed complete dictionary：The corresponding characteristic vector of training image is arranged according to classification, it is identical
The feature column vector of classification is discharged successively together, constructs the excessively complete dictionary A=[χ required for rarefaction representation1,χ2,...
χn], χiRepresent one type training sample, χi=[α1,α2,...αk], the total classification number of training sample is n；
Step 203：Obtain the excessively complete dictionary in new local：By each the training sample in sample to be sorted and excessively complete dictionary
This subtraction calculations residual values, set threshold value T, and residual values are proposed more than the training sample corresponding to threshold value, and composition is new
The excessively complete dictionary A in local1=[χ1,χ2,...χk]；
Step 204：Using based on the dilute optimization method y=Ax of dual Local Search Memetic Algorithm for Solving, sparse solution is obtained
X, y is original test image in optimization method, and A is the excessively complete dictionary of training image composition, and x is the corresponding sparse solutions of y；
In an embodiment of the present invention, the sparse solution for arriving of solving-optimizing equation is as shown in Figure 3；
Step 205：Design classification function δi,（δi∈Rm×n）, coefficients all kinds of in sparse solution x are proposed respectively, construct new
Sparse solution δi(x).The coefficient value for there was only a class in new sparse solution is not zero, and the value of remainder is zero；Therefore, press
Test sample is reconstructed according to below equation
Formula（9）
The process simplification of final image classification is the following problem of solution：
Formula（10）
Wherein, | | | |1Represent L1- norm, | | | |2Represent L2- norm, A is the excessively complete dictionary of training image composition.
Calculate original sample y to be sorted and reconstructed sampleTwo poor norms, the classification that the minima for obtaining is located is
The classification that test sample belongs to；
Step 206：Terminate the process of the sparse grader of training；
As shown in figure 4,
Described step 204, comprises the steps：
Step 301：Start with based on the dilute optimization method of Memetic Algorithm for Solving of dual Local Search, obtain sparse
Solution；
Step 302：The selection of individuality and coding, by the excessively complete dictionary A in local1In training sample position as individuality
Encoded, coded system adopts decimal coded mode, if including five sample positions in each individuality；
In an embodiment of the present invention, five positions are encoded per each individuality of class sample to be sorted, are included per a generation
50 individualities, calculated altogether for 200 generations, and its coded system is as shown in Figure 5.
Step 303：The corresponding training sample of five coding sites is proposed into the new dictionary A of composition2, calculated using match tracing
Method-MP, solving-optimizing problem y=Ax obtains the sparse coefficient under the dictionary；
Step 304：The selection of fitness function：By sample y to be sorted and reconstructed sampleDifference two norms be set to adapt to
Degree function, the value of two norms is less to represent that fitness is higher；
Step 305：Judge whether fitness value meets first end condition-residual values less than setting value or reach most
Big iterationses, if meeting, jump directly to step 310, otherwise continue executing with step 306；
Step 306：Select, sample is selected according to the height of fitness value, retain excellent individual, i.e. fitness compared with
High individuality；
Step 307：Intersect, the mode of intersection is handed over the part behind cross point to randomly generate a cross point
Change, two adjacent individualities intersect two-by-two, and in an embodiment of the present invention, crossover probability is set to 0.6；
Step 308：Variation, the mode of variation enters row variation to randomly select single-point, in an embodiment of the present invention, variation
Probability is set to 0.01；
Step 309：First stage Local Search, when obtaining preferably individual after each iteration, in local dictionary A1In
Using the neighborhood position of left and right n of each position as Local Search candidate constituency, again according to the height of fitness in candidate regions
It is low to be selected, update existing more excellent individuality；After first stage Local Search is completed, global search process steps are returned to
304, individual fitness value is calculated, carry out the judgement of next execute instruction；
In an embodiment of the present invention, using the neighborhood position of left and right n of each position in individuality as Local Search candidate
Constituency, selection mode is as shown in Figure 6；
Step 310：The Local Search of second stage, after all iterative process are completed, according to the search in step 309
Method, in whole excessively complete dictionary A a Local Search is carried out again；
Step 311：Judge whether to meet end condition-residual values less than setting value or reach maximum iteration time, if full
Foot carries out step 312, otherwise return to step 310；
Step 312：The final global optimum for meeting condition of output is individual, and the position in individuality is required for us
The position that sparse coefficient is located；
Step 313：Terminate the process using the Memetic Algorithm for Solving sparse solutions based on dual Local Search.
Part of the present embodiment without detailed narration belongs to the known conventional means of the industry, does not describe one by one here.
