CN102902978A - Object-oriented high-resolution remote-sensing image classification method - Google Patents

Abstract

The invention provides an object-oriented high-resolution remote-sensing image classification method. The method comprises the steps of S1, conducting segmentation processing on images to be processed to obtained a plurality of subimage objects; S2, obtaining feature information of subimage objects; and S3, classifying subimage objects according to the obtained feature information, wherein images to be processed are high-resolution remote-sensing images, the feature information of subimage objects comprises spectral information, shape information and texture information of subimage objects. According to the method, on the basis of object-oriented classification, a classification method combining probabilistic latent semantic analysis and a support vector machine is introduced, the problem that 'the same features with different classifications' and 'the same classifications with different features' are not high in identification ratio in the prior art is solved, the classification precision of high-resolution remote-sensing images is greatly improved, advantages of latent semantic analysis (LSA) and advantages of probabilistic latent semantic analysis (PLSA) are combined, and the problems of overfitting and local optimum which are caused by random initialization are effectively solved.

Description

A kind of OO high-resolution remote sensing image sorting technique

Technical field

The present invention relates to field of remote sensing image processing, relate in particular to a kind of OO high-resolution remote sensing image sorting technique.

Background technology

Can find out from the evolution of Remote Image Classification, sorting technique mainly is divided into three levels: (1) based on the sorting technique of pixel, a lot of traditional Classifying Method in Remote Sensing Image belong to this level, and this technology is quite ripe; (2) based on the sorting technique of object or primitive, be the sorting technique of a kind of higher level of growing up of recent two decades, OO sorting technique just belongs to this level; (3) based on the knowledge classification method, be a new development trend of Remote Image Classification, its theory also is in the discussion stage, and it is not very extensive using.

The wave band that high-resolution remote sensing image contains is less, spectral information is abundant not as spatial information, if still adopting traditional method based on pixel classifies, namely only utilize the spectral information of pixel, the abundant atural object spatial information that does not utilize high-resolution remote sensing image to provide, such as shape information, texture information and contextual information etc., can cause nicety of grading low, the data resource waste.The proposition of OO sorting technique, can be with processing unit---object, corresponding with entity in the real world, for the attributive character of utilizing entity when analyzing provides may.OO sorting technique is utilized the internal characteristics (spectrum, shape and texture) of object simultaneously, and topological characteristic and contextual feature have improved the precision of high-definition remote sensing Data classification to a great extent.

Based on the sorting technique of statistical model in remote sensing image classification in occupation of critical role, mainly comprise supervision and unsupervised classification.Supervised classification method commonly used has the methods such as maximum likelihood classification, minimum distance classification, Support Vector Machine, neural network.Unsupervised classification has the methods such as K-mean cluster, Fuzzy C-Means Clustering and IsoData.Each target is described by a stack features in these traditional target identification methods, namely use " feature-target " this simple statement model, when the target of identification " other with the characteristic foreign peoples " or " generic different characteristic ", can not effectively improve the accuracy rate of identification.

Summary of the invention

The technical matters that the present invention mainly solves provides a kind of OO high-resolution remote sensing image sorting technique, in order to solve the problem that can not effectively improve the accuracy rate of identification in the prior art when identifying the target of " other with the characteristic foreign peoples " or " generic different characteristic ".

For addressing the above problem, a kind of technical scheme that the present invention adopts is: a kind of OO high-resolution remote sensing image sorting technique is provided, comprises:

S1, pending Image Segmentation Using is processed, obtained a plurality of subimage objects;

S2, obtain the characteristic information of described subimage object;

The characteristic information that S3, basis get access to is classified to described subimage object;

Wherein, described pending image is high-resolution remote sensing image, and the characteristic information of described subimage object comprises spectral information, shape information and the texture information of subimage object.

Wherein, described S1 comprises:

S11, described pending image mixed open and close re-establishing filter and process, obtain filtering image, and utilize the Sobel operator to calculate the gradient image of described filtering image;

S12, obtain the gradient image that calculates through above-mentioned, and described gradient image is carried out watershed transform, obtain the initial segmentation image, set up the Region adjacency graph of described initial segmentation image;

S13, described initial segmentation image is carried out that the zone merge to be processed until cost function hour stops, obtain final segmentation result, generate a plurality of subimage objects.

