CN105787943A - SAR image registration method based on multi-scale image block characteristics and sparse expression - Google Patents

Abstract

The invention discloses an SAR image registration method based on multi-scale image block characteristics and sparse expression, and mainly solves the problem that an existing registration method is poor in effect when applied to SAR image registration. The SAR image registration method comprises the steps of: 1) inputting two SAR images, selecting any one as a reference image, and using the other one as an image to be registered; 2) selecting characteristic points of the reference image; 3) utilizing the multi-scale image block characteristics to construct characteristic point descriptors of the reference image and the image to be registered; 4) establishing matching point pairs between the reference image and the image to be registered; 5) removing abnormal points in the matching point pairs; 6) establishing an affine transformation model according to the finally obtained matching point pairs, adopting a least square method to obtain geometric deformation parameters, and obtaining a registration result. Compared with the prior art, the robustness to spot noise is improved, the accuracy and the matching precision of the matching point pairs are improved, and the SAR image registration method can be applied to image fusion and change detection.

Description

SAR image registration method based on multi-scale image block feature and rarefaction representation

Technical field

The invention belongs to technical field of image processing, be specifically related to the method for registering images during radar image processes, can be used for image co-registration and change-detection.

Background technology

Synthetic aperture radar SAR system is because of its round-the-clock, round-the-clock, has the features such as penetrance and is widely used in military and civilian neighborhood.SAR image registration apply as SAR image in key link, it is two width to the same scenery taking from different time, different visual angles or several SAR image are mated, the process of superposition.

For image registration problem, the method proposed at present substantially can be divided into two classes: based on the method for registering of gray scale and feature based.The half-tone information of image is directly utilized, the registration parameter such as translation corresponding when searching out Optimum Matching by setting up certain similarity measure between image pixel, rotation based on the method for registering of gray scale.The most frequently used method for registering based on gray scale is based on the method for registering of mutual information.Although this method comparison is directly perceived, it is easy to realize, but computation complexity is high, it is easy to be absorbed in locally optimal solution, and easily affected by noise.The method for registering of feature based is owing to being not directly placed on image intensity value, but acts on the feature of image itself, thus grey scale change has stronger adaptive capacity, and amount of calculation is little, it is possible to process the registration problems between image.The method for registering of the most frequently used feature based is based on the method for registering of Scale invariant features transform SIFT feature.Yet with there is speckle noise in SAR image, the method for registering of feature based is when processing SAR image registration, it is more likely that being detected by speckle noise is characteristic point, thus substantial amounts of error matching points can be brought, and the registration result led to errors.

Summary of the invention

It is an object of the invention to propose a kind of SAR image registration method based on multi-scale image block feature and rarefaction representation, during to solve prior art carries out SAR image registration, a large amount of error matching points occurs, cause the problem that registration accuracy is not high.

The technical thought realizing the object of the invention is: utilize the characteristic point that space correlation Sexual behavior mode is highly reliable, multi-scale image block feature is adopted to form feature descriptor, best matching double points is obtained according to the minimum difference criterion that rarefaction representation technology calculates, effectively strengthening the robustness to speckle noise, implementation step includes as follows:

(1) inputting two width images, an optional width is as reference picture I₁, using another width as image I subject to registration₂；

(2) reference picture characteristic point is chosen:

(2a) SIFT algorithm is adopted to extract reference picture I₁Characteristic point, and by I₁All characteristic points leave in the first set R；

(2b) from reference picture characteristic point set R, a characteristic point r is arbitrarily chosen_i, utilize Stationary Wavelet Transform method to calculate the spatial coherence ρ (r of each Feature point correspondence_i)；

(2c) threshold value E=0.05 is set, if the ρ (r obtained_i) meet ρ (r_i) >=E, then by reference picture characteristic point r_iRetain, otherwise, delete this characteristic point；

(2d) travel through all characteristic points of reference picture, repeat step (2b)-(2c), the reference picture characteristic point after being screened；

