CN103177444A - Automatic SAR (synthetic-aperture radar) image rectification method - Google Patents

Abstract

The invention provides an automatic SAR (synthetic-aperture radar) image rectification method. In the method, the problems encountered when feature points are selected by SIFT (scale-invariant feature transform) rectifying points under the condition of SAR image complex landforms are analyzed, the feature points are combined with the Delaunay triangulation network and coded, and triangle similarity of coded feature points are judged through automatic threshold values according to the Delaunay triangularization principle. Since SIFT rectification and the Delaunay triangulation network have robustness to geometric changes such as rotation, wrong rectifying points rectified by the SIFT can be filtered by the SIFT-Delaunay encryption algorithm according to the similarity among triangles, and accurate rectification can be achieved.

Description

A kind of SAR automatic image registration method

Technical field

The invention belongs to the signal processing technology field, especially relate to a kind of SAR automatic image registration method.

Background technology

SAR image registration is two width of the Same Scene that will obtain under different periods, different visual angles or different sensors or the process that multiple image superposes, namely by seeking a kind of spatial alternation, make the corresponding point that represent same target in two width images reach consistent on the locus.In change detection, use widely in a plurality of fields such as fusing image data and target identification due to image registration techniques, is one of basic task of image processing, thereby obtained research widely.

Image registration generally has two kinds of methods: methods based on domain and based on the image registration of feature.Wherein its principle of methods based on domain is to complete registration by the similarity that compares pixel grey scale in two width images.These class methods do not need image is carried out pre-service, the image registration accuracy less to size, that grey scale change is little is high, have living space correlation method, not displacement method and spectrum correlation method etc., these methods are all to utilize the half-tone information of image, registration accuracy is high, but responsive for image change, complexity is high, and is more responsive to the target speed ratio.For visual angle, gray scale, structural change all for larger remote sensing images registration accuracy relatively poor, therefore be not suitable for SAR image registration.Thereby at present generally all be based on the method for feature for the registration of SAR image.

Method based on feature does not directly operate grayscale image, but by extracting some common traits on reference picture and image subject to registration, set up the corresponding relation between feature, and then realize the registration of image, commonly used have a feature, a contour feature, edge feature and major component feature etc., method based on feature is not very sensitive to image change and image rotation, but registration is subjected to the impact of similar features, and unique point easily goes wrong.

SAR automatic image registration problem is one of gordian technique of SAR image interpretation, is well solved for a long time.Yardstick invariant features conversion SIFT algorithm is a kind of more successful autoregistration algorithm, but due to the complicacy of scene, the SIFT algorithm can not solve the problem of similar features diverse location.

Summary of the invention

Technical matters to be solved by this invention is to overcome the deficiencies in the prior art, the present invention proposes a kind of SAR automatic image registration method.Described method is based on Delaunay triangle topological theory, a kind of autoregistration algorithm of SIFT-Delaunay coding has been proposed, test by measured data, the present invention can solve the problem in SIFT algorithm registration in SAR image affined transformation, realizes the SAR automatic image registration.

For solving the problems of the technologies described above, the technical solution adopted in the present invention is:

A kind of SAR automatic image registration method, described method step is as follows:

Steps A, the SIFT characteristic matching:

Adopt the SIFT algorithm to extract the local feature of SAR image, seek extreme point, extracting position, yardstick, rotational invariants parameter at metric space; Two images subject to registration are extracted respectively the SIFT feature, and compare one by one, the unique point that satisfies the coupling requirement in two width images is kept, what do not meet the demands deletes;

Step B, the SIFT-Delaunay coding

SIFT coupling for two width images in steps A is right, adopts the coupling of SIFT-Delaunay coding eliminating error right;

Step B-1 to the SIFT unique point in every width image, adopts Delaunay triangulation network Incremental insertion method, sets up the Delaunay triangulation network;

Step B-2 to each triangle of the Delaunay triangulation network, utilizes the corresponding relation on three summits in triangle, sets up inner affine Transform Model, picks out the unique point that does not satisfy proportionate relationship;

