CN101853304B - Remote sensing image retrieval method based on feature selection and semi-supervised learning - Google Patents

Abstract

The invention discloses a remote sensing image retrieval method based on feature selection and semi-supervised learning. In the method, an optimal color feature and an optimal textural feature are selected respectively by utilizing a clustering method according to a minimum description length criterion and an improved Davies-Bouldin index; and then an appropriate semi-supervised learning method is selected according to the binarization weight of the optimal color feature and the optimal textural feature for carrying out remote sensing image retrieval. Compared with the traditional remote sensing image retrieval method, the invention not only can greatly improve the retrieval quality, but also can reduce the calculate quantity in the retrieval process and improve the retrieval speed.

Description

Remote sensing image retrieval method based on feature selecting and semi-supervised learning

Technical field

The present invention relates to image search method, relate in particular to a kind of remote sensing image retrieval method.

Background technology

Along with the continuous development of remote sensing technology, the remote sensing images quantity that obtain every day sharply increases, and the research of automatic inquiry of remote sensing images and retrieval technique is become gradually the problem of urgent need research.At present, Chinese scholars has proposed a lot of methods and has carried out the retrieval of content-based remote sensing images (CBIR), as textural characteristics based on the Gabor conversion, color characteristic and textural characteristics combination, textural characteristics and spatial information merge, histogram feature similarity measurement method, and based on GIS space method of semantic etc.Propositions such as Zhu Bin utilize the Gabor textural characteristics to carry out retrieval [the Bin Zhu of aerial image, Marshall R, Hsinchun C.Creating a large-scale content-based airphoto image digital library IEEE Trans onimage processing, 2000, vol.9, no.1:163-167.]; The Gabor textural characteristics is merged in propositions such as Lu Lizhen and color characteristic carries out remote Sensing Image Retrieval, and adopt the linear weighted function of texture and color characteristic Euclidean distance to measure similarity [Lu Lizhen, Liu Renyi, Liu Nan. the remote sensing image retrieval method of a kind of Fusion of Color and textural characteristics, China's image graphics journal (A), 2004,9 (3): 328-332.]; Ceng Zhiming etc. utilize improved co-occurrence matrix textural characteristics to carry out large scale remote Sensing Image Retrieval [Ceng Zhiming, Li Feng, Fu Kun, Deng. a kind of texture feature extraction algorithm of large scale remote sensing images information retrieval based on contents, Wuhan University's journal (information science version), 2005,30 (12): 1080-1083.].Bao Qian and Guo Ping be at the single band remote Sensing Image Retrieval, studied respectively based on the similarity measurement of proper vector with based on the similarity measurement of probability, finds χ ²Statistical distance is more effective to first kind of similarity measurement with similar included angle cosine tolerance, and based on the computing method of K-neighbour rule to the effective [Bao Qian of second kind of similarity measurement, Guo Ping. based on histogrammic remote sensing images similarity retrieval method relatively, the remote sensing journal, 2006,10 (6): 893-900.].Ferecatu and Boujemaa propose to utilize the method for active relevant feedback to carry out interactive remote Sensing Image Retrieval [Marin Ferecatu, NozhaBoujemaa.Interactive remote sensing image retrieval using active relevance feedback.IEEE Transactionson geoscience and remote sensing, 2007, vol.45, no.4:818-826.].

CBIR relies on feature extraction and High-dimensional Index Technology to retrieve, the method that adopts is: system extracts some low layer visual signatures (as color, texture, shape etc.) automatically from each width of cloth image, form with high dimension vector deposits database in, and relatively the similarity of these features obtains result for retrieval then.Mainly concentrate on Feature Extraction and fusion aspect at content-based remote Sensing Image Retrieval Study on Technology in the above-mentioned prior art, but all do not notice such fact: dissimilar searched targets, feature should be different.For same width of cloth image, different features is also different aspect the validity of describing its content, can represent that the feature of searched targets content should effectively improve retrieval performance if therefore extract.