(2e) the Euclidean distance E between any two characteristic point in reference picture characteristic point set is calculated after above-mentioned screening respectively_dIf, E_d>=15, then retain the two characteristic point, otherwise remove；

(2f) front 10 characteristic points in characteristic point set step (2e) obtained are as final reference picture characteristic point；

(3) multiscale image block feature construction reference picture characteristic point descriptor is utilized:

(3a) arbitrarily choose a reference picture characteristic point a, take image block P (a) of this characteristic point surrounding neighbors 15 × 15；

(3b) adopt Stationary Wavelet Transform that image block P (a) is carried out multi-resolution decomposition, obtain the image block P of three different decomposition yardsticks_s(a), s=3,4,5；

(3c) the grey level histogram vector H of above three different decomposition yardstick image block is calculated respectively_sA (), it can be used as the gray feature of feature descriptor；

(3d) the gradient orientation histogram vector G of above three different decomposition yardstick image block is calculated respectively_sA (), it can be used as the Gradient Features of feature descriptor；

(3e) gray feature corresponding for different decomposition yardstick image block and Gradient Features are together in series, obtain reference picture characteristic point a characteristic of correspondence descriptor F (a)={ H₃(a),H₄(a),H₅(a),G₃(a),G₄(a),G₅(a)}；

(4) for any one image characteristic point b subject to registration, its characteristic point descriptor F (b) is obtained according to the operation that step (3) is same, utilize the similarity between reference picture characteristic point descriptor F (a) and image characteristic point descriptor F (b) subject to registration, set up the matching double points between reference picture and image subject to registration；

(5) abnormity point in the matching double points that removal step (4) obtains:

(5a) from the matching double points that step (4) obtains, a reference picture characteristic point r is arbitrarily chosen_c, the reference picture characteristic point mated takes 3 neighborhood points of this characteristic point arest neighbors, and these 3 the neighborhood points taken is mapped to and reference picture characteristic point r_cBecome the image characteristic point t subject to registration of match point_cNeighborhood in, calculate this reference picture characteristic point r_cAnd the geometry cost between its neighborhood point

$O(r_c,r_c^k,t_c,t_{m\left(k\right)})=\underset{k=1}{\overset{3}{\Σ}}\frac{\left|\right|r_c-r_c^k\left|\right|-\left|\right|t_c-t_{m\left(k\right)}\left|\right|}{\left|\right|r_c-r_c^k\left|\right|+\left|\right|t_c-t_{m\left(k\right)}\left|\right|},$

Wherein,Represent reference picture characteristic point r_cKth nearest neighbor point, t_cRepresent and reference picture characteristic point r_cThe image characteristic point subject to registration matched, m () represents adaptation function, | | | | represent Euclidean distance, c represents the index of matching double points, its span is 1 to 10, and k represents the index of the nearest neighbor point that the c reference picture characteristic point take, and its span is 1 to 3；

(5b) travel through and all mated reference picture characteristic point, repeat step (5a), mated the geometry cost between reference picture characteristic point and its respective neighborhood point, using all characteristic points corresponding for geometry Least-cost value as datum mark set, be expressed as:

$m_{base}=\{(r_c,t_c),(r_c^1,t_{m\left(1\right)}),...,(r_c^k,t_{m\left(k\right)}),...,(r_c^3,t_{m\left(3\right)})\},$

Wherein, (r_c,t_c) represent matching double points, t_m(k)Represent and neighborhood pointCorresponding match point；

(5c) following formula is adopted to calculate residue match point to the geometry cost between datum mark:

$O(r_{c^{\′}},r_{base}^o,t_{c^{\′}},t_{m\left(o\right)})=\underset{o=1}{\overset{4}{\Σ}}\frac{\left|\right|r_{c^{\′}}-r_{base}^o\left|\right|-\left|\right|t_{c^{\′}}-t_{m\left(o\right)}\left|\right|}{\left|\right|r_{c^{\′}}-r_{base}^o\left|\right|+\left|\right|t_{c^{\′}}-t_{m\left(o\right)}\left|\right|},(r_{c^{\′}},t_{c^{\′}})\∉m_{base}$