Wherein, described Region adjacency graph generates by the following method: for the every number of sub images in the described initial segmentation image is given a numbering, and set up the Region adjacency graph of initial segmentation image according to information such as the border length of described subimage, area, adjacent sub-images numberings.

Wherein, among the described S13 described initial segmentation image carried out that the zone merge to be processed until cost function value hour stops to comprise:

Choosing the subimage in the described initial segmentation image is current object, merges with current object and with the object of described current object similarity maximum, upgrades the data of Region adjacency graph, obtains merging rear split image;

Judge whether cost function value is minimum value; If judge described cost function value minimum value, then stop to merge; Be not minimum value if judge described cost function value, then continue two objects of similarity maximum in the split image after the described merging are merged the data of processing and upgrading Region adjacency graph.

The calculating formula of similarity of wherein, described initial segmentation image being carried out adopting when the zone merges processing is:

$\mathrm{\δ}(i,j)=\frac{\left|\right|R_i\left|\right|\·\left|\right|R_j\left|\right|}{\left|\right|R_i\left|\right|+\left|\right|R_j\left|\right|}{[\mathrm{\μ}\left(R_i\right)-\mathrm{\μ}\left(R_j\right)]}^2\·I(i,j)$

Wherein, R _i, R _jRepresent regional i and regional j; || R _i||, || R _j|| represent respectively number of pixels among regional i and the regional j; μ (R _i) and μ (R _j) then represent respectively the spectrum average of regional i and regional j; If R _iAnd R _jAdjacent, I (i, j)=1 then, otherwise I (i, j)=∞.

Wherein, described S3 is specially:

S31, a plurality of subimage objects that will obtain are divided into training set and test set, respectively feature---the co-occurrence frequency matrix of object of calculation training collection and test set;

The PLSA model parameter of S32, use latent semantic analysis LSA initialization training set, then utilize the PLSA model of EM algorithm match training set, obtain characteristic information at the distribution probability P (w|z) on the potential classification and potential classification the distribution probability P (z|o on the training set object _Train);

Distribution probability P (w|z) on potential classification is constant for S33, maintenance characteristic information obtained above, use the PLSA model parameter of latent semantic analysis LSA initialization test collection, utilize the PLSA model of EM algorithm match test set, obtain the distribution probability P (z|o of potential classification on the test set object _Test);

S34, use SVM are to the distribution probability P (z|o of potential classification on the training set object _Train) differentiate study, generate the SVM discrimination model, and to the distribution probability P (z|o of potential classification on the test set object _Test) test, determine the classification of tested object, and with the different atural objects of different colours mark, realize the classification to described subimage object.

Wherein, the computing formula of LSA initialization PLSA model parameter is as follows:

$P\left(w_j\right|z_k)=e^{{(t_j\·v_k)}^2}/\underset{k=1}{\overset{K}{\mathrm{\Σ}}}{e}^{{(t_j\·v_k)}^2}$

$P\left(d_i\right|z_k)=e^{{(x_i\·u_k)}^2}/\underset{k=1}{\overset{K}{\mathrm{\Σ}}}{e}^{{(x_j\·u_k)}^2}$

Wherein, P (w _j| z _k) representation feature w _jAt potential classification z _kOn distribution probability, P (d _i| z _k) indicated object d _iAt potential classification z _kOn distribution probability, t _jThe proper vector in the feature object co-occurrence frequency matrix, x _iThe object vector in the feature object co-occurrence frequency matrix, v _kThe right proper vector after the co-occurrence matrix svd, u _kBe the left eigenvector after the co-occurrence matrix svd, K is the number of potential classification.

Wherein, the computing formula of described cost function is:

$J_l=\frac{\underset{r=0}{\overset{K_l}{\mathrm{\Σ}}}{J}_w^r}{J_t}=\frac{\underset{r=0}{\overset{K_l}{\mathrm{\Σ}}}\underset{i=0}{\overset{N_r}{\mathrm{\Σ}}}{[x_i^r-{\mathrm{\μ}}^r]}^2}{\underset{i=0}{\overset{N}{\mathrm{\Σ}}}{[x_i-\mathrm{\μ}]}^2}$

Wherein, K _lExpression remaining area quantity,

The spectrum divergence that represents regional r, J _tThe spectrum divergence that represents whole image-region, N _rRepresent the number of pixels that regional r contains, Represent i grey scale pixel value among the regional r, μ ^rThe spectrum average that represents regional r, total pixel count in the N presentation video, x _iI grey scale pixel value in the presentation video, the spectrum average of μ presentation video.