Relevant feedback (Relevance feedback) is a learning strategy the most frequently used among the CBIR, and it relies on the man-machine interaction process, and the user constantly feeds back, and its performance increases and improves along with the feedback samples collection, but also can increase user's burden simultaneously greatly.Provide a large amount of heavy burdens of marker samples in order to reduce the user because of repeatedly feeding back, also there is the scholar to propose to utilize the semi-supervised learning strategy to carry out image retrieval, the main thought of this strategy is to utilize a large amount of unmarked examples to assist study to a small amount of underlined example, whole learning process does not need manual intervention, only unmarked example is utilized based on learning algorithm self, for example, Yao etc. have proposed a kind of medical image search method (SEMI-SECC) [Jian Yao that corrects output encoder based on the semantic error of semi-supervised learning, Zhongfei Zhang, Antani S, et al.Automatic Medical ImageAnnotation and Retrieval using SEMI-SECC[C] .Proceedings of IEEE International Conference onMultimedia and Expo, Piscaaway, NJ, United States:IEEE Press, 2006:2005-2008.].Because in content-based remote Sensing Image Retrieval, usually has only seldom example sample (sometimes even have only an example goal sample), and to obtain that more mark example sample is also very difficult, therefore the retrieval of adopting semi-supervised learning to carry out remote sensing images is one and more reasonably selects.

Summary of the invention

We know: dissimilar searched targets, feature is different, and for the description of same searched targets content, the validity of different characteristic is different, carry out image retrieval if can find out the feature that can represent the searched targets content, that just can improve retrieval performance greatly.

Based on such thinking, the present invention tries hard to provide a kind of remote sensing image retrieval method in conjunction with feature selecting, promptly at image to be retrieved, selects and can represent that the feature of searched targets content carries out image retrieval.

The present invention utilizes the method for cluster analysis to carry out feature selecting.

As everyone knows, cluster is a kind of typical unsupervised learning method, and it arrives some significant set to image clustering according to picture material; In cluster process, to determine to need clusters number given in advance by artificial usually, this has not only increased user's burden, but also may introduce the interference of human factor to cluster result; In addition, the purpose of image clustering is divided into a plurality of clusters according to certain criterion with image set, make that the image similarity that is positioned at same clustering cluster is big as far as possible, and the image similarity that is positioned at different bunches is as far as possible little, therefore, in order correctly to estimate the cluster effect, thereby carry out feature selecting objectively, it is very important to choose suitable cluster validity index.

The present invention utilizes minimum description length (MDL) criterion to determine clusters number and according to Davies-Bouldin index (the following DB index that all is called for short) validity of cluster is estimated, thereby finds the characteristics of image that can represent the searched targets content.Here said minimum description length criterion is a prior art, but particular content list of references [Horst B, AlesL, Alexander S.MDL principle for robust vector quantisation.Pattern Analysis﹠amp; Applications, 1999,2:59-72, Springer-Verlag London Limited.]; The DB index is an index commonly used when weighing the cluster effect, by the ratio value representation that scatters between distribution and class in the class, the more little expression cluster of ratio effect good more [Davies D.L., Bouldin D.W..A cluster separation measure.1979.IEEETrans.Pattern Anal.Machine Intell.1 (4) .224-227].Consider that remote Sensing Image Retrieval not exclusively is unsupervised, the initial given example of user can be used as weak heuristic information, and characteristics of image should help the difference of this image subblock and other image blocks.Therefore we have carried out certain improvement to existing DB index, thereby more help feature selecting, specific as follows: the interior spread values of class of only calculating the target subclass at example user's image subblock place, and the interior spread values of the class that does not comprise non-target subclass, spread values also includes only spread values between class between non-target subclass and this target subclass between class, and do not comprise spread values between class between the non-target subclass, so not only can give prominence to the importance of target subclass and the difference between target subclass and the non-target subclass, but also can reduce calculated amount.

By above feature selecting, select can represent the characteristics of image of searched targets content after, just can use existing the whole bag of tricks structure respective classified device to carry out image retrieval.