Wherein, r_c′And t_c′Represent remaining reference picture characteristic point and image characteristic point subject to registration in matching double points respectively,Represent the o reference picture characteristic point in datum mark set, t_m(o)Representing and the match point of the o reference picture Feature point correspondence in collection on schedule set, c ' represents the index of residue matching double points, and its span is 1 to 6, and o represents the index of datum mark, and its span is 1 to 4；

(5d) threshold value E is set_o=0.03, if the geometry cost obtainedMeetThen by (r_c′,t_c′) as correct match point, otherwise, this matching double points is deleted；

(5e) repeat step (5c)-(5d), obtain reference picture and the final matching double points of image subject to registration；

(6) according to final matching double points obtained above, set up affine Transform Model, calculate the geometric deformation parameter of image subject to registration, and utilize this geometric deformation parameter, image subject to registration is carried out geometric transformation, obtains registration result.

The present invention compared with prior art has the advantage that

First, owing to the present invention is in the process to SAR image registration, adopt space correlation Sexual behavior mode characteristic point, utilize multiscale image block feature construction feature descriptor simultaneously, overcoming prior art cannot the deficiency of accurate description characteristic point attribute as block message only with single scalogram, make the present invention improve the significance of characteristic point, enhance the robustness to speckle noise.

Second, owing to the present invention utilizes the minimum difference criterion based on rarefaction representation technology to set up matching double points, adopt the geometrical-restriction relation between match point and its neighborhood point to filter feature abnormalities point simultaneously, overcoming prior art adopts Euclidean distance ratio method easily to occur the deficiency of error matching points in the process set up matching double points so that the present invention improves the accuracy of match point.

Accompanying drawing explanation

Fig. 1 is the flowchart of the present invention；

Fig. 2 is first group of the simulation experiment result figure of the present invention；

Fig. 3 is second group of the simulation experiment result figure of the present invention.

Detailed description of the invention

Below in conjunction with accompanying drawing, the present invention is described further:

With reference to Fig. 1, the enforcement step of the present invention is as follows:

Step 1, inputs two width images, and an optional width is as reference picture I₁, using another width as image I subject to registration₂。

Two width images of input are to intercept respectively in the different polarization modes obtained at certain airborne radar or different phase two width SAR image.

Step 2, chooses reference picture characteristic point.

2.1) SIFT algorithm is adopted to extract reference picture I₁Characteristic point, by I₁All characteristic points leave in the first set R；

2.2) from reference picture characteristic point set R, a characteristic point r is arbitrarily chosen_i, utilize Stationary Wavelet Transform method to calculate the spatial coherence ρ (r of each Feature point correspondence_i)；

2.2a) utilize Stationary Wavelet Transform that input picture is carried out s Scale Decomposition, obtain the detail pictures that input picture is 3 kinds different on different scale；

2.2b) define arbitrary image pixel x amplitude M under s yardstick_sX () is expressed as:

$M_s\left(x\right)=\sqrt{|W_s^h{|}^2+|W_s^v{|}^2+|W_s^d{|}^2}$

Wherein,It is illustrated respectively in s yardstick hypograph in the horizontal direction, vertical direction and diagonally adjacent detailed information, | | represent absolute value operation computing；

Following formula 2.2c) is adopted to calculate spatial coherence ρ (x) of pixel x:

$\mathrm{\ρ}\left(x\right)=\underset{s=3}{\overset{5}{\Π}}{M}_s\left(x\right)$

Wherein, ∏ () represents multiplication operations；

2.3) threshold value E=0.05 is set, if the ρ (r obtained_i) meet ρ (r_i) >=E, then by reference picture characteristic point r_iRetain, otherwise, delete this characteristic point；

2.4) the traversal all characteristic points of reference picture, repeat step 2.2)-2.3), the reference picture characteristic point after being screened；

2.5) the Euclidean distance E between any two characteristic point in reference picture characteristic point set is calculated after above-mentioned screening respectively_dIf, E_d>=15, then retain the two characteristic point, otherwise remove；

2.6) using step 2.5) front 10 characteristic points in the characteristic point set that obtains are as final reference picture characteristic point.