The invention has the beneficial effects as follows: be different from prior art and when the target of identification " other with the characteristic foreign peoples " or " generic different characteristic ", can not effectively improve the accuracy rate of identification, the invention provides a kind of OO high-resolution remote sensing image sorting technique, on the basis of object-oriented classification, introduced the sorting technique of probability latent semantic analysis and Support Vector Machine combination, solved to a certain extent " other with the feature foreign peoples " and " generic different feature " problem that discrimination is not high in the prior art, the nicety of grading of high-resolution remote sensing image is greatly improved, and combine the advantage of LSA and PLSA, effectively solved over-fitting and local optimal problem that random initializtion causes.

Description of drawings

Fig. 1 is the process flow diagram of OO high-resolution remote sensing image sorting technique in the embodiment;

Fig. 2 is the process flow diagram of OO high-resolution remote sensing image sorting technique in another embodiment;

Fig. 3 is the process flow diagram of image segmentation in the above-mentioned specific embodiment;

Fig. 4 is the object-oriented classification process figure of PLSA and SVM combination.

Embodiment

By describing technology contents of the present invention, structural attitude in detail, being realized purpose and effect, below in conjunction with embodiment and cooperate accompanying drawing to give in detail explanation.

See also Fig. 1, present embodiment provides a kind of OO high-resolution remote sensing image sorting technique, comprising:

S1, pending Image Segmentation Using is processed, obtained a plurality of subimage objects;

S2, obtain the characteristic information of described subimage object;

The characteristic information that S3, basis get access to is classified to described subimage object;

Wherein, described pending image is high-resolution remote sensing image, and the characteristic information of described subimage object comprises spectral information, shape information and the texture information of subimage object, and wherein spectral information comprises each band spectrum average, brightness; Shape information comprises area, border length, shape index, degree of compacting; Texture information comprises contrast, entropy, moment of inertia.

See also Fig. 2 and Fig. 3, in certain embodiments, described S1 comprises S11, S12 and S13, and literary composition specific as follows is described.

S11, described pending image mixed open and close re-establishing filter and process, obtain filtering image, and utilize the Sobel operator to calculate the gradient image of described filtering image.Concrete, the present invention at first opens the reconstruction computing to original image, splits reconstructed image again and closes the reconstruction computing, obtains filtering image.Morphological reconstruction by opening and close that to rebuild operation definition as follows:

$\left\{\begin{array}{c}{O}^R\left(f\right)=R(O(f,B),f)\\ C^R\left(f\right)=R(C(f,B),f)\end{array}\right.$

Wherein, f represents original image, and B represents structural element, and O represents opening operation, and C represents closed operation, O ^R(f) reconstruction, C are opened in expression ^R(f) reconstruction is closed in expression.Morphology mixing Opening and closing by reconstruction is defined as and opens first the secondary reconstruction of closing afterwards, that is: M ^R(f)=R (C (O ^R(f), B), O ^R(f)).Then utilize the gradient image of Sobel operator calculation of filtered image.

S12, obtain the gradient image that calculates through above-mentioned, and described gradient image is carried out watershed transform, obtain the initial segmentation image, set up the Region adjacency graph of described initial segmentation image.In specific embodiment, gradient image is implemented watershed transform, comprise following three steps:

(1) each pixel of grid scanning gradient image G (I), find out all minimum points: to (x, y) ∈ G (I), definition (x ', y ') be its neighborhood territory pixel, x '={ x-1, x, x+1} arranged when adopting 8 neighborhood; Y '={ y-1, y, y+1}.If exist arbitrarily (x ', y ') ∈ N (x, y), so that G (x ', y ')＜G (x, y), then G (x, y) is labeled as non-zone minimum (NARM), and puts into fifo queue Q.When the Q non-NULL, head of the queue element G (x ', y ') falls out, scan its field pixel (x ", y ") ∈ N (x ', y '), if G (x ", y ") is labeled as sky, and G (x, y)=G (x ", y "), then with G (x ", y ") is labeled as NARM and is put into formation Q tail end.