Comprehensive above the analysis, the present invention carries out remote Sensing Image Retrieval as follows:

A kind of remote sensing image retrieval method based on feature selecting and semi-supervised learning, at first select the feature of image to be retrieved, retrieve according to the latent structure respective classified device of selecting then, it is characterized in that: the feature of described selection image to be retrieved is meant: according to MDL criterion and improved DB index, select the optimum color characteristic and the optimum textural characteristics of image to be retrieved by the method for cluster analysis; Specifically realize by following each step:

Step 1) is carried out piecemeal with image to be retrieved;

Step 2) extracts each color characteristic and the textural characteristics of image to be retrieved respectively;

Step 3) is determined clusters number k according to the minimum description length criterion, specifically according to following each step:

Step 31) according to m cluster centre of maximum distance criterion initialization;

Step 32) sets a certain cluster centre C arbitrarily _j, calculate according to following formula The expression hypothesis is with C _jWhen removing, the total variation of code length before and after removing:

${\mathrm{\Δl}}_{C_j}=-L_0-n_j{\log }_2{p}_j+\underset{q=1,q\≠j}{\overset{m}{\mathrm{\Σ}}}{n}_{\mathrm{jq}}{\log }_2\left(\frac{n_q+n_{\mathrm{jq}}}{\left|I\right|}\right)+\underset{x\∈c_j}{\mathrm{\Σ}}\underset{i=1}{\overset{d}{\mathrm{\Σ}}}\frac{{(x_i-c_{\mathrm{iq}})}^2-{(x_i-c_{\mathrm{ij}})}^2}{2\left(\ln 2\right){\mathrm{\σ}}^2}$

Wherein, L ₀The code length at expression clustering cluster center:

$L_0=\sqrt{9\×\mathrm{\σ}\×\underset{j=1}{\overset{m}{\mathrm{\Σ}}}\left({-n}_j{\log }_2{p}_j\right)+\underset{q=1,q\≠j}{\overset{m}{\mathrm{\Σ}}}{n}_{\mathrm{jq}}{\log }_2\left(\frac{n_q+n_{\mathrm{jq}}}{\left|I\right|}\right)+\underset{x\∈c_j}{\mathrm{\Σ}}\underset{i=1}{\overset{d}{\mathrm{\Σ}}}\frac{{(x_i-c_{\mathrm{iq}})}^2-{(x_i-c_{\mathrm{ij}})}^2}{2\left(\ln 2\right){\mathrm{\σ}}^2}/m};$

n _qRepresent that q gathers the number of class sample; n _JqIt is j cluster centre and second neighbour's reference point is the number of samples of q cluster centre that the arest neighbors reference point is satisfied in expression, and the dimension of d representation feature, x are bunch C _jIn the sample member, x _iBe the numerical value of i feature; c _IqThe value of representing the i dimension of q cluster centre, c _IjThe value of representing the i dimension of j cluster centre; | I| represents total number of samples; p _jRepresent C _jGather class sample shared proportion in population sample; σ is the variance of sample data, and span is [0.1,0.2];

Step 33) obtains in the determining step 33

Whether, in this way, then remove cluster centre C less than 0 _jAs not, then keep cluster centre C _j

Step 34) iteration execution in step 32-step 33, up to there not being redundant cluster centre, the cluster centre number that remain this moment is needs definite clusters number k;

The clusters number k that step 4) is determined according to step 3 utilizes the K-means clustering method respectively each feature of extracting in the step 2 to be carried out cluster;

The improved DB index of each feature that step 5) obtains in the calculation procedure 4 respectively according to following formula, and select the textural characteristics of improved DB index minimum in the color characteristic of improved DB index minimum in the color characteristic and the textural characteristics respectively, as optimum color characteristic and optimum textural characteristics:

$S_t=\frac{1}{\left|C_t\right|}\underset{x\∈C_t}{\mathrm{\Σ}}D(x,p_t)$

${\mathrm{DB}}_c=\frac{1}{k-1}\underset{i\≠t}{\underset{i=1}{\overset{k}{\mathrm{\Σ}}}}\frac{1/S_t}{1/D(p_i,p_t)}=\frac{1}{k-1}\underset{i\≠t}{\overset{k}{\underset{i=1}{\mathrm{\Σ}}}}\frac{D(p_i,p_t)}{S_t}$

${\mathrm{DB}}_t=\frac{1}{k-1}\underset{i\≠t}{\overset{k}{\underset{i=1}{\mathrm{\Σ}}}}\frac{S_t}{D(p_i,p_t)}$

Wherein, D () be one apart from operator, for color characteristic, D () expression histogram is handed over distance; And for textural characteristics, D () represents Euclidean distance; T is bunch numbering of target subclass; S _tBe among the target subclass t all samples to the mean distance of cluster centre; | C _t| be the number of samples among the target subclass t; p _tIt is the cluster centre of target subclass t; K represents total clusters number; p _iThe cluster centre of representing non-target subclass; DB _cThe improved DB index of expression color characteristic; DB _tThe improved DB index of expression textural characteristics.