Step B-3, to each diabolo, more corresponding limit and angle one by one judge whether to meet the following conditions:

|L1-L′1|<T1,|L2-L′2|<T1,|L3-L′3|<T1

|A1-A′1|<T2,|A2-A′2|<T2,|A3-A′3|<T2

T1 wherein, T2 be according to all corresponding sides and angle set threshold value, L1, L2, L3 are respectively benchmark image leg-of-mutton three limits, and L ' 1, L ' 2, L ' 3 is leg-of-mutton three limits of image subject to registration, A1, A2, A3 is leg-of-mutton three angles of benchmark image, A ' 1, and A ' 2, and A ' 3 is leg-of-mutton three angles of image subject to registration;

If do not satisfy condition, continue to search three summits corresponding other triangles respectively, if all triangles that point connects do not satisfy respective conditions, delete this point;

If satisfy condition, keep these leg-of-mutton three summits; Keep the corresponding point that all satisfy condition, at last with these reference point as registration, two width images are carried out registration.

In described steps A, the concrete steps of SIFT characteristic matching are as follows:

Steps A-1, the generation of SIFT proper vector

Steps A-1-1 utilizes Gaussian difference pyrene and image convolution to obtain difference of Gaussian DOG image, judges position and the yardstick of extreme point by the gray-scale value that compares pixel in the DOG image; Utilize the unsettled unique point of stability metric deletion, determine feature point set;

Steps A-1-2 utilizes the gradient direction distribution feature of unique point neighborhood territory pixel, for each unique point assigned direction parameter, makes operator have the yardstick unchangeability, determines unique point principal direction;

Steps A-1-3, coordinate axis is rotated to unique point principal direction, get the window of 4 * 4 m multiple size centered by unique point, m is positive integer, calculates the gradient orientation histogram of 8 directions in each image blockage of 4 * 4, draw the accumulated value of each gradient direction, form a Seed Points, 8 directions of each Seed Points form 8 * m dimension SIFT proper vector, with the length normalization method of proper vector, obtain the SIFT proper vector;

Steps A-2, the coupling of proper vector

Adopt Euclidean distance as the similarity measurement of two width image SIFT unique points, the preset proportion threshold value, it is right to obtain mating by similarity measurement.

the invention has the beneficial effects as follows: the present invention proposes a kind of SAR automatic image registration method, the described methods analyst SIFT registration point problem that the unique point selection runs under SAR image MODEL OVER COMPLEX TOPOGRAPHY, by being combined coding with the Delaunay triangulation network, utilize Delaunay trigonometric ratio principle, judge coding characteristic point triangle similarity by automatic threshold, because SIFT registration and the Delaunay triangulation network all have robustness to Geometrical changes such as rotations, therefore the SIFT-Delaunay encryption algorithm can effectively utilize similarity between triangle to the misregistration point filtering of SIFT registration, reach the purpose of accurate registration.

Description of drawings

Fig. 1 is the Delaunay triangulation network with misregistration point; Fig. 1 a is the triangulation network of benchmark image, and Fig. 1 b is the triangulation network of image subject to registration.

Fig. 2 is a diabolo corresponding in the Delaunay triangulation network; Fig. 2 a is a triangle of benchmark image, and Fig. 2 b is a triangle of image subject to registration.

Fig. 3 is the Delaunay triangulation network that misregistration point has been eliminated; Fig. 3 a is the triangulation network of benchmark image, and Fig. 3 b is the triangulation network of registered images.

Fig. 4 is benchmark SAR image.

Fig. 5 is SAR image subject to registration.

Fig. 6 is images after registration.