Step 3: utilize multiscale image block feature construction reference picture characteristic point descriptor.

3.1) arbitrarily choose a reference picture characteristic point a, take image block P (a) of this characteristic point surrounding neighbors 15 × 15；

3.2) adopt Stationary Wavelet Transform that image block P (a) is carried out multi-resolution decomposition, obtain the image block P of three different decomposition yardsticks_s(a), s=3,4,5；

3.3) the grey level histogram vector H of above three different decomposition yardstick image block is calculated respectively_sA (), it can be used as the gray feature of characteristic point descriptor；

3.4) the gradient orientation histogram vector G of above three different decomposition yardstick image block is calculated respectively_sA (), it can be used as the Gradient Features of characteristic point descriptor；

3.5) gray feature corresponding for different decomposition yardstick image block and Gradient Features are together in series, obtain reference picture characteristic point a characteristic of correspondence descriptor F (a)={ H₃(a),H₄(a),H₅(a),G₃(a),G₄(a),G₅(a)}。

Step 4: for any one image characteristic point b subject to registration, its characteristic point descriptor F (b) is obtained according to the operation that step (3) is same, utilize the similarity between reference picture characteristic point descriptor F (a) and image characteristic point descriptor F (b) subject to registration, set up the matching double points between reference picture and image subject to registration.

The method forming matching double points has a variety of, common are arest neighbors method, Euclidean distance ratio etc., adopts but be not limited to set up based on the minimum difference criterion of rarefaction representation technology matching double points in this example, and it is as follows that it is embodied as step:

4.1) from the reference picture characteristic point that step (2) obtains, a characteristic point r is arbitrarily chosen_i, calculate its characteristic of correspondence descriptor F (r_i)；

4.2) set maximum side-play amount between reference picture and image subject to registration as l=100, image subject to registration is chosen the window W of L × L size_iAs characteristic point r_iRegion of search, and by pixel { V all in this region_q: q=1,2 ..., Q, Q=L × L} is all as characteristic point r_iCandidate matches point, wherein L=2 × l+1；

4.3) for above-mentioned each candidate matches point V_q, take the window of this pixel surrounding neighbors 20 × 20, and by all pixel u in this window_nCharacteristic of correspondence descriptor F (u_n) as V_qCorresponding sparse dictionary D_q={ F (u_n), n=1,2 ..., J, J=20 × 20}；

4.4) utilize orthogonal matching pursuit algorithm to solve following formula, obtain characteristic point r_iAt sparse dictionary D_qUnder sparse vector α_q

${\mathrm{\α}}_q=\underset{{\mathrm{\α}}_q}{\mathrm{argmin}}\left|\right|F\left(r_i\right)-D_q{\mathrm{\α}}_q\left|\right|s.t.\left|\right|{\mathrm{\α}}_q|{|}_0\≤C$

Wherein, argmin () representative function reaches the value of independent variable during minima, | | | | represent Euclidean distance, | | | |₀Representing zero norm, C represents degree of rarefication；

4.5) cycling among windows W_iInterior all pixels, repeat step 4.3)-4.4), obtain characteristic point r_iSparse vector α under Q different sparse dictionary_q；

4.6) it is calculated as follows characteristic point r_iWith any candidate matches point V_qBetween reconstructed error, and using the pixel with minimal reconstruction error as characteristic point r_iMatch point m (r_i)

$m\left(r_i\right)=\underset{q}{\mathrm{argmin}}\left|\right|F\left(r_i\right)-D_q{\mathrm{\α}}_q\left|\right|;$

4.7) travel through all reference picture characteristic points, repeat step 4.1)-4.6), obtain the match point that all reference picture characteristic points are corresponding in image subject to registration, all characteristic points of image subject to registration left in the second set T.