Because the remote sensing images reflection is the physical characteristics that ground covers, and therefore shows existing colouring information in the image, texture information is also arranged, so the present invention selects an optimum color characteristic and an optimum textural characteristics.Certainly, at some in particular cases, lake for example, as long as a feature is just enough, so the present invention also considers the weight of different characteristic, is changed to 0 by the binaryzation method feature weight that weight is low and gets final product.

Consider when the learning method of using relevant feedback commonly used is retrieved, in the process of each feedback, need the user to come the positive counter-example sample of mark, increased user's burden greatly; Simultaneously, in content-based remote Sensing Image Retrieval, usually has only seldom training sample (sometimes even have only a training sample), and the training sample that obtains a large amount of marks is also very difficult, therefore, the preferred semi-supervised learning method of the present invention is carried out image retrieval, particularly coorinated training method (Co-training) and self-training learning method wherein.Below the substance of these two kinds of semi-supervised learning methods is briefly introduced:

The coorinated training method is based on following hypothesis: feature space can be divided into two naturally, and two sorters are trained in these two sub-feature spaces.In the process of coorinated training, each sorter is by adding the training sample set that is enlarged oneself by the sample of the determined high confidence level of another sorter, and iteration successively is up to there not being more unmarked sample;

In the process of self-training study, use initial disaggregated model of flag data sample architecture earlier, remove to estimate the label of unmarked data then with this model, select with suitable selection criterion and correct be labeled data and they are joined in the training set, iteration is up to satisfying certain end condition successively.In the self-training learning process, need to determine that a threshold value Th is used as the condition threshold value that iteration stops, this threshold value can distinguish the non-target subclass the most close with the target subclass, and this threshold value Th sets as follows:

$\mathrm{Th}=\frac{D_1}{D_1+D_2}\×D_{12}$

Wherein, D ₁, D ₂Be respectively the radius of target bunch and the most adjacent non-target bunch, D ₁₂It is the distance between target bunch center and the most adjacent non-target bunch center.When the radius of asking for bunch, sample is to the maximum distance at center in generally selecting for use bunch, but may there be the small amount of noise sample in considering bunch, can adopt the pivot analysis method, the maximum distance that finds in the sample with the distance nearest preceding K% in bunch center as bunch radius, wherein the K value is below 100, can select according to actual needs.

Because the inventive method has been selected optimum color characteristic and optimum textural characteristics, and the weight of these two features may have influence on the retrieval effectiveness of semi-supervised learning method, therefore, the present invention utilizes the improved DB index that obtains in the feature selecting to determine the weight of optimum color characteristic and optimum textural characteristics; Consider simultaneously for color characteristic and textural characteristics, the represented feature difference of DB index of identical numerical value is a non-equivalence in the different codomains, therefore need determine feature weight by non-uniform quantizing, here adopt binarization method to determine feature weight: for color characteristic, when the inverse of the improved DB index of selected optimal characteristics less than threshold value T ₁The time, illustrating that target subclass and the difference of non-target subclass in color space are not clearly, this moment, the weight of color characteristic was made as 0, otherwise was 1; For textural characteristics, when the inverse of the improved DB index of selected optimal characteristics less than threshold value T ₂The time, illustrating that target subclass and the difference of non-target subclass in the textural characteristics space are not clearly, this moment, the weight of textural characteristics was made as 0, otherwise was 1.

