CN105005962A - Island-reef remote sensing image registration method based on hierarchical screening strategy - Google Patents

Abstract

The invention relates to an island-reef remote sensing image registration method based on a hierarchical screening strategy, which comprises the steps of matching affine invariant features; establishing a geometric transformation model; carrying out primary screening on matching results of the affine invariant features under constraints of the geometric transformation model; and controlling a screening result by further using the island-reef area overlapping degree on the basis of a primary screening result, ensuring to screen out all wrong feature point pairs, and finally finishing island-reef registration. According to the invention, double difficulties of textural feature deficiency and textural feature instability, which are specifically owned by island-reed remote sensing image matching, are considered, the feature point pairs, which obviously do not conform to spatial distribution, are screened out through establishing the geometric constraint model, then carrying out secondary screening on the feature point pairs by taking the island-reef area overlapping degree as a constraint on the basis of geometric transformation matrix constraint screening, and only correct feature point pairs are reserved so as to complete accurate island-reef registration. The method provided by the invention is high in adaptability, can complete island-reef image registration accurately, and can meet requirements of actual production.

Description

Based on the islands and reefs Remote Sensing Image Matching method of hierarchical screening strategy

Technical field

The present invention relates to a kind of islands and reefs Remote Sensing Image Matching method, particularly a kind of islands and reefs remote sensing image affine invariants method for registering based on hierarchical screening strategy.

Background technology

Image registration is the process of two width obtained under different time, different sensors or different condition or multiple image being carried out mating, and has been widely used in the fields such as remotely-sensed data analysis, computer vision, image procossing.Islands and reefs registration application is in the environmental monitoring of islands and reefs, the aspect such as resource management and disaster alarm.Along with the development in earth observation field, can obtain the islands and reefs remote sensing image under the different image-forming condition of a large amount of the same area, the registration for these islands and reefs images is the basis to islands and reefs change determination and analysis.

The image registration mode of feature based obtains extensive concern.When textural characteristics is sufficient, image registration is relatively easy; When image lacks effective textural characteristics, automatic image registration is a difficulty and has the problem of challenge, in most cases needs the artificial result revising coupling.Local invariant feature, particularly affine invariants, can to a certain extent, overcome the problem that texture lacks because its to occurring in image covering, the mixed and disorderly and content change of background has stronger robustness.MSER (Maximally Stable ExternalRegion) operator is considered to optimum feature extraction operator (Mikolajczyk, 2005).The most outstanding (Mikolajczyk, 2005) that SIFT (ScaleInvariant Feature Transformation) descriptor shows in feature interpretation.In conjunction with both, be referred to as affine invariants coupling, the very sane autoregistration being applied to image.In this technical foundation, select RANSAC (Random SampleConsensus) algorithm to screen feature point pairs further, more excellent matching effect can be obtained.But, for islands and reefs Image Matching, directly carry out affine coupling in conjunction with MSER operator and SIFT descriptor, carry out RANSAC screening further, be often still difficult to obtain satisfied effect.This is that the very big technical difficulty existed due to islands and reefs Remote Sensing Images Matching causes, and roughly can be summarized as two aspects:

1) islands and reefs image lacks textural characteristics.There is the waters that large area lacks textural characteristics in islands and reefs image, and reef platform under water do not have obvious textural characteristics yet, makes textural characteristics mainly concentrate on the grey sand island part of the exposure water surface.Lacking the image of textural characteristics is difficult to provide enough information for Image registration.

2) islands and reefs textural characteristics is unstable.On the one hand, the dynamic change of islands and reefs environment can disturb the characteristic matching of islands and reefs image, as the change of the water level of islands and reefs, the spray etc. of wave strike reef dish formation.On the other hand, islands and reefs affect by artificial exploitation, and because islands and reefs have important value in military strategy and tourist resources, the construction of the heavy construction such as airport and harbour and reconstructing frequently, makes the islands and reefs remote sensing image information difference of different times larger.And, Part Development (as marine reclamation land engineering) makes surface partial shape and area of islands and reefs also occur significantly to change, the textural characteristics that these factors show as image on remote sensing image is unstable, cause the feature point pairs generating a large amount of mistake, the registration of severe jamming image.