Step 5: the abnormity point in the matching double points that removal step (4) obtains.

5.1) from the matching double points that step (4) obtains, a reference picture characteristic point r is arbitrarily chosen_c, the reference picture characteristic point mated takes 3 neighborhood points of this characteristic point arest neighbors, and these 3 the neighborhood points taken is mapped to and reference picture characteristic point r_cBecome the image characteristic point t subject to registration of match point_cNeighborhood in, calculate this reference picture characteristic point r_cAnd the geometry cost between its neighborhood point

$O(r_c,r_c^k,t_c,t_{m\left(k\right)})=\underset{k=1}{\overset{3}{\Σ}}\frac{\left|\right|r_c-r_c^k\left|\right|-\left|\right|t_c-t_{m\left(k\right)}\left|\right|}{\left|\right|r_c-r_c^k\left|\right|+\left|\right|t_c-t_{m\left(k\right)}\left|\right|},$

Wherein,Represent reference picture characteristic point r_cKth nearest neighbor point, t_cRepresent and reference picture characteristic point r_cThe image characteristic point subject to registration matched, m () represents adaptation function, | | | | represent Euclidean distance, c represents the index of matching double points, its span is 1 to 10, and k represents the index of the nearest neighbor point that the c reference picture characteristic point take, and its span is 1 to 3；

5.2) travel through and all mated reference picture characteristic point, repeat step 5.1), mated the geometry cost between reference picture characteristic point and its respective neighborhood point, using all characteristic points corresponding for geometry Least-cost value as datum mark set, be expressed as:

$m_{base}=\{(r_c,t_c),(r_c^1,t_{m\left(1\right)}),...,(r_c^k,t_{m\left(k\right)}),...,(r_c^3,t_{m\left(3\right)})\},$

Wherein, (r_c,t_c) represent matching double points, t_m(k)Represent and neighborhood pointCorresponding match point；

5.3) following formula is adopted to calculate residue match point to the geometry cost between datum mark:

$O(r_{c^{\′}},r_{base}^o,t_{m\left(o\right)})=\underset{o=1}{\overset{4}{\Σ}}\frac{\left|\right|r_{c^{\′}}-r_{base}^o\left|\right|-\left|\right|t_{c^{\′}}-t_{m\left(o\right)}\left|\right|}{\left|\right|r_{c^{\′}}-r_{base}^o\left|\right|+\left|\right|t_{c^{\′}}-t_{m\left(o\right)}\left|\right|},(r_{c^{\′}},t_{c^{\′}})\∉m_{base}$

Wherein, r_c′And t_c′Represent remaining reference picture characteristic point and image characteristic point subject to registration in matching double points respectively,Represent the o reference picture characteristic point in datum mark set, t_m(o)Representing and the match point of the o reference picture Feature point correspondence in collection on schedule set, c ' represents the index of residue matching double points, and its span is 1 to 6, and o represents the index of datum mark, and its span is 1 to 4；

5.4) threshold value E is set_o=0.03, if the geometry cost obtainedMeetThen by (r_c′,t_c′) as correct match point, otherwise, this matching double points is deleted；

5.5) step 5.3 is repeated)-5.4), obtain reference picture and the final matching double points of image subject to registration.

Step 6: according to final matching double points, obtain registration result.

Affine Transform Model is set up according to final matching double points obtained above, calculating the geometric deformation parameter of image subject to registration, this example specifically adopts method of least square computational geometry deformation parameter, and utilizes this geometric deformation parameter, image subject to registration is carried out geometric transformation, obtains registration result.

Below in conjunction with experiment simulation, effect of the present invention is described further.

1. simulated conditions:

The Simulation Experimental Platform of the present invention adopts Intel (R) Pentium (R) CPUG32403.10GHz, inside saves as 4GB, runs the PC of Windows7, and programming language is Matlab2011b.