(2) mark minimum point and its field pixel, and the field pixel that will be labeled as sky is put into fifo queue Q, the neighborhood territory pixel that is labeled as NARM is put into ordered queue OQ: each pixel of grid scanning gradient image G (I), if G is (x, y) be labeled as sky, G (x, y) ∈ CB then _i(i=1 begins), and G (x, y) put into fifo queue Q.When the Q non-NULL, head of the queue element G (x ', y ') dequeue, scan its field pixel (x ", y ") ∈ N (x ', y '), if G (x ", y ") is labeled as sky, then G (x ", y ") ∈ CB _i, and with G (x ", y ") puts into formation Q; If G (x ", y ") is labeled as NARM, then G (x ", y ") ∈ CB _i, and with G (x ", y ") puts into a gray scale ordered queue OQ.

(3) the field pixel of pixel among the mark ordered queue OQ: when the OQ non-NULL, head of the queue element G (x, y) dequeue, G (x, y) ∈ CB _k(k=1,2 ..., i), scan its field pixel (x ', y ') ∈ N (x, y), if G (x ', y ') be labeled as NARM, G (x ', y ') CB then _k

Through behind the watershed transform, each zone of initial segmentation image has been endowed a numbering.Each zone is considered as a summit, then connects two corresponding summits if two zones are adjacent, set up the Region adjacency graph of initial segmentation image according to information such as the border length of described subimage, area, adjacent sub-images numberings.

In the art, common sorting technique is that each pixel to image carries out feature calculation, then by sample learning and sorter training, is that each pixel is distributed category label, thereby obtains classification results.Consider the large scale characteristic of remote sensing images, the calculated amount that this process need consumption is very large.If therefore first image is divided into several zones, and then is classified in the zone, calculated amount will reduce greatly so.The present invention just is being based on such thinking, adopts OO classificating thought, at first Image Segmentation Using is obtained some homogeneous regions, then extracts provincial characteristics, again to each region allocation category label, sets up the Region adjacency graph of initial segmentation image.On OO basis, utilize simultaneously internal characteristics (spectrum, shape and texture), topological characteristic and the contextual feature of object, can improve to a great extent the precision of high-definition remote sensing Data classification.

S13, described initial segmentation image is carried out that the zone merge to be processed until cost function hour stops, obtain final segmentation result, generate a plurality of subimage objects.

In the above-described embodiment, described initial segmentation image is carried out that the zone merge to be processed until cost function value hour stops to comprise:

Choosing the subimage in the described initial segmentation image is current object, merges with current object and with the object of described current object similarity maximum, upgrades the data of Region adjacency graph, obtains merging rear split image;

Judge whether cost function value is minimum value; If judge described cost function value minimum value, then stop to merge; Be not minimum value if judge described cost function value, then continue two objects of similarity maximum in the split image after the described merging are merged the data of processing and upgrading Region adjacency graph.

In the above-described embodiment, regional similarity measurement directly affects order and the number of times that merges as the criterion of zone merging, and the calculating formula of similarity of employing was when described initial segmentation image was carried out zone merging processing:

$\mathrm{\δ}(i,j)=\frac{\left|\right|R_i\left|\right|\·\left|\right|R_j\left|\right|}{\left|\right|R_i\left|\right|+\left|\right|R_j\left|\right|}{[\mathrm{\μ}\left(R_i\right)-\mathrm{\μ}\left(R_j\right)]}^2\·I(i,j)$

Wherein, R _i, R _jRepresent regional i and regional j; || R _i||, || R _j|| represent respectively number of pixels in regional i and the j territory, district; μ (R _i) and μ (R _j) then represent respectively the spectrum average of regional i and regional j; If R _iAnd R _jAdjacent, I (i, j)=1 then, otherwise I (i, j)=∞.