The effect of the present invention can be further illustrated by following emulation experiment：
1. experiment condition and content：
Experiment condition：
It is core22.4GHZ, emulated using Matlab2010 in internal memory 2G, WINDOWSXP system in CPU.
Experiment content：
Present invention experiment SAR image storehouse used includes 5 class SAR images, respectively cities and towns, farmland, bridge, mountains and rivers and water
Domain, image size is 256 × 256, and sum is respectively 560,633,367,600,600, as shown in Figure 7.In every class image with
Machine selects 200 width as training image, and remaining is used as test image.
(1) comparison of accuracy of classifying：Respectively use set forth herein algorithm and orthogonal matching pursuit algorithm, least square
Method, step is moved orthogonal matching pursuit algorithm, base tracing algorithm, Memetic algorithms etc. and carries out image classification, the final classification knot of comparison
Really.
(2) precision ratio of SAR image retrieval is calculated：
Recall ratio and the standard evaluation methodology that precision ratio is in information retrieval, are increasingly used and are based on now
In the middle of the image retrieval of content, herein using precision ratio as retrieval result evaluation criterion, it reflect retrieval result it is accurate
Property.Then precision ratio is defined as：
Formula（11）
Wherein, Q represents whole image data base, and A represents the set of associated picture, and B represents the image collection for retrieving.
2. experimental result：
(1) this 5 class SAR image is classified with above-mentioned carried method, as a result as shown in fig. 7, wherein green line generation
Classification results of the table based on orthogonal matching pursuit algorithm；Blue line is represented based on the classification results of method of least square；Black line is
The result that base method for tracing is obtained；Yellow line is that step moves the result that orthogonal matching pursuit algorithm is obtained；Red line be based on
The classification results of Memetic algorithms；Pink colour line is the proposed Memetic based on dual Local Search and sparse coding
Method obtained by classification results.
Knowable to Comparative result table 1, evolution algorithm can play preferable effect, phase in the optimization problem of image classification
Than Memetic algorithm, it is proposed that method can obtain higher classification accuracy, and robustness is higher.Concrete classification
Accuracy is (%)：
Table 1
(2) use set forth herein the result of SAR image retrieval that obtains of method as shown in table 2, calculate per class SAR image
To precision ratio be：
Table 2
In summary, set forth herein method solve the problems, such as SAR image retrieve when, classification accuracy, retrieve phase
Like property, computation complexity, as a result preferably effect is all served in terms of robustness.
Claims (2)
1. the SAR image search method based on sparse coding classification, is characterized in that：At least comprise the steps：
Step 101：Start based on the SAR image retrieval of sparse coding classification；
Step 102：Training image to be chosen in SAR image storehouse, these images are read in, each image size of selection is 256 ×
256, and the method filtered using exquisite Lee carries out pretreatment to it, reduces what coherent speckle noise was caused to image classification result
Affect, the window size of wave filter is set as 7 × 7；
Step 103：Feature extraction is carried out using the method for gray level co-occurrence matrixes to pretreated training image, 0 ° is chosen,
45 °, 90 °, the energy, entropy, contrast, local similarity, related each five features on 135 ° of directions, each width training image pair
The column vector that dimension is 20 should be obtained；
If pij(d, θ) is represented in given space length d and direction θ, with gray scale i as initial point, the probability of gray level j, i=is occurred
1,2 ..., G, j=1,2 ..., G, G are the maximum of gray level in institute's image under consideration region, then gray level co-occurrence matrixes are a G
The square formation of × G, note
The value of θ is set as 0 °, 45 °, 90 °, 135 °, in order to obtain image textural characteristics in all directions, to above-mentioned 4
Individual direction constructs respectively gray level co-occurrence matrixes and extracts corresponding characteristic vector；From feature include following five kinds of features：
1) energy, also known as angular second moment：
It is the tolerance of gradation of image distributing homogeneity or flatness, when in gray level co-occurrence matrixes Elemental redistribution relatively concentrate on lead it is right
When near linea angulata, gradation of image distribution uniform in regional area is illustrated, image is presented thinner texture, the value of angular second moment
It is larger；
2) entropy：
Entropy is the tolerance that image has quantity of information, is the characteristic parameter for measuring grey level distribution randomness, characterizes image
The complexity of middle texture；The gray scale of image is more uniform, and entropy is less, and the texture of image is more complicated, and entropy is then bigger；The opposing party
Face, entropy can also measure the randomness of image texture, and entropy is bigger, and gray scale distribution randomness is big in representative image；