2. emulation content and interpretation of result:

Emulation 1, apply the method for registering images based on SIFT-OCT, method for registering images based on BFSIFT respectively, SAR image is carried out registration by method for registering images and the present invention based on NDSS-SIFT, result is as shown in Figure 2, wherein Fig. 2 (a) is based on the match point line graph of SIFT-OCT method, Fig. 2 (b) is based on the match point line graph of BFSIFT method, Fig. 2 (c) is based on the match point line graph of NDSS-SIFT method, and Fig. 2 (d) is the match point line graph of the present invention.In Fig. 2, yellow solid line represents correct matching double points, and red line represents error matching points pair.

Emulation 2, apply the method for registering images based on SIFT-OCT, method for registering images based on BFSIFT respectively, SAR image is carried out registration by method for registering images and the present invention based on NDSS-SIFT, result is as shown in Figure 3, wherein Fig. 3 (a) is based on the match point line graph of SIFT-OCT method, Fig. 3 (b) is based on the match point line graph of BFSIFT method, Fig. 3 (c) is based on the match point line graph of NDSS-SIFT method, and Fig. 3 (d) is the match point line graph of the present invention.In Fig. 3, yellow solid line represents correct matching double points, and red line represents error matching points pair.

Can be seen that from Fig. 2 (a)-2 (c) and Fig. 3 (a)-3 (c), SAR image pair is surveyed for 2 groups, all there is relatively more error matching points in SIFT-OCT method, BFSIFT method and 3 kinds of algorithms of NDSS-SIFT method.Wherein, SIFT-OCT method effect is worst, and the error matching points comprised is maximum.Error matching points in BFSIFT method and NDSS-SIFT method is to relatively fewer.

From Fig. 2 (d) and Fig. 3 (d) it can be seen that survey SAR image pair for 2 groups, the registration result that the present invention obtains is more accurate, does not comprise error matching points.This comes from the present invention and utilizes spatial coherence to choose reliable characteristic point, utilizes multiscale image block feature construction feature descriptor simultaneously, enhances the significance of feature descriptor and the robustness to speckle noise.Additionally, utilize the minimum difference criterion based on rarefaction representation technology to set up matching double points, improve the accuracy of matching double points, thus solving problem when prior art is applied to SAR image registration, a large amount of error matching points pair occurring.

Claims (3)

1., based on a SAR image registration method for multi-scale image block feature and rarefaction representation, comprise the steps:

(1) inputting two width images, an optional width is as reference picture I₁, using another width as image I subject to registration₂；

(2) reference picture characteristic point is chosen:

(2a) SIFT algorithm is adopted to extract reference picture I₁Characteristic point, and by I₁All characteristic points leave in the first set R；

(2b) from reference picture characteristic point set R, a characteristic point r is arbitrarily chosen_i, utilize Stationary Wavelet Transform method to calculate the spatial coherence ρ (r of each Feature point correspondence_i)；

(2c) threshold value E=0.05 is set, if the ρ (r obtained_i) meet ρ (r_i) >=E, then by reference picture characteristic point r_iRetain, otherwise, delete this characteristic point；

(2d) travel through all characteristic points of reference picture, repeat step (2b)-(2c), the reference picture characteristic point after being screened；

(2e) the Euclidean distance E between any two characteristic point in reference picture characteristic point set is calculated after above-mentioned screening respectively_dIf, E_d>=15, then retain the two characteristic point, otherwise remove；

(2f) front 10 characteristic points in characteristic point set step (2e) obtained are as final reference picture characteristic point；

(3) multiscale image block feature construction reference picture characteristic point descriptor is utilized:

(3a) arbitrarily choose a reference picture characteristic point a, take image block P (a) of this characteristic point surrounding neighbors 15 × 15；

