CN106127224A - Based on the semi-supervised polarization SAR sorting technique quickly auctioning figure - Google Patents

Abstract

The invention discloses a kind of semi-supervised polarization SAR sorting technique based on quick auction figure.Mainly solve prior art configuration figure matrix complex degree high, the problem that classification accuracy rate is low.The steps include: 1) polarimetric SAR image data is split；2) in figure after singulation, pixel adds spatial information；3) the pixel data acquisition system after carrying out spatial information weighting is asked for similar neighbouring matrix；4) optimize similar neighbouring matrix, obtain auction figure matrix；5) auction figure matrix is carried out sparse process, obtain sparse auction figure matrix；6) utilize sparse auction figure logm to carry out semisupervised classification according to collection, obtain each pixel sorted class label matrix；(7) all of pixel is coloured by the class label utilizing each pixel samples point, the image after output category.The present invention improves classification accuracy rate, decreases the classification time, can be used for the terrain classification to Polarimetric SAR Image and target recognition.

Description

Based on the semi-supervised polarization SAR sorting technique quickly auctioning figure

Technical field

The invention belongs to technical field of image processing, relate to the application in image classification field, a kind of half prison Superintend and direct polarization SAR sorting technique, can be used for the terrain classification to Polarimetric SAR Image and target recognition.

Background technology

Polarization SAR be one be to utilize synthetic aperture principle, it is achieved high-resolution microwave imaging radar, not only there is whole day Time, round-the-clock, resolution is high, can the advantage such as side-looking imaging, SAR image has abundant detailed information, important texture simultaneously Feature and obvious atural object geometry, can be widely applied to the numerous areas such as military affairs, agricultural, navigation, geographical supervision.In the world Remote sensing fields is highly valued, and therefore Classification of Polarimetric SAR Image has become an important research side of polarization SAR information processing To.

The purpose of Classification of Polarimetric SAR Image is to utilize the airborne or polarization measurement data of borne polarization SAR sensor acquisition, According to the character of pixel, determine the classification belonging to each pixel.It is an important content of image interpretation technology, is that other should Basis.In recent years, updating and developing along with polarization SAR sensor, substantial amounts of polarization SAR sorting technique is carried Going out, they are filled with new vitality for the research of radar image classification with application.

Semisupervised classification is by having marker samples to classify a large amount of unmarked samples on a small quantity, and semisupervised classification is for subtracting Few labeled cost, improves Learning machine performance and has the most great practical significance, along with the extensive application of the method, based on figure The method of model the most increasingly receives publicity, and the method builds as summit first by having marker samples and unmarked sample One graph model, gives weights to the limit between each summit, and weights represent the similarity between two samples.It After, by limit, the class label having marker samples is passed to unmarked sample according to majorized function i.e. grader, thus to nothing Marker samples is classified.

Although sorting technique based on figure all achieves higher classification accuracy rate on image is classified, but yet suffers from Two shortcomings: one is in patterning process, ignores the spatial information of image, causing the detail section to image to divide has necessarily Error.Two is that patterning process time complexity is high, and operand is big.

Summary of the invention

Present invention aims to the deficiency of above-mentioned prior art, propose a kind of quick semi-supervised polarization based on figure SAR sorting technique, to reduce the time complexity of composition, and improves the classification accuracy rate of image.

For achieving the above object, technical solution of the present invention includes the following:

(1) from the polarimetric SAR image data file of computer hard disc, obtain coherence property data X of Polarimetric SAR Image, Utilize Pauli decomposition method that coherence property data carry out process and obtain Pauli RGB figure, by super-pixel split-run, this figure is entered Row segmentation；

(2) the Pauli RGB figure after utilizing segmentation carries out spatial information weighting to each pixel:

(2a) Pauli RGB is schemed each pixel x_iM pixel X of surrounding_m={ x_i1,x_i2,x_ij…,x_im} ∈R^d×mRepresent, x_ij∈X_mRepresent pixel x_iJth pixel around, j=1,2 ..., m, m are pixel x_iSurrounding pixel The number of point, x_iIt it is the column vector of d dimension；