Embodiment

Below in conjunction with accompanying drawing, a kind of SAR automatic image registration method that the present invention is proposed is elaborated:

A kind of SAR automatic image registration method, step is as follows:

Steps A, the SIFT characteristic matching

The SIFT Method And Principle is a kind of method of extracting image local feature, utilizes multi-scale method that the grey scale change of neighborhood is described as the vector with certain invariance, with this vector point as the SIFT feature;

Steps A-1, the determining of the detection of unique point and exact position thereof

At first utilize Gaussian difference pyrene and image convolution to obtain difference of Gaussian DoG image, determine position and the yardstick of extreme point by the gray-scale value that compares pixel in the DoG image; Then utilize the unsettled unique point of stability metric deletion finally to determine feature point set;

Difference of Gaussian DoG is defined as

D(x,y,σ)=[G(x,y,k)-G(x,y,σ)]*I(x,y)=L(x,y,kσ)-L(x,y,σ)

Wherein:

L(x,y,σ)=G(x,y,σ)*I(x,y)，

I (x, y) is the gaussian kernel under the different scale factor

G(x,y,σ)=exp[-(x ²+y ²)/2σ ²]/2πσ ²，

(x, y) is volume coordinate, and k, σ are the yardstick coordinates;

The DOG pyramid subtracts each other by adjacent metric space function in gaussian pyramid and obtains, and the scale factor of the pyramidal ground floor of DOG is consistent with the ground floor of gaussian pyramid;

In order the extreme point in DOG space to be detected, need to each pixel in the DOG metric space with 9 neighbor pixels in his adjacent 8 pixels and levels altogether 26 pixels compare, to guarantee Local Extremum can be detected at metric space and two-dimensional space;

Because the DOG value is more responsive to noise and edge, therefore, also need further checking could accurately orientate unique point as by the Local Extremum that obtains with upper type, accurately be defined as position and the yardstick of unique point by the three-dimensional quadratic function of match, simultaneously can remove the low extreme point of contrast and unsettled edge respective point, to strengthen accuracy and the stability of coupling, improve antijamming capability.

Steps A-2, the determining of unique point direction

Utilize the gradient direction distribution feature of unique point neighborhood territory pixel, be each unique point assigned direction parameter, make operator have the yardstick unchangeability:

$m(x,y)=\sqrt{{(L(x+1,y)-(x-1,y))}^2+{((x,y+1)-(x,y-1))}^2}$

θ(x,y)=tan2(L(x+1,y)-(x-1,y))/((x,y+1)-(x,y-1))

Wherein, m (x, y) represents Grad, and θ (x, y) represents gradient direction; L yardstick used is each unique point yardstick at place separately, in actual computation, samples in the neighborhood window centered by unique point, and adds up the gradient direction of neighborhood territory pixel with gradient orientation histogram; The scope of histogram of gradients is 0 to 360 degree, wherein post of every 10 degree, 36 posts altogether; The direction of peak value in histogram of gradients when having the peak value of another suitable main peak value 80%, can be regarded as this direction the auxiliary direction of this unique point; A unique point can be designated as multiple directions, a principal direction, and more than one auxiliary direction can strengthen the robustness of program like this;

Steps A-3, the generation of proper vector

Coordinate axis is rotated to unique point principal direction, keeping its rotational invariance, then centered by unique point to get 8 * 8 window as example, a pixel of every little lattice representative feature vertex neighborhood place metric space wherein, the direction of arrow represents the gradient direction of this pixel, and length represents the gradient-norm value; Calculate the gradient orientation histogram of 8 directions in each image blockage of 4 * 4, draw the accumulated value of each gradient direction, form a Seed Points; Formed 2 * 2 Seed Points, 8 directions of each Seed Points can generate 32 data altogether, form one 32 dimension SIFT proper vector; This mode has strengthened the antijamming capability of algorithm, to locating unique point devious, fault-tolerance is preferably arranged also; This moment, the SIFT proper vector of generation has been removed the impact of the geometry deformations such as change of scale and rotation, then with the length normalization method of proper vector, just can further remove the impact of illumination variation;