(3b) adopt Stationary Wavelet Transform that image block P (a) is carried out multi-resolution decomposition, obtain the image block P of three different decomposition yardsticks_s(a), s=3,4,5；

(3c) the grey level histogram vector H of above three different decomposition yardstick image block is calculated respectively_sA (), it can be used as the gray feature of feature descriptor；

(3d) the gradient orientation histogram vector G of above three different decomposition yardstick image block is calculated respectively_sA (), it can be used as the Gradient Features of feature descriptor；

(3e) gray feature corresponding for different decomposition yardstick image block and Gradient Features are together in series, obtain reference picture characteristic point a characteristic of correspondence descriptor F (a)={ H₃(a),H₄(a),H₅(a),G₃(a),G₄(a),G₅(a)}；

(4) for any one image characteristic point b subject to registration, its characteristic point descriptor F (b) is obtained according to the operation that step (3) is same, utilize the similarity between reference picture characteristic point descriptor F (a) and image characteristic point descriptor F (b) subject to registration, set up the matching double points between reference picture and image subject to registration；

(5) abnormity point in the matching double points that removal step (4) obtains:

(5a) from the matching double points that step (4) obtains, a reference picture characteristic point r is arbitrarily chosen_c, the reference picture characteristic point mated takes 3 neighborhood points of this characteristic point arest neighbors, and these 3 the neighborhood points taken is mapped to and reference picture characteristic point r_cBecome the image characteristic point t subject to registration of match point_cNeighborhood in, calculate this reference picture characteristic point r_cAnd the geometry cost between its neighborhood point

$O(r_c,r_c^k,t_c,t_{m\left(k\right)})=\underset{k=1}{\overset{3}{\Σ}}\frac{\left|\right|r_c-r_c^k\left|\right|-\left|\right|t_c-t_{m\left(k\right)}\left|\right|}{\left|\right|r_c-r_c^k\left|\right|+\left|\right|t_c-t_{m\left(k\right)}\left|\right|},$

Wherein,Represent reference picture characteristic point r_cKth nearest neighbor point, t_cRepresent and reference picture characteristic point r_cThe image characteristic point subject to registration matched, m () represents adaptation function, | | | | represent Euclidean distance, c represents the index of matching double points, its span is 1 to 10, and k represents the index of the nearest neighbor point that the c reference picture characteristic point take, and its span is 1 to 3；

(5b) travel through and all mated reference picture characteristic point, repeat step (5a), mated the geometry cost between reference picture characteristic point and its respective neighborhood point, using all characteristic points corresponding for geometry Least-cost value as datum mark set, be expressed as:

$m_{base}=\{(r_c,t_c),(r_c^1,t_{m\left(1\right)}),...,(r_c^k,t_{m\left(k\right)}),...,(r_c^3,t_{m\left(3\right)})\},$

Wherein, (r_c,t_c) represent matching double points, t_m(k)Represent and neighborhood pointCorresponding match point；

(5c) following formula is adopted to calculate residue match point to the geometry cost between datum mark:

$O(r_{c^{\′}},r_{base}^o,t_{c^{\′}},t_{m\left(o\right)})=\underset{o=1}{\overset{4}{\Σ}}\frac{\left|\right|r_{c^{\′}}-r_{base}^o\left|\right|-\left|\right|t_{c^{\′}}-t_{m\left(o\right)}\left|\right|}{\left|\right|r_{c^{\′}}-r_{base}^o\left|\right|+\left|\right|t_{c^{\′}}-t_{m\left(o\right)}\left|\right|},\left(r_{c^{\′}},t_{c^{\′}}\right)\∉m_{base}$

Wherein, r_c′And t_c′Represent remaining reference picture characteristic point and image characteristic point subject to registration in matching double points respectively,Represent the o reference picture characteristic point in datum mark set, t_m(o)Representing and the match point of the o reference picture Feature point correspondence in collection on schedule set, c ' represents the index of residue matching double points, and its span is 1 to 6, and o represents the index of datum mark, and its span is 1 to 4.