The criterion of in the present invention stop area merging is another the crucial problem in the region merging algorithm, hour stop when the cost functional value in the above-described embodiment, namely when spectrum divergence and remaining area ratio and hour stop area merge, wherein, remaining area is than referring to that the remaining area number accounts for the ratio that merges the forefoot area sum.The computing formula of described cost function is:

$J_l=\frac{\underset{r=0}{\overset{K_l}{\mathrm{\Σ}}}{J}_w^r}{J_t}=\frac{\underset{r=0}{\overset{K_l}{\mathrm{\Σ}}}\underset{i=0}{\overset{N_r}{\mathrm{\Σ}}}{[x_i^r-{\mathrm{\μ}}^r]}^2}{\underset{i=0}{\overset{N}{\mathrm{\Σ}}}{[x_i-\mathrm{\μ}]}^2}$

Wherein, K _lExpression remaining area quantity,

The spectrum divergence that represents regional r, J _tThe spectrum divergence that represents whole image-region, N _rRepresent the number of pixels that regional r contains,

Represent i grey scale pixel value among the regional r, μ ^rThe spectrum average that represents regional r, total pixel count in the N presentation video, x _iI grey scale pixel value in the presentation video, the spectrum average of μ presentation video.

See also Fig. 4, described S3 comprises S31, S32, S33 and S34.Specific as follows described.

S31, a plurality of subimage objects that will obtain are divided into training set and test set, respectively feature---the co-occurrence frequency matrix of object of calculation training collection and test set.The co-occurrence frequency entry of a matrix element of feature---object has reflected the number of times that certain eigenwert occurs in certain object, its computation process is as follows:

Every kind of feature utilizing the K-means algorithm that step 2 is extracted is carried out separately cluster, and all cluster centre values are as this feature characteristic of correspondence collection V.Then the feature s that extracts is quantified as some concrete numerical value v among the character pair collection V by the nearest neighbouring rule of following formula _i:

Wherein, dist (x, y) is the Euclidean distance between x and the y, N _VSize for word set V.

After feature carried out quantification, just can be according to the v that obtains _iCalculated characteristics frequency matrix H (s):

H(s)=(h ^l(s)) _l=1,…,L， $h^l\left(s\right)=(h_1^l\left(s\right),...,h_{N_V^l}^l\left(s\right))$

Wherein, as Q (s)=v _iThe time, h _iOtherwise be 0 (s)=1; The species number of L representation feature,

The size that represents gained word set V behind the l kind feature clustering.

The PLSA model parameter of S32, use latent semantic analysis LSA initialization training set, then utilize the PLSA model of EM algorithm match training set, obtain characteristic information at the distribution probability P (w|z) on the potential classification and potential classification the distribution probability P (z|o on the training set object _Train).

Distribution probability P (w|z) on potential classification is constant for S33, maintenance characteristic information obtained above, use the PLSA model parameter of latent semantic analysis LSA initialization test collection, utilize the PLSA model of EM algorithm match test set, obtain the distribution probability P (z|o of potential classification on the test set object _Test).

S34, use SVM are to the distribution probability P (z|o of potential classification on the training set object _Train) differentiate study, generate the SVM discrimination model, and to the distribution probability P (z|o of potential classification on the test set object _Test) test, determine the classification of tested object, and with the different atural objects of different colours mark, realize the classification to described subimage object.

Concrete, the computing formula of LSA initialization PLSA model parameter is as follows:

$P\left(w_j\right|z_k)=e^{{(t_j\·v_k)}^2}/\underset{k=1}{\overset{K}{\mathrm{\Σ}}}{e}^{{(t_j\·v_k)}^2}$

$P\left(d_i\right|z_k)=e^{{(x_i\·u_k)}^2}/\underset{k=1}{\overset{K}{\mathrm{\Σ}}}{e}^{{(x_j\·u_k)}^2}$

Wherein, P (w _j| z _k) representation feature w _jAt potential classification z _kOn distribution probability, P (d _i| z _k) indicated object d _iAt potential classification z _kOn distribution probability, t _jThe proper vector in the feature object co-occurrence frequency matrix, x _iThe object vector in the feature object co-occurrence frequency matrix, v _kThe right proper vector after the co-occurrence matrix svd, u _kBe the left eigenvector after the co-occurrence matrix svd, K is the number of potential classification.

Be not difficult to find out by above-mentioned, the present invention has been incorporated herein the problem that can not effectively improve the accuracy rate of identification when sorting technique that the probability latent semantic analysis is combined with Support Vector Machine and latent semantic analysis (LSA) initialization PLSA model parameter technological means solve the target of the identification " other with the characteristic foreign peoples " mentioned in the background technology or " generic different characteristic ".