Summary of the invention

The technical problem to be solved in the present invention is: overcome prior art shortcoming, proposes a kind of islands and reefs Remote Sensing Image Matching method based on hierarchical screening strategy.

In order to solve above technical matters, the islands and reefs Remote Sensing Image Matching method based on hierarchical screening strategy provided by the invention, comprises the following steps:

The first step, affine invariants mate---and extract affine invariants point to two width islands and reefs remote sensing images, pixel-level generation proper vector, mates according to the unique point of proper vector to two width islands and reefs remote sensing images, obtains some feature point pairs;

Second step, set up geometric transformation model---select affine Transform Model as geometric transformation model, from two width islands and reefs remote sensing images, extract islands and reefs border respectively by natural fracture method, and resolve according to two the islands and reefs boundary informations obtained the model parameter obtaining affine Transform Model;

3rd step, first screening feature point pairs---according to described affine Transform Model, the unique point of in feature point pairs is carried out affined transformation, conversion calculates the Euclidean distance between the unique point of mating with it afterwards, as the error amount of feature point pairs, the feature point pairs accounting for total sample number 8%-12% is chosen from small to large with error amount, error threshold increases gradually, continuous calculating error values be less than the unique point belonging to main image in the feature point pairs of this error threshold form the spatial cohesion of point set, the error threshold corresponding when spatial cohesion reaches maximum is as screening threshold value, screen out the feature point pairs that error amount is greater than this screening threshold value,

4th step, postsearch screening feature point pairs---using islands and reefs area Duplication as constraint condition, use RANSAC iterative algorithm to screen feature point pairs;

5th step, islands and reefs Remote Sensing Image Matching---after the screening of the 3rd step and the 4th step, the feature point pairs according to retaining resolves homography matrix, completes the registration of islands and reefs remote sensing image with the homography matrix resolving acquisition.

In order to solve above technical matters, the present invention also has following further feature:

1, in the described first step, the feature point pairs quantity of generation is not less than 4.。

2, in described second step, affine transformation matrix model obtains according to following formula:

$H=\left[\begin{array}{ccc}{x}_0& m_{1}cos\mathrm{\α}& m_{1}sin\mathrm{\α}\\ y_0& -m_{2}sin\mathrm{\α}& m_2\cos \mathrm{\α}\end{array}\right]$

Above parameter can be resolved according to the islands and reefs boundary information of two width images, and wherein, α is rotation angle, m ₁and m ₂for scale parameter, x ₀and y ₀for translation parameters.

3, in the first step, use MSER operator to extract affine invariants point to two width islands and reefs remote sensing images, then with SIFT descriptor, affine invariants point is described, pixel-level generation 128 dimensional feature vector.

4, in the 3rd step, front n error amount minimal characteristic point centering belong to the unique point of main image form the spatial cohesion SDQ of point set _nobtain according to following formula:

${\mathrm{SDQ}}_n=\sqrt{\frac{\underset{i=1}{\overset{n}{\Σ}}{(x_i-{\stackrel{\‾}{x}}_{wmc})}^2+\underset{i=1}{\overset{n}{\Σ}}{(y_i-{\stackrel{\‾}{y}}_{wmc})}^2}{n}}$

Wherein, the weighted center point coordinate that before representing, n error amount minimal characteristic point is right, (x _i, y _i) be the coordinate that i-th error minimal characteristic point centering belongs to the unique point of main image.

5, in described 4th step, using islands and reefs area Duplication as constraint condition, the concrete grammar using RANSAC iterative algorithm to screen feature point pairs is as follows:

A, start first time RANSAC iteration using RANSAC initial threshold as screening parameter, the value of RANSAC initial threshold is 0.01;

B, RANSAC threshold value is as screening parameter, RANSAC algorithm is utilized feature point pairs to be divided into intra-office point to drawn game exterior point pair, screen out point not in the know right, according to intra-office point to solving homography matrix, and utilize homography matrix to convert main image, calculate the area Duplication on island in the island and image subject to registration converting rear image;

C, progressively increase RANSAC threshold value, and repeat step b, when islands and reefs area Duplication obviously declines, stop RANSAC iteration;

A RANSAC threshold value before d, islands and reefs area Duplication obviously decline is best RANSAC threshold value, screens out the point not in the know be separated out according to this best RANSAC threshold value right.