According to above analysis, it is as follows just can to draw preferred version of the present invention:

At first,, select the optimum color characteristic and the optimum textural characteristics of image to be retrieved, promptly carry out according to above-mentioned steps 1-step 5 by the method for cluster analysis according to MDL criterion and improved DB index;

Then, select suitable semi-supervised learning method, and utilize the semi-supervised learning method of choosing to carry out image retrieval according to the weight of optimum color characteristic and optimum textural characteristics; Specifically realize by following each step:

Step 6) is calculated the binaryzation weight of optimum color characteristic and optimum textural characteristics respectively according to improved DB index, and concrete grammar is as follows:

For color characteristic, when the inverse of the improved DB index of selected optimal characteristics less than pre-set threshold T ₁The time, illustrating that target subclass and the difference of non-target subclass in color space are not clearly, this moment, the weight of color characteristic was made as 0, otherwise was 1; For textural characteristics, when the inverse of the improved DB index of selected optimal characteristics less than pre-set threshold T ₂The time, illustrating that target subclass and the difference of non-target subclass in the textural characteristics space are not clearly, this moment, the weight of textural characteristics was made as 0, otherwise was 1;

Step 7) is chosen suitable semi-supervised learning method and is retrieved, and is specially: when the binaryzation weight of optimum color characteristic and textural characteristics all is 1, select the coorinated training method to retrieve; And when the weight of a certain feature in optimum color characteristic and the textural characteristics was 0, selecting the self-training method to rely on weight separately was that 1 feature is retrieved.

The present invention utilizes clustering method to select optimum color characteristic and textural characteristics respectively at first according to minimum description length criterion and improved Davies-Bouldin index; Select suitable semi-supervised learning method to carry out the retrieval of remote sensing images according to the color characteristic of optimum and the binaryzation weight of textural characteristics then.The prior art of comparing, the present invention not only can improve retrieval quality greatly, can also effectively reduce calculated amount in the retrieving, improves the speed of retrieval.

Description of drawings

Fig. 1 is the process flow diagram of the specific embodiment of the invention.

Embodiment

Below in conjunction with accompanying drawing technical scheme of the present invention is elaborated:

Use method of the present invention that the different faces of land is covered (land cover) and carried out the retrieval experiment, the retrieval in wherein existing soil erosion zone also has the residential area, forest land, the retrieval of general objectivess such as lake culture in enclosure; The concrete retrieval according to following each step:

Step 1) is carried out piecemeal with image to be retrieved;

In this embodiment,, taked overlapping partition strategy for fear of same target branch being gone among the different fritters, every block size is length=min (128, sample image is long), wide=min (128, sample image is wide), the overlapping 1/2 long 1/2 wide pixel that multiply by between piece and the piece;

Step 2) extracts each color characteristic and the textural characteristics of image to be retrieved respectively;

In this embodiment, HSI color characteristic, Lab color characteristic, Glcm textural characteristics and Gabor textural characteristics have been extracted respectively;

Step 3) is determined clusters number k according to the minimum description length criterion, specifically according to following each step:

Step 31) according to m cluster centre of maximum distance criterion initialization;

Step 32) sets a certain cluster centre C arbitrarily _j, calculate according to following formula

The expression hypothesis is with C _jWhen removing, the total variation of code length before and after removing:

$\mathrm{\Δ}{l}_{C_j}=-L_0-n_j{\log }_2{p}_j+\underset{q=1,q\≠j}{\overset{m}{\mathrm{\Σ}}}{n}_{\mathrm{jq}}{\log }_2\left(\frac{n_q+n_{\mathrm{jq}}}{\left|I\right|}\right)+\underset{x\∈c_j}{\mathrm{\Σ}}\underset{i=1}{\overset{d}{\mathrm{\Σ}}}\frac{{(x_i-c_{\mathrm{iq}})}^2-{(x_i-c_{\mathrm{ij}})}^2}{2\left(\ln 2\right){\mathrm{\σ}}^2}$

Wherein, L ₀The code length at expression clustering cluster center:

$L_0=\sqrt{9\×\mathrm{\σ}\×\underset{j=1}{\overset{m}{\mathrm{\Σ}}}({-n}_j{\log }_2{p}_j+\underset{q=1,q\≠j}{\overset{m}{\mathrm{\Σ}}}{n}_{\mathrm{jq}}{\log }_2\left(\frac{n_q+n_{\mathrm{jq}}}{\left|I\right|}\right)+\underset{x\∈c_j}{\mathrm{\Σ}}\underset{i=1}{\overset{d}{\mathrm{\Σ}}}\frac{{(x_i-c_{\mathrm{iq}})}^2-{(x_i-c_{\mathrm{ij}})}^2}{2\left(\ln 2\right){\mathrm{\σ}}^2})/m};$

n _qRepresent that q gathers the number of class sample; n _JqIt is j cluster centre and second neighbour's reference point is the number of samples of q cluster centre that the arest neighbors reference point is satisfied in expression, and the dimension of d representation feature, x are bunch C _jIn the sample member, x _iBe the numerical value of i feature; c _IqThe value of representing the i dimension of q cluster centre, c _IjThe value of representing the i dimension of j cluster centre; | I| represents total number of samples; p _jRepresent C _jGather class sample shared proportion in population sample; σ is the variance of sample data, and span is [0.1,0.2], and in this embodiment, the value of σ is 0.12;

Step 33) obtains in the determining step 33

Whether, in this way, then remove cluster centre C less than 0 _jAs not, then keep cluster centre C _j

Step 34) iteration execution in step 32-step 33, up to there not being redundant cluster centre, the cluster centre number that remain this moment is needs definite clusters number k;

The clusters number k that step 4) is determined according to step 3 utilizes the K-means clustering method respectively each feature of extracting in the step 2 to be carried out cluster;

The improved DB index of each feature that step 5) obtains in the calculation procedure 4 respectively according to following formula, and select the textural characteristics of improved DB index minimum in the color characteristic of improved DB index minimum in the color characteristic and the textural characteristics respectively, as optimum color characteristic and optimum textural characteristics:

$S_t=\frac{1}{\left|C_t\right|}\underset{x\∈C_t}{\mathrm{\Σ}}D(x,p_t)$

${\mathrm{DB}}_c=\frac{1}{k-1}\underset{i\≠t}{\overset{k}{\underset{i=1}{\mathrm{\Σ}}}}\frac{1/S_t}{1/D(p_i,p_t)}=\frac{1}{k-1}\underset{i\≠t}{\overset{k}{\underset{i=1}{\mathrm{\Σ}}}}\frac{D(p_i,p_t)}{S_t}$

${\mathrm{DB}}_t=\frac{1}{k-1}\underset{i\≠t}{\overset{k}{\underset{i=1}{\mathrm{\Σ}}}}\frac{S_t}{D(p_i,p_t)}$

Wherein, D () be one apart from operator, for color characteristic, D () expression histogram is handed over distance; And for textural characteristics, D () represents Euclidean distance; T is bunch numbering of target subclass; S _tBe among the target subclass t all samples to the mean distance of cluster centre; | C _t| be the number of samples among the target subclass t; p _tIt is the cluster centre of target subclass t; K represents total clusters number; p _iThe cluster centre of representing non-target subclass; DB _cThe improved DB index of expression color characteristic; DB _tThe improved DB index of expression textural characteristics;

Step 6) is calculated the binaryzation weight of optimum color characteristic and optimum textural characteristics respectively according to improved DB index, and concrete grammar is as follows:

For color characteristic, when the inverse of the improved DB index of selected optimal characteristics less than pre-set threshold T ₁The time, illustrating that target subclass and the difference of non-target subclass in color space are not clearly, this moment, the weight of color characteristic was made as 0, otherwise was 1; For textural characteristics, when the inverse of the improved DB index of selected optimal characteristics less than pre-set threshold T ₂The time, illustrating that target subclass and the difference of non-target subclass in the textural characteristics space are not clearly, this moment, the weight of textural characteristics was made as 0, otherwise was 1;

In this embodiment, threshold value T ₁, T ₂Value get 2 and 3 respectively;

Step 7) is chosen suitable semi-supervised learning method and is retrieved, and is specially: when the binaryzation weight of optimum color characteristic and textural characteristics all is 1, select the coorinated training method to retrieve; And when the weight of a certain feature in optimum color characteristic and the textural characteristics was 0, selecting the self-training method to rely on weight separately was that 1 feature is retrieved;

If when selecting the self-training method to retrieve in this step, determine in the cluster process threshold value Th as stopping criterion for iteration according to following formula:

$\mathrm{Th}=\frac{D_1}{D_1+D_2}\×D_{12}$

Wherein, D ₁, D ₂Be respectively the sample farthest that finds in the sample of target bunch and the most adjacent non-target bunch preceding K% that this bunch of middle distance center is nearest and the distance between this bunch center, K " 100; D ₁₂It is the distance between target bunch center and the most adjacent non-target bunch center; In this embodiment, K gets 95.