3) contrast, also known as the moment of inertia：
Local gray level change total amount in its phenogram picture, reflects the definition of image and the rill depth of texture；Texture
Rill depth, contrast is big, and effect is clear；Conversely, contrast is little, then rill is shallow, and effect is obscured；That is, if the deviation from right
The unit of linea angulata have higher value, i.e. gradation of image value changes quickly, then contrast value has larger value；
4) local similarity：
Local similar performance portrays the textural characteristics of regional area, is to discriminate between the important measure of different target；
5) it is related
Wherein,
Therefore, after gray level co-occurrence matrixes have been calculated, each image can obtain 20 features, i.e. dimensionality reduction to a dimension be 20
Column vector；
Step 104：Complete dictionary was constituted using the characteristic vector obtained in step 103, sparse grader was trained；
Step 105：Whole SAR picture libraries are classified using the sparse grader for training；
Step 106：The category and corresponding sparse solution of each image in storage SAR image storehouse；
Step 107：Import test SAR image, it is desirable to which its size is equal in magnitude with the SAR image in image library, according to step 102
In method process is filtered to test image；
Step 108：Method according to the gray level co-occurrence matrixes in step 103 carries out feature extraction to test image, obtains corresponding
20 features；
Step 109：Test image is judged whether in the SAR image storehouse that step 106 is obtained, if so, then directly carry out step
110, otherwise into step 111；
Step 110：Directly extract and test image identical SAR image in the SAR image storehouse for having category；
Step 111：Test image is classified using the sparse grader for training before；
Step 112：The class label and its corresponding sparse solution of storage test image；
Step 113：The image of identical category is proposed in the SAR image storehouse for having category according to the classification of test image, figure is carried out
As similarity mode；Test image and the Euclidean distance of the sparse solution of generic image are calculated first, are then found in sparse solution
The position that greatest coefficient is located, calculates the difference of two positions, and the module of similarity is set as Euclidean distance and position difference
Product absolute value inverse, be worth bigger, represent that Similarity value is higher, similarity expression formula is：
WhereinFor Euclidean distance expression formula, xiRepresent i-th in the corresponding sparse solution of test image
Sparse coefficient on individual position, mxFor the position that greatest coefficient is located；uiRepresent the dilute of the training image generic with test sample
Discongest the sparse coefficient on i-th position, nuFor the position that its greatest coefficient is located,
Step 114：By the Similarity value for obtaining, arranged according to order from big to small, returned retrieval result；
Step 115：Terminate the SAR image search method of sparse coding classification；
Described step 104, comprises the steps：
Step 201：Start to build complete dictionary and train sparse grader；
Step 202：Constructed complete dictionary：The corresponding characteristic vector of training image is arranged according to classification, identical category
Feature column vector discharge successively together, construct the excessively complete dictionary A=[x required for rarefaction representation1, x2..., xn],
χiRepresent one type training sample, xi=[a1, a2..., ak], the total classification number of training sample is n；
Step 203：Obtain the excessively complete dictionary in new local：By each the training sample phase in sample to be sorted and excessively complete dictionary
Subtract calculating residual values, set threshold value T, residual values are proposed more than the training sample corresponding to threshold value, constitute new office
The excessively complete dictionary A in portion1=[x1, x2..., xn1]；
Step 204：Using based on the dilute optimization method y=Ax of dual Local Search Memetic Algorithm for Solving, sparse solution x is obtained, it is excellent
It is original test image to change y in equation, and A is the excessively complete dictionary of training image composition, and x is the corresponding sparse solutions of y；
Step 205：Design classification function δi, (δi∈Rm×n), all kinds of coefficients in sparse solution x are proposed respectively, construct new dilute
Discongest δiX (), the coefficient value for there was only a class in new sparse solution is not zero, the value of remainder is zero；Therefore, according to
Lower expression re-formation goes out test sample
The process simplification of final image classification is the following problem of solution：
Wherein, | | | |1Represent L1- norm, | | | |2Represent L2- norm, A is the excessively complete dictionary of training image composition；
Calculate original sample y to be sorted and reconstructed sampleTwo poor norms, the classification that the minima for obtaining is located is test specimens
Originally the classification for belonging to；
Step 206：Terminate the process of the sparse grader of training.