6, in described 4th step, second time iteration and RANSAC threshold value afterwards obtain according to following formula:

Wherein, TR _currentfor the RANSAC threshold value of current iteration, TR _previousfor the RANSAC threshold value of last iteration.

7, in described 4th step, islands and reefs area Duplication Ratio obtains according to following formula:

$Ratio=\frac{R_{\mathrm{\μ}a}\∩R_{\left(H^T\mathrm{\μ}bH\right)}}{R_{\mathrm{\μ}a}\∪R_{\left(H^T\mathrm{\μ}bH\right)}}\×100\%$

Wherein, R _{μ a}for reference territory, islands and reefs face area, for conversion territory, islands and reefs face area.

Beneficial effect of the present invention is as follows:

1), the present invention is directed to the feature that islands and reefs remote sensing image textural characteristics is rare and textural characteristics is unstable, on the basis of affine invariants coupling, propose twice screening strategy from " space distribution control " to " control of islands and reefs degree of overlapping ", thus remove all error characteristics point pair, complete the registration of islands and reefs remote sensing image.

2), the coarse sizing method proposed by setting up geometric transformation model discrimination feature point pairs of the present invention, and determine to screen threshold value in conjunction with feature point pairs number and spatial cohesion, on the whole the Space Consistency of feature point pairs is verified, eliminate the error characteristic point pair that space distribution is inconsistent, improve the correctness of subsequent match.

3), the present invention is directed to the select permeability that RANSAC screens threshold value, propose and control automatically to select optimum screening threshold method according to territory, islands and reefs face Duplication, guarantee the reliability of the selection result, ensure the precision of final islands and reefs remote sensing image.

4), the inventive method strong adaptability, be proven, use the inventive method can registration islands and reefs remote sensing image accurately, the needs of actual production can be met, compare traditional Remote Sensing Image Matching method and there is stronger practicality.

Accompanying drawing explanation

Below in conjunction with accompanying drawing, the present invention is further illustrated.

Fig. 1 is method for registering process flow diagram of the present invention.

Fig. 2 is test block bitmap area and raw video figure in the embodiment of the present invention.

Fig. 3-a is China Nan Zi island remote sensing image in 2005.

Fig. 3-b is China Nan Zi island remote sensing image in 2009.

Fig. 3-c is China Nan Zi island image islands and reefs Boundary Extraction result in 2005.

Fig. 3-d is China Nan Zi island image islands and reefs Boundary Extraction result in 2009.

Fig. 3-e is that China Nan Zi island image affine invariants point in 2005 extracts result.

Fig. 3-f is that China Nan Zi island image affine invariants point in 2009 extracts result.

Fig. 3-g is China Nan Zi island image affine invariants matching result.

Fig. 3-h is the result after China Nan Zi island image screens for the first time.

Fig. 3-i is that to carry out RANSAC threshold value be result after the postsearch screening of 0.6 to China Nan Zi island image.

Fig. 3-j is that to carry out RANSAC threshold value be result after the postsearch screening of 0.7 to China Nan Zi island image.

Fig. 4 is that islands and reefs overlapping area ratio is with RANSAC changes of threshold figure.

Embodiment

Elaborate the present invention with reference to the accompanying drawings below, make technology path of the present invention and operation steps more clear.

As shown in Figure 2, embodiment selects the Nan Zi island being positioned at China South Sea as study area, and longitude is east longitude 114 degree 20 points, and latitude is north latitude 11 degree 26 points, area about 0.13 square kilometre.This study area belongs to Chinese Hainan Province Sansha City, and island is oblong, height above sea level average out to 3.9 meters.