In the art, probability latent semantic analysis (PLSA) is introduced a potential semantic layer between " feature-target " statement model, strengthened the statement ability to target.In the above-described embodiment, add up each eigenwert at the condition distribution probability P (w|z) on the potential classification of difference and the condition distribution probability P (z|o) of each object on different potential classifications, then with the distribution probability P (z|o) of object on the potential classification of difference as test data, use Support Vector Machine SVM to classify.SVM is as a kind of new Data Classification Technology, solved to a great extent the problem that other classification exist, such as non-linear, Model Selection, mistake study, multidimensional and local minimum problems, the more important thing is, SVM can not cause Hughes's phenomenon, for limited training sample, the increase of intrinsic dimensionality can not reduce nicety of grading.Experiment showed, that this method can more effectively distinguish the atural object of " other with the characteristic foreign peoples " and " generic different characteristic ", improve nicety of grading.

And maximum (EM) algorithm random initializtion of the expectation that the maximal possibility estimation in the PLSA algorithm adopts PLSA parameter can produce the problem of over-fitting and local optimum, makes classification results unstable.Present known LSA is consistent in form with PLSA, and corresponding parameter exists certain contact, so this method is utilized LSA initialization PLSA parameter, then the parameter randomization that first LSA is obtained is used for parameter corresponding to initialization PLSA model.Improved PLSA algorithm uses the LSA dimensionality reduction at initial phase, and the potential semantic relation between can discovery object early combines the advantage of LSA and PLSA, has effectively solved over-fitting and local optimal problem that random initializtion causes.

In sum, the invention provides a kind of OO high-resolution remote sensing image sorting technique, on the basis of object-oriented classification, introduced the sorting technique of probability latent semantic analysis and Support Vector Machine combination, solved to a certain extent " other with the feature foreign peoples " and " generic different feature " problem that discrimination is not high in the prior art, the nicety of grading of high-resolution remote sensing image is greatly improved, and combine the advantage of LSA and PLSA, effectively solved over-fitting and local optimal problem that random initializtion causes.

The above only is embodiments of the invention; be not so limit claim of the present invention; every equivalent structure or equivalent flow process conversion that utilizes instructions of the present invention and accompanying drawing content to do; or directly or indirectly be used in other relevant technical fields, all in like manner be included in the scope of patent protection of the present invention.

Claims (8)

1. an OO high-resolution remote sensing image sorting technique is characterized in that, comprising:

S1, pending Image Segmentation Using is processed, obtained a plurality of subimage objects;

S2, obtain the characteristic information of described subimage object;

The characteristic information that S3, basis get access to is classified to described subimage object;

Wherein, described pending image is high-resolution remote sensing image, and the characteristic information of described subimage object comprises spectral information, shape information and the texture information of subimage object.

2. OO high-resolution remote sensing image sorting technique according to claim 1 is characterized in that, described S1 comprises:

S11, described pending image mixed open and close re-establishing filter and process, obtain filtering image, and utilize the Sobel operator to calculate the gradient image of described filtering image;

S12, obtain the gradient image that calculates through above-mentioned, and described gradient image is carried out watershed transform, obtain the initial segmentation image, set up the Region adjacency graph of described initial segmentation image;

S13, described initial segmentation image is carried out that the zone merge to be processed until cost function hour stops, obtain final segmentation result, generate a plurality of subimage objects.

3. OO high-resolution remote sensing image sorting technique according to claim 2, it is characterized in that, described Region adjacency graph generates by the following method: for the every number of sub images in the described initial segmentation image is given a numbering, and set up the Region adjacency graph of initial segmentation image according to information such as the border length of described subimage, area, adjacent sub-images numberings.

4. OO high-resolution remote sensing image sorting technique according to claim 3 is characterized in that, among the described S13 described initial segmentation image is carried out that the zone merge to be processed until cost function value hour stops to comprise:

Choosing the subimage in the described initial segmentation image is current object, merges with current object and with the object of described current object similarity maximum, upgrades the data of Region adjacency graph, obtains merging rear split image;

Judge whether cost function value is minimum value; If judge described cost function value minimum value, then stop to merge; Be not minimum value if judge described cost function value, then continue two objects of similarity maximum in the split image after the described merging are merged the data of processing and upgrading Region adjacency graph.