(2b) x is used_iM pixel around is to this pixel x_iCarry out spatial information weighting:All of pixel set expression after spatial information weights isN is the number of all pixels in image；

(3) the pixel data acquisition system X after all spatial informations being weighted^*It is divided into L data subset, and to L number Utilize according to subset Nystrom approach method to ask for similar neighborhoods matrix simultaneously, obtain the similar neighborhoods that the l data subset is asked for Matrix isWherein l ∈ 1,2 ... L；U represents A^lApproximation characteristic Vector, Λ is diagonal matrix, B=U Λ U^T, B is the similar neighborhoods matrix in the l data subset between r sample, C Being the similar neighborhoods matrix in the l data subset between r sample and remaining q-r sample, q is the l data Total number of samples in subset；

(4) Game Theory is used to optimize the formula similar neighborhoods matrix A to asking in step (3)^lIt is optimized, is clapped Sell figure matrix S^l,l∈1,2,…L；

(5) to the auction figure matrix S in step (4)^lCarry out sparse process, obtain sparse auction figure matrix W^l；

(6) the sparse auction figure matrix W of the data subset asked in step (5) is utilized^l, L data subset is carried out half Supervised classification merges the classification results of each data subset, obtains each pixel sorted class label matrix Y；

(7) all of pixel is carried out by the class label utilizing each pixel sample point obtained in step (6) Color also calculates classification accuracy rate, the image after output category.

The invention have the advantage that

1, the present invention, by processing original image pixels point data, enters spatial information weighting to it, and then Construct and compare to tradition figure matrix and more can accurately represent the auction sparse graph matrix of similarity between pixel, improve figure The degree of accuracy divided as details.

2, in terms of structural map matrix, the present invention utilizes the method construct figure matrix that Nystrom approaches, and utilizes sample data Similar diagram matrix between similar diagram matrix between concentrated part sampled point, and sampled point and remaining sample point is estimated whole The similar diagram matrix of individual sample data set, compares to tradition patterning process and greatly reduces time of composition, thus reduce right The time of image classification.

Accompanying drawing explanation

The flowchart of Fig. 1 present invention；

The existing method of Fig. 2 and the experimental result picture of the inventive method.

Detailed description of the invention

With reference to Fig. 1, the present invention to be embodied as step as follows:

Step 1, Image semantic classification

(1a) from the polarimetric SAR image data file of computer hard disc, obtain the coherence property data of Polarimetric SAR Image X；

(1b) utilize Pauli decomposition method that the coherence property data of Polarimetric SAR Image are carried out process and obtain Pauli RGB Figure

(1c) by super-pixel split-run, its Pauli RGB figure is divided into 50 pieces, the Polarimetric SAR Image used in this experiment

It is 120 × 80 farmland analogous diagram；

Pauli RGB figure after step 2, utilization segmentation carries out spatial information weighting to each pixel

(2a) Pauli RGB is schemed each pixel x_iM pixel X of surrounding_m={ x_i1,x_i2,x_ij…,x_im} ∈R^d×mRepresent, wherein x_ij∈X_mRepresent pixel x_iJth pixel around, j=1,2 ..., m, d are pixel number evidence Dimension, m is pixel x_iThe number of surrounding pixel point, wherein d=9, m=4；

(2b) x is used_iM pixel around is to this pixel x_iCarry out spatial information weighting:All of pixel set expression after spatial information weights isN=9600 is the number of all pixels in image, and takes the pixel work of 1% at random For training sample, remaining pixel is test sample.