The inventive method can combine with existing C BIR system fully, thereby realizes the remote Sensing Image Retrieval of robotization.

By the inventive method and existing relevant feedback method are retrieved contrast test, can find that the inventive method is suitable with the relevant feedback method on the index of recall ratio and precision ratio, but the time spent in retrieval is far below relevant feedback method required time, and compare relevant feedback method based on man-machine interaction, the inventive method does not need repeatedly man-machine interaction, has alleviated user's burden.

Claims (5)

1. remote sensing image retrieval method based on feature selecting and semi-supervised learning, at first select the feature of image to be retrieved, retrieve according to the latent structure respective classified device of selecting then, it is characterized in that: the feature of described selection image to be retrieved is meant: according to minimum description length criterion and improved Davies-Bouldin index, select the optimum color characteristic and the optimum textural characteristics of image to be retrieved by the method for cluster analysis; Specifically realize by following each step:

Step 1) is carried out piecemeal with image to be retrieved;

Step 2) extracts each color characteristic and the textural characteristics of image to be retrieved respectively;

Step 3) is determined clusters number k according to the minimum description length criterion, specifically according to following each step:

Step 31) according to m cluster centre of maximum distance criterion initialization;

Step 32) sets a certain cluster centre C arbitrarily _j, calculate according to following formula

The expression hypothesis is with C _jWhen removing, the total variation of code length before and after removing:

$\mathrm{\Δ}{l}_{C_j}=-L_0-n_j{\log }_2{p}_j+\underset{q=1,q\≠j}{\overset{m}{\mathrm{\Σ}}}{n}_{\mathrm{jq}}{\log }_2\left(\frac{n_q+n_{\mathrm{jq}}}{\left|I\right|}\right)+\underset{x\∈c_j}{\mathrm{\Σ}}\underset{i=1}{\overset{d}{\mathrm{\Σ}}}\frac{{(x_i-c_{\mathrm{iq}})}^2-{(x_i-c_{\mathrm{ij}})}^2}{2\left(\ln 2\right){\mathrm{\σ}}^2}$

Wherein, L ₀The code length at expression clustering cluster center:

$L_0=\sqrt{9\×\mathrm{\σ}\×\underset{j=1}{\overset{m}{\mathrm{\Σ}}}(-n_j{\log }_2{p}_j+\underset{q=1,q\≠j}{\overset{m}{\mathrm{\Σ}}}{n}_{\mathrm{jq}}{\log }_2\left(\frac{n_q+n_{\mathrm{jq}}}{\left|I\right|}\right)+\underset{x\∈c_j}{\mathrm{\Σ}}\underset{i=1}{\overset{d}{\mathrm{\Σ}}}\frac{{(x_i-c_{\mathrm{iq}})}^2-{(x_i-c_{\mathrm{ij}})}^2}{2\left(\ln 2\right){\mathrm{\σ}}^2})/m};$

n _qRepresent that q gathers the number of class sample; n _JqIt is j cluster centre and second neighbour's reference point is the number of samples of q cluster centre that the arest neighbors reference point is satisfied in expression, and the dimension of d representation feature, x are bunch C _jIn the sample member, x _iBe the numerical value of i feature; c _IqThe value of representing the i dimension of q cluster centre, c _IjThe value of representing the i dimension of j cluster centre; | I| represents total number of samples; p _jRepresent C _jGather class sample shared proportion in population sample; σ is the variance of sample data, and span is [0.1,0.2];

Step 33) obtains in the determining step 33

Whether, in this way, then remove cluster centre C less than 0 _jAs not, then keep cluster centre C _j

Step 34) iteration execution in step 32-step 33, up to there not being redundant cluster centre, the cluster centre number that remain this moment is needs definite clusters number k;

The clusters number k that step 4) is determined according to step 3 utilizes the K-means clustering method respectively each feature of extracting in the step 2 to be carried out cluster;