5. OO high-resolution remote sensing image sorting technique according to claim 4 is characterized in that, the calculating formula of similarity that described initial segmentation image is carried out adopting when the zone merges processing is:

$\mathrm{\δ}(i,j)=\frac{\left|\right|R_i\left|\right|\·\left|\right|R_j\left|\right|}{\left|\right|R_i\left|\right|+\left|\right|R_j\left|\right|}{[\mathrm{\μ}\left(R_i\right)-\mathrm{\μ}\left(R_j\right)]}^2\·I(i,j)$

Wherein, R _i, R _jRepresent regional i and regional j; || R _i||, || R _j|| represent respectively number of pixels among regional i and the regional j; μ (R _i) and μ (R _j) then represent respectively the spectrum average of regional i and regional j; If R _iAnd R _jAdjacent, I (i, j)=1 then, otherwise I (i, j)=∞.

6. OO high-resolution remote sensing image sorting technique according to claim 1 is characterized in that, described S3 is specially:

S31, a plurality of subimage objects that will obtain are divided into training set and test set, respectively feature---the co-occurrence frequency matrix of object of calculation training collection and test set;

The PLSA model parameter of S32, use latent semantic analysis LSA initialization training set, then utilize the PLSA model of EM algorithm match training set, obtain characteristic information at the distribution probability P (w|z) on the potential classification and potential classification the distribution probability P (z|o on the training set object _Train);

Distribution probability P (w|z) on potential classification is constant for S33, maintenance characteristic information obtained above, use the PLSA model parameter of latent semantic analysis LSA initialization test collection, utilize the PLSA model of EM algorithm match test set, obtain the distribution probability P (z|o of potential classification on the test set object _Test);

S34, use SVM are to the distribution probability P (z|o of potential classification on the training set object _Train) differentiate study, generate the SVM discrimination model, and to the distribution probability P (z|o of potential classification on the test set object _Test) test, determine the classification of tested object, and with the different atural objects of different colours mark, realize the classification to described subimage object.

7. OO high-resolution remote sensing image sorting technique according to claim 6 is characterized in that, the computing formula of LSA initialization PLSA model parameter is as follows:

$P\left(w_j\right|z_k)=e^{{(t_j\·v_k)}^2}/\underset{k=1}{\overset{K}{\mathrm{\Σ}}}{e}^{{(t_j\·v_k)}^2}$

$P\left(d_i\right|z_k)=e^{{(x_i\·u_k)}^2}/\underset{k=1}{\overset{K}{\mathrm{\Σ}}}{e}^{{(x_j\·u_k)}^2}$

Wherein, P (w _j| z _k) representation feature w _jAt potential classification z _kOn distribution probability, P (d _i| z _k) indicated object d _iAt potential classification z _kOn distribution probability, t _jThe proper vector in the feature object co-occurrence frequency matrix, x _iThe object vector in the feature object co-occurrence frequency matrix, v _kThe right proper vector after the co-occurrence matrix svd, u _kBe the left eigenvector after the co-occurrence matrix svd, K is the number of potential classification.

8. according to claim 2 to 5 each described OO high-resolution remote sensing image sorting techniques, it is characterized in that, the computing formula of described cost function is:

$J_l=\frac{\underset{r=0}{\overset{K_l}{\mathrm{\Σ}}}{J}_w^r}{J_t}=\frac{\underset{r=0}{\overset{K_l}{\mathrm{\Σ}}}\underset{i=0}{\overset{N_r}{\mathrm{\Σ}}}{[x_i^r-{\mathrm{\μ}}^r]}^2}{\underset{i=0}{\overset{N}{\mathrm{\Σ}}}{[x_i-\mathrm{\μ}]}^2}$

Wherein, K _lExpression remaining area quantity,

The spectrum divergence that represents regional r, J _tThe spectrum divergence that represents whole image-region, N _rRepresent the number of pixels that regional r contains,

Represent i grey scale pixel value of regional r, μ ^rThe spectrum average that represents regional r, total pixel count in the N presentation video, x _iI grey scale pixel value in the presentation video, the spectrum average of μ presentation video.

Images