Steps A-4, the coupling of proper vector

After the SIFT proper vector of two width images to be matched generates, must carry out similarity measurement, generally adopt the distance metric function as instrument, can obtain coupling potential between image by similarity measurement; Here adopt Euclidean distance as the similarity measurement between two width images, after obtaining the SIFT proper vector, adopt preferential k-d tree approximation BBF searching algorithm, carry out first search and search 2 approximate KNN unique points of each unique point, in these two unique points, if nearest distance less than certain proportion threshold value, is accepted this a pair of match point except near in proper order distance; Reduce this proportion threshold value, SIFT match point number can reduce, but more stable; Then eliminate mispairing, when obtaining potential coupling by similarity measurement, need to improve robustness according to how much restrictions and other additional constraint eliminating errors coupling, so just utilize the SIFT Feature Correspondence Algorithm to realize images match;

In the SAR imaging process, adjustment due to flight path, can cause the SAR image of different flight numbers to have the different problem in pitch angle, also there is simultaneously the problem that similar scene is arranged in complex scene, namely the matching degree of the registration point of diverse location is very high, and this can exist with regard to the registration that makes similarity measurement the accurate problem that mismatches;

For the problems referred to above, we introduce Delaunay triangle topological theory, only have under the condition of affined transformation at two width images, utilize the Delaunay triangulation to carry out the impact point coupling;

Step B, the SIFT-Delaunay coding

Corresponding one by one by the reference mark that top step obtains, but the complicacy due to scene, the proper vector of some unique points is similar, but the geometric position is ungenuine corresponding, because our data used have affine unchangeability, therefore adopt the Delaunay triangulation network Incremental insertion method based on the Qi algorithm, set up the Delaunay triangulation network, as shown in Figure 1, each triangle to Fig. 1 b, the ranks number that utilize three summits are corresponding with the triangle of corresponding Fig. 1 a benchmark image, set up inner affine Transform Model, pick out the unique point that does not satisfy proportionate relationship;

Build the Delaunay triangulation network, the corresponding more than one triangle of each point, utilize the coordinate coding of corresponding point to determine corresponding leg-of-mutton two other point of each point, utilize the corresponding benchmark images in three summits and image subject to registration to determine the affine Transform Model parameter of this triangle inside, because corresponding point are consistent basically, so parameter is got all corresponding leg-of-mutton averages; Specific implementation is as follows:

More corresponding triangle one by one, each triangle has three limits and three angles, as shown in Figure 2, to benchmark triangular graph 2a, judges whether triangular graph 2b subject to registration meets the following conditions:

|L1-L′1|<T1,|L2-L′2|<T1,|L3-L′3|<T1

|A1-A′1|<T2,|A2-A′2|<T2,|A3-A′3|<T2

T1 wherein, T2 are the threshold value of doing according to whole average adjustment of corresponding sides and angle, wherein T1, T2 be according to all corresponding sides and angle set threshold value, L1, L2, L3 is respectively benchmark image leg-of-mutton three limits, and L ' 1, and L ' 2, L ' 3 is leg-of-mutton three limits of image subject to registration, A1, A2, A3 is leg-of-mutton three angles of benchmark image, A ' 1, and A ' 2, and A ' 3 is leg-of-mutton three angles of image subject to registration; If top condition all satisfies, remember this leg-of-mutton three summits, if do not satisfy continue to search three summits corresponding other triangles respectively, if all triangles that point connects do not satisfy, delete this point; Write down the corresponding point that all satisfy condition.As shown in Figure 3, be the Delaunay triangulation network after misregistration point is eliminated; Fig. 3 a is the triangulation network of benchmark image, and Fig. 3 b is the triangulation network of registered images.

For the validity of verification algorithm, we provide one group of experiment, and experimental data is the SAR image of Same Scene, and difference is that two groups of image attitude angle are different, and is as Fig. 4, shown in Figure 5.

Respectively two width images are carried out the search of SIFT unique point, and 2 stack features points are carried out similar calculating, complicacy due to ground, some unique points are not real registration point, so we encode by the Delaunay triangulation network after to registration, the eliminating error match point obtains last registration results as shown in Figure 6.