2. the SAR image search method classified based on sparse coding according to claim 1, is characterized in that：Described step
Rapid 204, comprise the steps：
Step 301：Start with based on the dilute optimization method of Memetic Algorithm for Solving of dual Local Search, obtain sparse solution；
Step 302：The selection of individuality and coding, by the excessively complete dictionary A in local1In the position of training sample carry out as individuality
Coding, coded system adopts decimal coded mode, if including five coding sites in each individuality；
Step 303：The corresponding training sample of five coding sites is proposed into the new dictionary A of composition2, using matching pursuit algorithm-
MP, solving-optimizing problem y=Ax obtains the A2Sparse coefficient under dictionary；
Step 304：The selection of fitness function：By sample y to be sorted and reconstructed sampleTwo norms of difference be set to fitness letter
Number, the value of two norms is less to represent that fitness is higher；
Step 305：Judge whether fitness value meets first end condition-residual values less than setting value or reach maximum and change
Generation number, if meeting, jumps directly to step 310, otherwise continues executing with step 306；
Step 306：Select, sample is selected according to the height of fitness value, retain excellent individual, i.e. fitness higher
It is individual；
Step 307：Intersect, the mode of intersection swaps the part behind cross point, phase to randomly generate a cross point
Adjacent individuality intersects two-by-two, and crossover probability is set to 0.6；
Step 308：Variation, the mode of variation enters row variation to randomly select single-point, and mutation probability is set to 0.01；
Step 309：First stage Local Search, when obtaining preferably individual after each iteration, in local dictionary A1It is middle by each
The neighborhood position of left and right n of position is selected as Local Search candidate regions, the height again according to fitness in candidate regions
Select, update existing more excellent individuality；After first stage Local Search is completed, global search process steps 304 are returned to, calculate individual
The fitness value of body, carries out the judgement of next execute instruction；
Step 310：The Local Search of second stage, after all iterative process are completed, according to the side of the search in step 309
Method, in whole excessively complete dictionary A a Local Search is carried out again；
Step 311：Judge whether to meet end condition-residual values less than setting value or reach maximum iteration time, if meet into
Row step 312, otherwise return to step 310；
Step 312：The final global optimum for meeting condition of output is individual, and the position in individuality is sparse required for us
The position that coefficient is located；
Step 313：Terminate the process using the Memetic Algorithm for Solving sparse solutions based on dual Local Search.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201310733522.8A CN103714148B (en) | 2013-12-24 | 2013-12-24 | SAR image search method based on sparse coding classification |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201310733522.8A CN103714148B (en) | 2013-12-24 | 2013-12-24 | SAR image search method based on sparse coding classification |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN103714148A CN103714148A (en) | 2014-04-09 |
| CN103714148B true CN103714148B (en) | 2017-05-10 |
Family
ID=50407123
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201310733522.8A Active CN103714148B (en) | 2013-12-24 | 2013-12-24 | SAR image search method based on sparse coding classification |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN103714148B (en) |
Families Citing this family (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104156918B (en) * | 2014-08-01 | 2017-02-15 | 西安电子科技大学 | SAR image noise suppression method based on joint sparse representation and residual fusion |
| CN105320964B (en) * | 2015-10-26 | 2019-06-04 | 中国矿业大学(北京) | Coal-rock identification method based on rarefaction representation |
| CN106951922A (en) * | 2017-03-16 | 2017-07-14 | 太原理工大学 | A kind of real-time screening system of astronomic graph picture based on SVMs |
| CN108846414B (en) * | 2018-05-22 | 2021-09-28 | 浙江理工大学 | SAR image subcategory classification method based on decision-level fusion idea |
| CN109471106B (en) * | 2018-11-22 | 2021-04-13 | 上海海事大学 | SAR ocean internal wave stripe recognition method combining clustering analysis and boundary tracking method |
| CN109815357B (en) * | 2019-01-28 | 2022-12-06 | 辽宁工程技术大学 | Remote sensing image retrieval method based on nonlinear dimension reduction and sparse representation |
| CN113032533B (en) * | 2021-05-31 | 2021-09-03 | 宏龙科技(杭州)有限公司 | Efficient retrieval method based on sparse vector matching |
| CN113870120B (en) * | 2021-12-07 | 2022-03-01 | 领伟创新智能系统(浙江)有限公司 | Processing surface texture inclination correction method based on pq-mean distribution |