(5d) threshold value E is set_o=0.03, if the geometry cost obtainedMeetThen by (r_c′,t_c′) as correct match point, otherwise, this matching double points is deleted；

(5e) repeat step (5c)-(5d), obtain reference picture and the final matching double points of image subject to registration；

(6) according to final matching double points obtained above, set up affine Transform Model, calculate the geometric deformation parameter of image subject to registration, and utilize this geometric deformation parameter, image subject to registration is carried out geometric transformation, obtains registration result.

2. SAR image registration method according to claim 1, it is characterised in that utilize Stationary Wavelet Transform method to calculate the spatial coherence of each Feature point correspondence in step (2b), carry out as follows:

2b1) utilize Stationary Wavelet Transform that input picture is carried out s Scale Decomposition, obtain the detail pictures that input picture is 3 kinds different on different scale；

2b2) define arbitrary image pixel x amplitude M under s yardstick_sX () is expressed as:

$M_s\left(x\right)=\sqrt{{\left|W_s^h\right|}^2+{\left|W_s^v\right|}^2+{\left|W_s^d\right|}^2}$

Wherein,It is illustrated respectively in s yardstick hypograph in the horizontal direction, vertical direction and diagonally adjacent detailed information, | | represent absolute value operation computing；

Following formula 2b3) is adopted to calculate spatial coherence ρ (x) of pixel x:

$\mathrm{\ρ}\left(x\right)=\underset{s=3}{\overset{5}{\Π}}{M}_s\left(x\right)$

Wherein, ∏ () represents multiplication operations.

3. the SAR image registration method based on multi-scale image block feature and rarefaction representation according to claim 1, it is characterised in that set up the matching double points between reference picture and image subject to registration in described step (4), carry out as follows:

4a) from the reference picture characteristic point that step (2) obtains, arbitrarily choose a characteristic point r_i, calculate its characteristic of correspondence descriptor F (r_i)；

4b) set maximum side-play amount between reference picture and image subject to registration as l=100, image subject to registration is chosen the window W of L × L size_iAs characteristic point r_iRegion of search, by pixel { V all in this region_q: q=1,2 ..., Q, Q=L × L} is all as characteristic point r_iCandidate matches point, wherein L=2 × l+1；

4c) for above-mentioned each candidate matches point V_q, take the window of this pixel surrounding neighbors 20 × 20, and by all pixel u in this window_nCharacteristic of correspondence descriptor F (u_n) as V_qCorresponding sparse dictionary D_q={ F (u_n), n=1,2 ..., J, J=20 × 20}；

4d) utilize orthogonal matching pursuit algorithm to solve following formula, obtain characteristic point r_iAt sparse dictionary D_qUnder sparse vector α_q

$\begin{array}{cc}{\mathrm{\α}}_q=\underset{{\mathrm{\α}}_q}{\arg \min }\left|\right|F\left(r_i\right)-D_q{\mathrm{\α}}_q\left|\right|& s.t.\left|\right|{\mathrm{\α}}_q|{|}_0\≤C\end{array}$

Wherein, argmin () representative function reaches the value of independent variable during minima, | | | | represent Euclidean distance, | | | |₀Representing zero norm, C represents degree of rarefication；

4e) cycling among windows W_iInterior all pixels, repeat step 4c)-4d), obtain characteristic point r_iSparse vector α under Q different sparse dictionary_q。

4f) it is calculated as follows characteristic point r_iWith any candidate matches point V_qBetween reconstructed error, and using the pixel with minimal reconstruction error as characteristic point r_iMatch point m (r_i)

$m\left(r_i\right)=\underset{q}{\arg \min }\left|\right|F\left(r_i\right)-D_q{\mathrm{\α}}_q\left|\right|;$

4g) travel through all reference picture characteristic points, repeat step (4a)-(4f), obtain the match point that all reference picture characteristic points are corresponding in image subject to registration, all characteristic points of image subject to registration are left in the second set T.