The improved Davies-Bouldin index of each feature that step 5) obtains in the calculation procedure 4 respectively according to following formula, and select the textural characteristics of improved Davies-Bouldin index minimum in the color characteristic of improved Davies-Bouldin index minimum in the color characteristic and the textural characteristics respectively, as optimum color characteristic and optimum textural characteristics:

$S_t=\frac{1}{\left|C_t\right|}\underset{x\∈C_t}{\mathrm{\Σ}}D(x,p_t)$

${\mathrm{DB}}_c=\frac{1}{k-1}\underset{\underset{i\≠t}{i=1}}{\overset{k}{\mathrm{\Σ}}}\frac{1/S_t}{1/D(p_i,p_t)}=\frac{1}{k-1}\underset{\underset{i\≠t}{i=1}}{\overset{k}{\mathrm{\Σ}}}\frac{D(p_i,p_t)}{S_t}$

${\mathrm{DB}}_t=\frac{1}{k-1}\underset{\underset{i\≠t}{i=1}}{\overset{k}{\mathrm{\Σ}}}\frac{S_t}{D(p_i,p_t)}$

Wherein, D () be one apart from operator, for color characteristic, D () expression histogram is handed over distance; And for textural characteristics, D () represents Euclidean distance; T is bunch numbering of target subclass; S _tBe among the target subclass t all samples to the mean distance of cluster centre; | C _t| be the number of samples among the target subclass t; p _tIt is the cluster centre of target subclass t; K represents total clusters number; p _iThe cluster centre of representing non-target subclass; DB _cThe improved Davies-Bouldin index of expression color characteristic; DB _tThe improved Davies-Bouldin index of expression textural characteristics.

2. according to claim 1 based on the remote sensing image retrieval method of feature selecting and semi-supervised learning, it is characterized in that: the latent structure respective classified device that described basis is selected is retrieved and is meant: the weight according to optimum color characteristic and optimum textural characteristics is selected suitable semi-supervised learning method, and utilizes the semi-supervised learning method of choosing to carry out image retrieval; Specifically realize by following steps:

Step 6) is calculated the binaryzation weight of optimum color characteristic and optimum textural characteristics respectively according to improved Davies-Bouldin index, and concrete grammar is as follows:

For color characteristic, when the inverse of the improved Davies-Bouldin index of selected optimal characteristics less than pre-set threshold T ₁The time, illustrating that target subclass and the difference of non-target subclass in color space are not clearly, this moment, the weight of color characteristic was made as 0, otherwise was 1; For textural characteristics, when the inverse of the improved Davies-Bouldin index of selected optimal characteristics less than pre-set threshold T ₂The time, illustrating that target subclass and the difference of non-target subclass in the textural characteristics space are not clearly, this moment, the weight of textural characteristics was made as 0, otherwise was 1;

Step 7) is chosen suitable semi-supervised learning method and is retrieved, and is specially: when the binaryzation weight of optimum color characteristic and textural characteristics all is 1, select the coorinated training method to retrieve; And when the weight of a certain feature in optimum color characteristic and the textural characteristics was 0, selecting the self-training method to rely on weight separately was that 1 feature is retrieved.

As described in the claim 2 based on the remote sensing image retrieval method of feature selecting and semi-supervised learning, it is characterized in that: the T of pre-set threshold described in the step 6 ₁, T ₂Value get 2 and 3 respectively.

As described in the claim 2 based on the remote sensing image retrieval method of feature selecting and semi-supervised learning, it is characterized in that: when selecting the self-training method to retrieve in the step 7, determine in the cluster process threshold value Th as stopping criterion for iteration according to following formula:

$\mathrm{Th}=\frac{D_1}{D_1+D_2}\×D_{12}$

Wherein, D ₁, D ₂Be respectively the sample farthest that finds in the sample of target bunch and the most adjacent non-target bunch preceding K% that this bunch of middle distance center is nearest and the distance between this bunch center, K " 100; D ₁₂It is the distance between target bunch center and the most adjacent non-target bunch center.

As described in the claim 4 based on the remote sensing image retrieval method of feature selecting and semi-supervised learning, it is characterized in that: described K value is 95.

Images