Step 3 constructs similar diagram matrix

(3a) the pixel data acquisition system X after all spatial informations being weighted^*Averagely it is divided into L=4 data subset；

(3b) utilize Nystrom approach method to ask for similar neighborhoods matrix L data subset simultaneously, obtain l number The similar neighborhoods matrix asked for according to subset isWherein l ∈ 1,2 ... L；U represents A^lApproximate characteristic vector, Λ is diagonal matrix, B=U Λ U^T, B is r in the l data subset Similar neighborhoods matrix between=100 samples, C is that in the l data subset, r sample is individual with remaining q-r Similar neighborhoods matrix between sample, q=2400 is total number of samples in the l data subset, whereinWhereinRepresent two sample pointsWithBetween Euclidean distance, take σ=1 during calculating；

Step 4 uses Game Theory to optimize the formula similar neighborhoods matrix A to asking in step (3)^lIt is optimized, obtains Auction figure matrix S^l,l∈1,2,…L。

Game Theory optimizes formula:Wherein P=(p₁,p₂,p_i…,p_q)^TRepresent l number According to the vector of similarity between pixel, p in subset_iFor the i-th element in P, i ∈ 1,2 ... q, π=(π₁,π₂,π_j…, π_q)^TFor A^lSuperior vector, π_iFor the jth element in π,J=1,2 ... q,For A^lIn i-th row jth row Element.

Step 5 is to the auction figure matrix S in step (4)^lCarry out sparse process, obtain sparse auction figure matrix W^l:

W^l=S^lH,

Wherein,It is sparse matrix, wherein a h_ijFor the i-th row jth row in sparse matrix H Element,Front 8 first prime elements of wherein often going in matrix H take 1, and remaining all takes 0.

Step 6 carries out semisupervised classification to data subset.

(6a) the sparse auction figure matrix W of the data subset asked in step (5) is utilized^l, L data subset is carried out half Supervised classification, formula is as follows:

$\underset{Y}{\arg min}\underset{ij}{\Σ}{W}^l\left|\right|y_i^l-y_j^l\left|\right|$

Wherein W^lThe form of expression be

WhereinFor W^lIn i-th row jth row element,It is that in the l data subset, ith pixel orders vegetarian refreshmentsClass Distinguishing label,It it is jth pixel in the l data subsetClass label, work as pixelAnd pixelSimilarity is relatively Height, i.e.The when of being worth the biggest, above-mentioned formula to be made to obtain optimal solution, thenWithDistance want near, takeI.e. i-th Individual pixelWith jth pixelBeing assigned to same class, after classification, the class label matrix of l data subset is

(6b) merge the classification results of 4 data subsets, obtain all pixels sorted class label matrix

Step 7 utilizes the class label of each pixel sample point obtained in step (6) to carry out all of pixel Colour and calculate classification accuracy rate, the image after output category.

The effect of the present invention can be illustrated by emulation experiment:

1. experiment condition

Experiment microcomputer CPU used is Intel Corei5-2430M internal memory 4GB, and programming platform is Matlab R2011b.

Experiment figure is the farmland analogous diagram of 120 × 80, and this figure one has 9600 pixels, seven class crops, takes it In 1% pixel as training sample, remaining is test sample.

2. experiment content

Contrast algorithm is: the semisupervised classification method based on low-rank expression figure published, half prison based on anchor point figure Superintend and direct sorting technique, semisupervised classification method based on auction figure, and the farmland emulation that the method for the present invention is to above-mentioned 120 × 80 Figure carries out semisupervised classification, as in figure 2 it is shown, wherein, Fig. 2 (a) PauliRGB schemes result, Fig. 2 (b) standard results figure, Fig. 2 (c) The classification results figure of the semisupervised classification method of figure is represented based on low-rank, Fig. 2 (d) semisupervised classification based on anchor point figure method Classification results figure, the classification results figure of Fig. 2 (e) semisupervised classification method based on auction figure, dividing of Fig. 2 (f) the inventive method Class result figure.Digital number in Fig. 2 represents the classification of crops

Fig. 2 (c), Fig. 2 (d), Fig. 2 (e), Fig. 2 (f) are compared respectively contrast visible with Fig. 2 (b), dividing of the inventive method Class result has promoted than existing three kinds of methods for the classification results of 7 class crops.To above-mentioned four kinds of methods to agriculture Field analogous diagram carries out time of semisupervised classification and classification accuracy rate is added up, and the results are shown in Table 1

The result that farmland analogous diagram is classified by the existing control methods of table 1 and the inventive method

As it can be seen from table 1 the present invention is higher compared to other method accuracy, and the time of operation is shorter.

Claims (4)

1. a semi-supervised polarization SAR sorting technique based on quickly auction figure, including:

(1) from the polarimetric SAR image data file of computer hard disc, obtain coherence property data X of Polarimetric SAR Image, utilize Pauli decomposition method carries out process and obtains Pauli RGB figure coherence property data, carries out this figure point by super-pixel split-run Cut；

(2) the Pauli RGB figure after utilizing segmentation carries out spatial information weighting to each pixel:

(2a) Pauli RGB is schemed each pixel x_i(x_iBe d dimension column vector) m pixel X of surrounding_m={ x_i1, x_i2,x_ij…,x_im}∈R^d×mRepresent, x_ij∈X_mRepresent pixel x_iJth pixel around, j=1,2 ..., m, m are pixel Point x_iThe number of surrounding pixel point；

(2b) x is used_iM pixel around is to this pixel x_iCarry out spatial information weighting: All of pixel set expression after spatial information weights isN is all in image The number of pixel, and every class pixel takes the pixel of 1% as training sample, and remaining pixel is test sample.

(3) by the pixel data acquisition system X after all addition spatial informations^*Averagely it is divided into L data subset, and to L data Subset utilizes Nystrom approach method to ask for similar neighborhoods matrix simultaneously, obtains the similar neighborhoods square that the l data subset is asked for Battle array isWherein l ∈ 1,2 ... L；U represents A^lApproximation characteristic to Amount, Λ is diagonal matrix, B=U Λ U^T, B is the similar neighborhoods matrix in the l data subset between r sample, and C is Similar neighborhoods matrix between r sample and remaining q-r sample in l data subset, q is the l data Concentrate total number of samples；

(4) Game Theory is used to optimize the formula similar neighborhoods matrix A to asking in step (3)^lIt is optimized, obtains auction figure Matrix S^l,l∈1,2,…L；

(5) to the auction figure matrix S in step (4)^lCarry out sparse process, obtain sparse auction figure matrix W^l；

(6) the sparse auction figure matrix W of the Sub Data Set asked in step (5) is utilized^l, L data subset is carried out semi-supervised point Class also merges the classification results of each data subset, obtains each pixel sorted class label matrix Y；

(7) all of pixel is coloured also by the class label utilizing each pixel sample point obtained in step (6) Calculate classification accuracy rate, the image after output category.

2. according to the semi-supervised polarization SAR data classification method based on quickly auction figure described in claims 1, wherein step (4) Game Theory in optimizes formula:Wherein P=(p₁,p₂,p_i…,p_q)^TRepresent l number According to the vector of similarity between pixel, p in subset_iFor the i-th element in P, i ∈ 1,2 ... q, π=(π₁,π₂,π_j…, π_q)^TFor A^lSuperior vector, π_iFor the jth element in π,J=1,2 ... q,For A^lIn i-th row jth row Element.

3. according to the semi-supervised polarization SAR data classification method based on quickly auction figure described in claims 1, wherein step (5) the sparse auction figure matrix W of l data subset in^l, equation below calculate:

W^l=S^lH,

Wherein,It is sparse matrix, wherein a h_ijFor the unit of the i-th row jth row in sparse matrix H Element, h_ij∈{0,1},Front 8 first prime elements of wherein often going in matrix H take 1, and remaining all takes 0.

4. according to the semi-supervised polarization SAR data classification method based on quickly auction figure described in claims 1, wherein step (6) in, L data subset being carried out semisupervised classification, its formula is as follows:

$\underset{Y}{\mathrm{argmin}}\underset{ij}{\Σ}{W}^l\left|\right|y_i^l-y_j^l\left|\right|$

WhereinFor W^lIn i-th row jth row element,It is that in the l data subset, ith pixel orders vegetarian refreshmentsClassification mark Sign,It it is jth pixel in the l data subsetClass label, work as pixelAnd pixelSimilarity is the highest, I.e.The when of being worth the biggest, above-mentioned formula to be made to obtain optimal solution, thenWithDistance want near, takeI.e. pixelAnd pixelIt is assigned to same class.