Experimental data is the 2 scape high-resolution remote sensing images covering this region, the part that the first scape (Fig. 2 lower left side and Fig. 3-a) is QuickBird remote sensing image panchromatic wave-band, and spatial resolution is 0.6 meter, images on January 27th, 2005; The part that second scape (on the right side of Fig. 2 bottom and Fig. 3-b) is WorldView-2 remote sensing image panchromatic wave-band, spatial resolution is 0.5 meter, and imaging time is on Dec 31st, 2009.

The present invention is based on the islands and reefs Remote Sensing Image Matching method (process flow diagram is shown in Fig. 1) of hierarchical screening strategy, comprise the following steps:

The first step, affine invariants mate---and use MSER operator to extract affine invariants point to two width islands and reefs remote sensing images, then with SIFT descriptor, affine invariants point is described, pixel-level generation 128 dimensional feature vector, mate according to the unique point of proper vector to two width islands and reefs remote sensing images, obtain some feature point pairs.

First ENVI software is used to cut out islands and reefs and surrounding body from original islands and reefs remote sensing image, then the affine invariants point (this step is by OpenCV storehouse C++ programming realization) of MSER operator extraction islands and reefs image is used, but involved affine invariants point is not limited to the acquisition of this algorithm in the present invention, in the present embodiment, two islands and reefs images generate 1562 and 2980 unique points respectively, and Fig. 3-e is that China Nan Zi island image affine invariants point in 2005 extracts result.Fig. 3-f is that China Nan Zi island image affine invariants point in 2009 extracts result.Then be described each unique point with SIFT descriptor, generating 128 is proper vector (this step is by OpenCV storehouse C++ programming realization), but in the present invention, involved proper vector is not limited to the acquisition of this algorithm.Calculate the Euclidean distance of the proper vector on image and all proper vectors on another image.When most adjacency is less than 0.85 with the ratio of time adjacency, think that this proper vector of two matches, generate a pair feature point pairs, the present embodiment symbiosis becomes 439 feature point pairs.Fig. 3-g is China Nan Zi island image affine invariants matching result.

Second step, set up geometric transformation model---select affine Transform Model as geometric transformation model, from two width islands and reefs remote sensing images, extract islands and reefs border respectively by natural fracture method, and resolve according to two the islands and reefs boundary informations obtained the model parameter obtaining affine Transform Model.

In this step, first from the islands and reefs remote sensing image of source, islands and reefs reef platform is extracted by natural fracture method, then ENVI software is used to carry out closing operation of mathematical morphology filtering, re-use ArcGIS software and carry out raster data vector quantization, obtain islands and reefs border, Fig. 3-c is China Nan Zi island image islands and reefs Boundary Extraction result in 2005, and Fig. 3-d is China Nan Zi island image islands and reefs Boundary Extraction result in 2009.

Then, progressively resolve affine invariants matrix according to two islands and reefs boundary informations, this step is overall by OpenCV storehouse C++ programming realization, and detailed step is as follows:

A1), computing center's point.Center point coordinate obtain according to following formula:

$\stackrel{\‾}{x}=\frac{1}{n}\underset{i=1}{\overset{n}{\Σ}}{x}_i,\stackrel{\‾}{y}=\frac{1}{n}\underset{i=1}{\overset{n}{\Σ}}{y}_i$

In the present embodiment, two islands and reefs border central point planimetric coordinatess are respectively (1168.60,1121.22) and (1138.82,1135.80).

A2), principal direction is determined.First determined the line beam equation of central point (x, y), line beam equation obtains according to following formula:

$xtan\mathrm{\θ}-y+\stackrel{\‾}{y}-\stackrel{\‾}{x}tan\mathrm{\θ}=0$

The intrafascicular all straight lines of calculated line and all frontier points hang down the quadratic sum of distance again, and when quadratic sum is minimum, this straight slope is border principal direction.Principal direction slope θ obtains according to following formula:

$P=\underset{i=1}{\overset{n}{\Σ}}{\[(x_i-\stackrel{\‾}{x})sin\mathrm{\θ}-(y_i-\stackrel{\‾}{y})cos\mathrm{\θ}\]}^2$

In the present embodiment, two islands and reefs border principal direction slopes are respectively 41.531273 and 40.528446.

A3), rotation angle is calculated.Calculate rotation angle parameter α in affine matrix, subtracted each other by principal direction slope and can obtain α value.In the present embodiment, two islands and reefs border principal direction rotation angle parameter are-1.002827.

A4), scale parameter is calculated.First brought in line beam equation by principal direction slope and can obtain two straight lines, using straight line and islands and reefs border transversal as major axis, in the present embodiment, major axis is respectively 1044.132202 and 1043.324707.And then definition is secondary direction with principal direction vertical direction, power is upwards crossed central point straight line and islands and reefs border transversal as minor axis, in the present embodiment, minor axis is respectively 527.977173 and 557.712952, last according to the ratio of major axis and the ratio of minor axis between two islands and reefs borders, try to achieve scale parameter in affine matrix.In the present embodiment, the scale parameter of major axis and minor axis is respectively 0.999226635 and 1.056320198.

A5), translation parameters is calculated.According to the rotation angle of having tried to achieve and scale parameter, after being brought into by central point, translation parameters obtains according to following formula:

$H=\left[\begin{array}{ccc}{x}_0& m_{1}cos\mathrm{\α}& m_{1}sin\mathrm{\α}\\ y_0& -m_{2}sin\mathrm{\α}& m_2\cos \mathrm{\α}\end{array}\right]$

In the present embodiment, on primary and secondary direction, translation parameters is respectively-7.970558821 and-27.94712967.

A6), affine transformation matrix is tried to achieve.Finally can obtain affine transformation matrix, in the present embodiment, affine transformation matrix is as follows:

$H=\left[\begin{array}{ccc}0.999073587& -0.018487441& -7.970558821\\ 0.017488204& 1.056158405& -27.94712967\end{array}\right].$

3rd step, first screening feature point pairs---according to described affine Transform Model, the unique point of in feature point pairs is carried out affined transformation, conversion calculates the Euclidean distance between the unique point of mating with it afterwards, as the error amount of feature point pairs, the feature point pairs accounting for total sample number 8%-12% is chosen from small to large with error amount, error threshold increases gradually, continuous calculating error values be less than the unique point belonging to main image in the feature point pairs of this error threshold form the spatial cohesion of point set, the error threshold corresponding when spatial cohesion reaches maximum is as screening threshold value, screen out the feature point pairs that error amount is greater than this screening threshold value.

In this step, the unique point coordinate in feature point pairs is brought in the affine transformation matrix in second step, try to achieve the Euclidean distance of transform characteristics point coordinate and unique point coordinate of the same name, as the error amount of feature point pairs.Error amount according to feature point pairs sorts to feature point pairs, the spatial cohesion SDQ of front n error minimal characteristic point (main image) _nobtain according to following formula:

${\mathrm{SDQ}}_n=\sqrt{\frac{\underset{i=1}{\overset{n}{\Σ}}{(x_i-{\stackrel{\‾}{x}}_{wmc})}^2+\underset{i=1}{\overset{n}{\Σ}}{(y_i-{\stackrel{\‾}{y}}_{wmc})}^2}{n}}$

Wherein, the weighted center point coordinate of n error minimal characteristic point before representing, (x _i, y _i) be the coordinate of i-th error minimal characteristic point.Keeping characteristics point is to the feature point pairs of the 8%-12% number of total sample number, and corresponding to the feature point pairs selecting spatial cohesion maximum, error amount is as screening threshold value, with this threshold value screening feature point pairs.Screening threshold value in the present embodiment is 100, and after screening, feature point pairs number is 35.Fig. 3-h is the result after China Nan Zi island image screens for the first time.

4th step, postsearch screening feature point pairs---using islands and reefs area Duplication as constraint condition, use RANSAC iterative algorithm to screen feature point pairs.

In this step, utilize islands and reefs area degree of overlapping as constraint, automatically choose optimum RANSAC and screen threshold value.First, choose minimum RANSAC threshold value, the selection result not keeping characteristics point to or only retain the correct feature point pairs of sub-fraction, resolve initial homography matrix, to calculate in image territory, islands and reefs face after homography matrix conversion with the overlapping area ratio in territory, islands and reefs face in another image; Then progressively increase threshold value, retain more proper characteristics point pair, when islands and reefs face overlapping area has significant change, program can stop.RANSAC threshold value is now appropriate threshold.Based on this threshold value, the screening of all feature point pairs is intra-office point drawn game exterior point by RANSAC, ensure that intra-office point is correct feature point pairs, can registration islands and reefs image accurately according to these feature point pairs.This step is overall by OpenCV storehouse C++ programming realization, and concrete steps are as follows:

B1), using RANSAC initial threshold as screening parameter start first time RANSAC iteration, in this example, the value of RANSAC initial threshold is 0.01;

B2), RANSAC threshold value as screening parameter, utilizing RANSAC algorithm feature point pairs to be divided into intra-office point to drawn game exterior point pair, screening out point not in the know right, according to intra-office point to solving homography matrix;

B3) utilize the homography matrix solved to convert main image, calculate the area Duplication on island in the island and image subject to registration converting rear image;

Islands and reefs area Duplication Ratio obtains according to following formula:

$Ratio=\frac{R_{\mathrm{\μ}a}\∩R_{\left(H^T\mathrm{\μ}bH\right)}}{R_{\mathrm{\μ}a}\∪R_{\left(H^T\mathrm{\μ}bH\right)}}\×100\%$

Wherein, R _{μ a}for reference territory, islands and reefs face area, for conversion territory, islands and reefs face area;

B4), progressively increase RANSAC threshold value, and repeat step b2, b3, when islands and reefs area Duplication obviously declines, stop RANSAC iteration;

Second time iteration and RANSAC threshold value afterwards obtain according to following formula:

Wherein, TR _currentfor the RANSAC threshold value of current iteration, TR _previousfor the RANSAC threshold value of last iteration;

B5), islands and reefs area Duplication obviously decline before a RANSAC threshold value be best RANSAC threshold value, screen out the point not in the know be separated out according to this best RANSAC threshold value right.

If Fig. 4 is that islands and reefs overlapping area ratio is with RANSAC changes of threshold figure, islands and reefs overlapping area signal when the upper left corner represents that RANSAC threshold value is 0.6 in figure, corresponding island area Duplication is 94.52%, islands and reefs overlapping area signal when being 0.7 that the lower left corner is RANSAC threshold value, obviously, when RANSAC threshold value increases to 0.7, islands and reefs area Duplication drops near 86%, decline fairly obvious, so 0.6 is best RANSAC threshold value.Fig. 3-i is that to carry out RANSAC threshold value be result after the postsearch screening of 0.6 to China Nan Zi island image,

5th step, islands and reefs Remote Sensing Image Matching---after twice screening, the feature point pairs according to retaining resolves homography matrix, completes the registration of islands and reefs remote sensing image with homography matrix.Fig. 3-j is that to carry out RANSAC threshold value be result after the postsearch screening of 0.7 to China Nan Zi island image.

After experience above-mentioned steps, the present embodiment finally retains 10 feature point pairs, and visually distinguishes that 10 feature point pairs are proper characteristics point pair, resolves homography matrix, complete the accuracy registration of China Nan Zi island remote sensing image with homography matrix according to these 10 feature point pairs.

In addition to the implementation, the present invention can also have other embodiments.All employings are equal to the technical scheme of replacement or equivalent transformation formation, all drop on the protection domain of application claims.

Claims (8)

1., based on an islands and reefs Remote Sensing Image Matching method for hierarchical screening strategy, comprise the following steps:

The first step, affine invariants mate---and extract affine invariants point to two width islands and reefs remote sensing images, pixel-level generation proper vector, mates according to the unique point of proper vector to two width islands and reefs remote sensing images, obtains some feature point pairs;

Second step, set up geometric transformation model---select affine Transform Model as geometric transformation model, from two width islands and reefs remote sensing images, extract islands and reefs border respectively by natural fracture method, and resolve according to two the islands and reefs boundary informations obtained the model parameter obtaining affine Transform Model;

3rd step, first screening feature point pairs---according to described affine Transform Model, the unique point of in feature point pairs is carried out affined transformation, conversion calculates the Euclidean distance between the unique point of mating with it afterwards, as the error amount of feature point pairs, the feature point pairs accounting for total sample number 8%-12% is chosen from small to large with error amount, error threshold increases gradually, continuous calculating error values be less than the unique point belonging to main image in the feature point pairs of this error threshold form the spatial cohesion of point set, the error threshold corresponding when spatial cohesion reaches maximum is as screening threshold value, screen out the feature point pairs that error amount is greater than this screening threshold value,

4th step, postsearch screening feature point pairs---using islands and reefs area Duplication as constraint condition, use RANSAC iterative algorithm to screen feature point pairs;

5th step, islands and reefs Remote Sensing Image Matching---after the screening of the 3rd step and the 4th step, the feature point pairs according to retaining resolves homography matrix, completes the registration of islands and reefs remote sensing image with the homography matrix resolving acquisition.

2. the islands and reefs Remote Sensing Image Matching method based on hierarchical screening strategy according to claim 1, it is characterized in that: in the described first step, the feature point pairs quantity of generation is not less than 4.

3. the islands and reefs Remote Sensing Image Matching method based on hierarchical screening strategy according to claim 1, it is characterized in that: in the first step, MSER operator is used to extract affine invariants point to two width islands and reefs remote sensing images, then with SIFT descriptor, affine invariants point is described, pixel-level generation 128 dimensional feature vector.

4. the islands and reefs Remote Sensing Image Matching method based on hierarchical screening strategy according to claim 1, it is characterized in that: in described second step, affine transformation matrix model obtains according to following formula:

Above parameter can be resolved according to the islands and reefs boundary information of two width images, and wherein, α is rotation angle, m ₁and m ₂for scale parameter, x ₀and y ₀for translation parameters.How to extract above parameter according to islands and reefs boundary information.

5. the islands and reefs Remote Sensing Image Matching method based on hierarchical screening strategy according to claim 1, is characterized in that: in the 3rd step, front n error amount minimal characteristic point centering belong to the unique point of main image form the spatial cohesion SDQ of point set _nobtain according to following formula:

Wherein, the weighted center point coordinate that before representing, n error amount minimal characteristic point is right, (x _i, y _i) be the coordinate that i-th error minimal characteristic point centering belongs to the unique point of main image.

6. the islands and reefs Remote Sensing Image Matching method based on hierarchical screening strategy according to claim 1, it is characterized in that: in described 4th step, using islands and reefs area Duplication as constraint condition, the concrete grammar using RANSAC iterative algorithm to screen feature point pairs is as follows:

A, start first time RANSAC iteration using RANSAC initial threshold as screening parameter, the value of RANSAC initial threshold is 0.01;

B, RANSAC threshold value is as screening parameter, RANSAC algorithm is utilized feature point pairs to be divided into intra-office point to drawn game exterior point pair, screen out point not in the know right, according to intra-office point to solving homography matrix, and utilize homography matrix to convert main image, calculate the area Duplication on island in the island and image subject to registration converting rear image;

C, progressively increase RANSAC threshold value, and repeat step b, when islands and reefs area Duplication obviously declines, stop RANSAC iteration;

A RANSAC threshold value before d, islands and reefs area Duplication obviously decline is best RANSAC threshold value, screens out the point not in the know be separated out according to this best RANSAC threshold value right.

7. the islands and reefs Remote Sensing Image Matching method based on hierarchical screening strategy according to claim 1, is characterized in that: in described 4th step, and second time iteration and RANSAC threshold value afterwards obtain according to following formula:

Wherein, TR _currentfor the RANSAC threshold value of current iteration, TR _previousfor the RANSAC threshold value of last iteration.

8. the islands and reefs Remote Sensing Image Matching method based on hierarchical screening strategy according to claim 6, is characterized in that: in described 4th step, islands and reefs area Duplication Ratio obtains according to following formula:

Wherein, R _{μ a}for reference territory, islands and reefs face area, for conversion territory, islands and reefs face area.