Can find out by registration results, the SAR data very complex that the present invention uses has been included the ocean, land, the land also has the city, Plain and mountain area, can say that the complicated situation in ground makes simple SIFT registration Algorithm lose efficacy, and has produced the local feature similar situation of diverse location; Simultaneously, image subject to registration and benchmark image attitude angle are inconsistent, and this has also brought interference to registration.The impact of the Geometrical changes such as anti-rotation is preferably arranged due to SIFT algorithm and Delaunay triangulation network feature, therefore select these two kinds of algorithms and proposed encryption algorithm based on SIFT-Delaunay in the present invention, filtering error characteristic point effectively is in the situation that complex background and the registration that Geometrical change can be correct is arranged.

Claims (2)

1. a SAR automatic image registration method, is characterized in that, described method step is as follows:

Steps A, the SIFT characteristic matching:

Adopt the SIFT algorithm to extract the local feature of SAR image, seek extreme point, extracting position, yardstick, rotational invariants parameter at metric space; Two images subject to registration are extracted respectively the SIFT feature, and compare one by one, the unique point that satisfies the coupling requirement in two width images is kept, what do not meet the demands deletes;

Step B, the SIFT-Delaunay coding

SIFT coupling for two width images in steps A is right, adopts the coupling of SIFT-Delaunay coding eliminating error right;

Step B-1 to the SIFT unique point in every width image, adopts Delaunay triangulation network Incremental insertion method, sets up the Delaunay triangulation network;

Step B-2 to each triangle of the Delaunay triangulation network, utilizes the corresponding relation on three summits in triangle, sets up inner affine Transform Model, picks out the unique point that does not satisfy proportionate relationship;

Step B-3, to each diabolo, more corresponding limit and angle one by one judge whether to meet the following conditions:

|L1-L′1|<T1,|L2-L′2|<T1,|L3-L′3|<T1

|A1-A′1|<T2,|A2-A′2|<T2,|A3-A′3|<T2

T1 wherein, T2 be according to all corresponding sides and angle set threshold value, L1, L2, L3 are respectively benchmark image leg-of-mutton three limits, and L ' 1, L ' 2, L ' 3 is leg-of-mutton three limits of image subject to registration, A1, A2, A3 is leg-of-mutton three angles of benchmark image, A ' 1, and A ' 2, and A ' 3 is leg-of-mutton three angles of image subject to registration;

If do not satisfy condition, continue to search three summits corresponding other triangles respectively, if all triangles that point connects do not satisfy respective conditions, delete this point;

If satisfy condition, keep these leg-of-mutton three summits; Keep the corresponding point that all satisfy condition, at last with these reference point as registration, two width images are carried out registration.

2. a kind of SAR automatic image registration method according to claim 1, is characterized in that, in described steps A, the concrete steps of SIFT characteristic matching are as follows:

Steps A-1, the generation of SIFT proper vector

Steps A-1-1 utilizes Gaussian difference pyrene and image convolution to obtain difference of Gaussian DOG image, judges position and the yardstick of extreme point by the gray-scale value that compares pixel in the DOG image; Utilize the unsettled unique point of stability metric deletion, determine feature point set;

Steps A-1-2 utilizes the gradient direction distribution feature of unique point neighborhood territory pixel, for each unique point assigned direction parameter, makes operator have the yardstick unchangeability, determines unique point principal direction;

Steps A-1-3, coordinate axis is rotated to unique point principal direction, get the window of 4 * 4 m multiple size centered by unique point, m is positive integer, calculates the gradient orientation histogram of 8 directions in each image blockage of 4 * 4, draw the accumulated value of each gradient direction, form a Seed Points, 8 directions of each Seed Points form 8 * m dimension SIFT proper vector, with the length normalization method of proper vector, obtain the SIFT proper vector;

Steps A-2, the coupling of proper vector

Adopt Euclidean distance as the similarity measurement of two width image SIFT unique points, the preset proportion threshold value, it is right to obtain mating by similarity measurement.

Images