| CN117557582B (en) * | 2024-01-02 | 2024-03-12 | 深圳翰博设计股份有限公司 | Artificial intelligence-based building landscape design image processing system |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101551809A (en) * | 2009-05-13 | 2009-10-07 | 西安电子科技大学 | Search method of SAR images classified based on Gauss hybrid model |
| CN102509288A (en) * | 2011-10-19 | 2012-06-20 | 西安电子科技大学 | SAR (Synthetic Aperture Radar) image change detection method based on Memetic kernel clustering |
| CN102968796A (en) * | 2012-11-30 | 2013-03-13 | 西安电子科技大学 | SAR (Synthetic Aperture Radar) image segmentation method based on sampling learning |
| CN103345511A (en) * | 2013-07-04 | 2013-10-09 | 西北工业大学 | Remote sensing image searching method based on sparse representation |
| CN103366371A (en) * | 2013-06-25 | 2013-10-23 | 西安电子科技大学 | K distribution and texture feature-based SAR (Synthetic Aperture Radar) image segmentation method |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8620093B2 (en) * | 2010-03-15 | 2013-12-31 | The United States Of America As Represented By The Secretary Of The Army | Method and system for image registration and change detection |
-
2013
- 2013-12-24 CN CN201310733522.8A patent/CN103714148B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101551809A (en) * | 2009-05-13 | 2009-10-07 | 西安电子科技大学 | Search method of SAR images classified based on Gauss hybrid model |
| CN102509288A (en) * | 2011-10-19 | 2012-06-20 | 西安电子科技大学 | SAR (Synthetic Aperture Radar) image change detection method based on Memetic kernel clustering |
| CN102968796A (en) * | 2012-11-30 | 2013-03-13 | 西安电子科技大学 | SAR (Synthetic Aperture Radar) image segmentation method based on sampling learning |
| CN103366371A (en) * | 2013-06-25 | 2013-10-23 | 西安电子科技大学 | K distribution and texture feature-based SAR (Synthetic Aperture Radar) image segmentation method |
| CN103345511A (en) * | 2013-07-04 | 2013-10-09 | 西北工业大学 | Remote sensing image searching method based on sparse representation |
Non-Patent Citations (1)
| Title |
|---|
| 一种快速的基于稀疏表示分类器;陈才扣 等;《南京大学学报（自然科学）》;20120131;第48卷(第1期);第70-76页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN103714148A (en) | 2014-04-09 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN103714148B (en) | SAR image search method based on sparse coding classification | |
| Xia et al. | AID: A benchmark data set for performance evaluation of aerial scene classification | |
| CN107885764B (en) | Rapid Hash vehicle retrieval method based on multitask deep learning | |
| Wang et al. | Word–sentence framework for remote sensing image captioning | |
| CN110929080B (en) | Optical remote sensing image retrieval method based on attention and generation countermeasure network | |
| CN112668494A (en) | Small sample change detection method based on multi-scale feature extraction | |
| CN110276316A (en) | A kind of human body critical point detection method based on deep learning | |
| Fan et al. | Unsupervised feature learning for land-use scene recognition | |
| CN108280187B (en) | Hierarchical image retrieval method based on depth features of convolutional neural network | |
| Grilli et al. | Application of machine and deep learning strategies for the classification of heritage point clouds | |
| He et al. | A dual global–local attention network for hyperspectral band selection | |
| CN111680176A (en) | Remote sensing image retrieval method and system based on attention and bidirectional feature fusion | |
| CN108154094B (en) | Hyperspectral image unsupervised waveband selection method based on subinterval division | |
| CN108984642A (en) | A kind of PRINTED FABRIC image search method based on Hash coding | |
| CN105894030B (en) | High-resolution remote sensing image scene classification method based on layering multiple features fusion | |
| CN114821198B (en) | Cross-domain hyperspectral image classification method based on self-supervision and small sample learning | |
| CN103336801A (en) | Multi-feature locality sensitive hashing (LSH) indexing combination-based remote sensing image retrieval method | |
| CN107958067A (en) | It is a kind of based on without mark Automatic Feature Extraction extensive electric business picture retrieval system | |
| Chen et al. | Measuring the effectiveness of various features for thematic information extraction from very high resolution remote sensing imagery | |
| CN106844739A (en) | A kind of Remote Sensing Imagery Change information retrieval method based on neutral net coorinated training | |
| Xie et al. | Spatial pooling of heterogeneous features for image applications | |
| CN103093243A (en) | High resolution panchromatic remote sensing image cloud discriminating method | |
| Thirumaladevi et al. | Remote sensing image scene classification by transfer learning to augment the accuracy | |
| CN115690549A (en) | Target detection method for realizing multi-dimensional feature fusion based on parallel interaction architecture model | |
| CN104881668A (en) | Method and system for extracting image fingerprints based on representative local mode |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |