CN103310453A - Rapid image registration method based on sub-image corner features - Google Patents

Abstract

The invention discloses a rapid image registration method based on sub-image corner features. The method includes the specific steps: firstly, selecting a reference sub-image and a to-be-registered sub-image, selecting one sub-image from a reference image as the reference sub-image, and selecting one sub-image with the same coordinate space as the reference sub-image from a to-be-registered image as the to-be-registered sub-image; secondly, extracting corners of the reference sub-image and the reference sub-image; thirdly, performing feature description on the corners extracted from the reference sub-image and the reference sub-image to obtain a feature vector of each corner; fourthly, performing similarity measurement and feature matching on the feature vectors of the corners on the reference sub-image and the reference sub-image to obtain K matching point pairs; and fifthly, adopting a least square method to compute a transformation matrix H between the reference image and the to-be-registered image based on the K matching point pairs, and registering the to-be-registered image onto the reference image based on the transformation matrix H. By the method, the requirement for image matching precision can be met, and image matching speed is increased greatly.

Description

A kind of fast image registration method based on the subimage Corner Feature

Technical field

The invention belongs to the image registration techniques field, be specifically related to a kind of fast image registration method based on the subimage Corner Feature.

Background technology

The application of image registration is very extensive, such as fields such as pattern-recognition, self-navigation, medical diagnosis, computer visions.Having carried out many research work aspect the registration of image, multiple method for registering images has been proposed at present.Most concentrates on feature extraction to the research of image registration, and feature is described, similarity measurement, relatively the waiting of multiple method for registering, and the real-time of less concern registration.

Common method for registering can be divided into two classes: the feature-based registration method, such as Harris cornerpoints method, SIFT method etc.; Based on the method for registering in zone, such as mutual information, FMT etc.Wherein the method for registering based on the zone does not need to extract feature, is applicable to half-tone information greater than the situation of structural information, and requires the gray scale function of two width of cloth images must be similar or statistical dependence at least; From geometric angle, it only can process the situation of translation and small angle rotation, and big angle rotary or yardstick convergent-divergent must mean the raising of computation complexity and time complexity, so the scope of application is narrower.And the feature-based registration method can the diverse image of registration two width of cloth natural qualities, and adapts to geometry and optical distortion complicated between two width of cloth images, therefore becomes the focus of Recent study.But its bottleneck is how correctly to detect characteristic of correspondence, and carries out the feature description of low complex degree, robust, to improve the efficient of match search.Because shooting environmental and scenery distribute, so that take the image that obtains in contrast, text structure, the aspect distributed poles such as textural characteristics are inhomogeneous.Directly when large figure carries out feature extraction (being called for short large figure method), may become on the contrary the noise spot of correct coupling in the unique point of feature Fuzzy extracted region, cause mismatch; In addition, a large amount of unique points has been expanded the scope of search volume, causes search efficiency and real-time to descend.

The technology that addresses this problem at present has: 1. " a kind of method for registering that mutual information is combined with template matches ", adopting mutual information is the template matches that similarity criteria carries out image, obtain candidate's Matching sub-image, spatial relationship by large figure remainder subject to registration and template and subgraph, obtain the large figure behind the registration, calculate respectively the mutual information according to the figure subject to registration in each candidate's subgraph registration situation, obtain the maximum corresponding subgraph of mutual information, determine final registration result.But the method mismatch phenomenon easily occurs, and the registration time is long when the little image of gray scale difference is carried out registration.2. based on the method for registering images of wavelet transformation, utilize wavelet coefficient to choose effective subgraph, and utilize wavelet transformation that image is divided into some levels, utilize cross-correlation coefficient as similarity measure, realize at last the registration of image by iterative refinement algorithm.But the method relates to Calculation of correlation factor, the wavelet coefficient subgraph is chosen and the more step consuming time such as iteration refinement, so that algorithm complex is high, realizes that difficulty is large, and real-time is poor.

Summary of the invention

Given this, the present invention proposes a kind of rapid registering method based on the subimage Corner Feature on the basis of improving based on the feature registration method, is intended to satisfy the robustness and the real-time that improve registration under the prerequisite of registration accuracy.

In order to solve the problems of the technologies described above, the present invention is achieved in that

A kind of fast image registration method based on the subimage Corner Feature, concrete steps comprise:

Step 1, choose with reference to subgraph and subgraph subject to registration;

From reference picture, choose a subgraph as the reference subgraph, from image subject to registration, choose a coordinate space with reference to the identical subgraph of subgraph as subgraph subject to registration;

Step 2, extract the angle point with reference to subgraph and subgraph subject to registration;

Step 3, the angle point that extracts on reference subgraph and the subgraph subject to registration is carried out feature describe, obtain the proper vector of each angle point;

Step 4, carry out similarity measurement and characteristic matching with subgraph to be matched with reference to the proper vector of angle point on the subgraph, finally obtain K matching double points;

The detailed process of this step is:

1) for each angle point p that extracts _i, seek and p _iP contiguous point consists of p _iApart from neighborhood, i=1,2 ... N, N by on two width of cloth subgraphs total number of extraction angle point;

2) calculate successively the mahalanobis distance of all the Corner Feature vectors in the reference subgraph of each Corner Feature vector in the subgraph subject to registration, with mahalanobis distance less than setting threshold d _Mth1Two angle points be defined as matching double points, a plurality of matching double points consist of the coupling formation;

3) in the coupling formation, reject not at the matching double points in respective distances field, obtain K matching double points;

Step 5, based on K matching double points, adopt least square method to calculate transformation matrix H between image subject to registration and the reference picture, utilize described transformation matrix H with image registration subject to registration to reference picture.

Further, of the present invention is strong, the obvious width of cloth subgraph of architectural feature of contrast on the reference picture with reference to subgraph.

Further, the process of choosing with reference to subgraph of the present invention is:

At first become n the subgraph that size is identical with reference to image segmentation, next calculates entropy and the average gradient of each subgraph, then selects the subgraph of entropy and average gradient sum maximum as the reference subgraph.

Further, the present invention extracts with reference to subgraph identical with the method for subgraph angle point subject to registration, and detailed process is:

At first, based on the Harris angular-point detection method, detect the subgraph angle point; Secondly, the method for taking the non-very big inhibition of neighborhood and total amount to suppress is screened the angle point that initial detecting goes out, and extracts the top n angle point; Then, remove the angle point that is positioned on the subgraph borderline region, thereby extract required angle point.

Further, subgraph borderline region of the present invention is for being the rectangular area of 4* σ apart from border width, and wherein σ is the Gaussian smoothing factor.

Further, the 12 dimension gradient vectors of utilizing angle point that the present invention is better represent the feature Description Matrix of angle point.

Further, the present invention is calculating mahalanobis distance d _MIn the process of (i, j), utilize orthogonal matrix P and diagonal matrix D to represent the inverse matrix C of covariance matrix C ^-1, convert the calculating of mahalanobis distance to Euclidean distance d _FCalculating;

$d_M(i,j)=\sqrt{{(X_i^1-X_j^2)}^t{C}^{-1}(X_i^1-X_j^2)}$

$=\sqrt{{(X_i^1-X_j^2)}^t{P}^t\sqrt{D}\·\sqrt{D}P(X_i^1-X_j^2)}$

$=d_E(\sqrt{D}{\mathrm{PX}}_i^1,\sqrt{D}P{X}_j^2)$

Wherein,

The proper vector that represents i angle point on the subgraph subject to registration,

Expression is with reference to the proper vector of j angle point on the subgraph.

Beneficial effect:

The first, the present invention is by extracting subgraph at large figure subject to registration with reference to large figure, and the passing threshold setting extracts the matching double points of high matching degree, thereby so that the present invention can be fast, the coupling between the accurate subject to registration and reference diagram.

The second, the present invention is at large figure subject to registration and chooses the subgraph of contrast and clear in structure with reference to large figure, then carry out subsequent treatment according to the feature-based registration method, be that the high operation of the computation complexities such as extract minutiae and match search is directly carried out at subgraph rather than large figure, improved the accuracy that character pair point detects, the faster more accurate transformation matrix that estimates has finally improved precision and the real-time of registration.

Three, select based on improved Harris angular-point detection method, its detection be gray scale and the violent maximal point of graded in the angle point subrange.The angle point number is more for extracting, excessive, the slow-footed problem of coupling computation complexity, the method that the present invention takes the non-very big inhibition of neighborhood and total amount to suppress is screened the angle point that initial detecting goes out, and remove being had a few in the 4* σ in border among the R (σ is the Gaussian smoothing factor), purpose is that the angle point of reservation is tried one's best at the middle position of entire image, avoid because rotation, translation, convergent-divergent etc. shift out the part angle point, improve the repetition rate of extracting angle point.

Four, the present invention carries out Steerable filter to each corner pixels, provides the total derivative along gradient direction; For tackling the gray difference that causes because of affined transformation, remove first order derivative, the proper vector take 12 dimension gradient vectors as angle point.Compare, based on the method for SIFT, its descriptor is 128 dimensions, and improved PCA-SIFT is 36 dimensions.The present invention has larger minimizing at algorithm efficiency and calculated amount.

Five, the present invention converts finding the solution of mahalanobis distance to the calculating of simple Euclidean distance, can avoid redundant inversion operation when hardware is realized, reduces the space requirement to internal memory, improves the real-time of algorithm.

Description of drawings

Fig. 1 is subgraph method registration schematic flow sheet.

Embodiment

Below in conjunction with the accompanying drawing embodiment that develops simultaneously, describe the present invention.

Suppose that at first distortion model is affined transformation between image, its mathematical notation is as follows:

$\left[\begin{array}{c}{x_i}^{\′}\\ {y_i}^{\′}\end{array}\right]=s\left[\begin{array}{cc}\cos \mathrm{\θ}& -\sin \mathrm{\θ}\\ \sin \mathrm{\θ}& \cos \mathrm{\θ}\end{array}\right]*\left[\begin{array}{c}{x}_i\\ y_i\end{array}\right]+\left[\begin{array}{c}{t}_x\\ t_y\end{array}\right]=\left[\begin{array}{cc}{a}_1& a_2\\ a_3& a_4\end{array}\right]*\left[\begin{array}{c}{x}_i\\ y_i\end{array}\right]+\left[\begin{array}{c}{t}_x\\ t_y\end{array}\right]---\left(1\right)$

Wherein, s is scale factor, and θ is rotation angle, t _xBe x direction translational movement, t _yBe y direction translational movement, (x _i', y _i') be the point on the image after the distortion, (x _i, y _i) be the point on the image before the distortion.

As shown in Figure 1, based on the method for registering images of Sub-Image Feature, concrete steps are:

Step 1, choose with reference to subgraph and subgraph subject to registration, namely from reference picture, choose a subgraph as the reference subgraph, from image subject to registration, choose a coordinate space with reference to the identical subgraph of subgraph as subgraph subject to registration.

The present invention preferably extracts two width of cloth subgraphs in the following ways:

Choose in the upper same coordinate space of image subject to registration and reference picture (image subject to registration be the reference picture distorted after image) at first that contrast is strong, architectural feature obviously, the identical image of size is as subgraph subject to registration with reference to subgraph.

Subgraph choose the method that adopts entropy and gradient-norm to combine:

First become n (generally the getting n 〉=4) subgraph that individual size is identical with reference to image segmentation, calculate again entropy and the average gradient of each subgraph, and then the subgraph of selecting entropy and average gradient sum maximum chooses a subgraph as effective subgraph subject to registration in image same coordinate subject to registration space as effectively with reference to subgraph.

Entropy can be used as the tolerance of image local area information, is normally defined

$E=-\underset{i=1}{\overset{L}{\mathrm{\Σ}}}P\left(\mathrm{bi}\right)\ln P\left(\mathrm{bi}\right)$

Wherein, the pixel progression of L-subgraph, the probability of P (bi)-i level pixel brightness value.

Average gradient has reflected the readability of image, also reflects simultaneously minor detail contrast and texture transformation feature in the image, and its formula is:

$\▿\stackrel{\‾}{G}=\frac{1}{M\×N}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{[\mathrm{\Δxf}{(i,j)}^2+\mathrm{\Δyf}{(i,j)}^2]}^{1/2}$

Wherein, Δ xf (i, j) and Δ yf (i, j) represent respectively the first order difference of pixel (i, j) on x direction and y direction, and M and N represent respectively line number and the columns of subimage in this formula.

According to entropy and the average gradient of each subgraph in the above-mentioned formula computed image, and get the effective registration subgraph of conduct of entropy and average gradient sum maximum.

The step of registration of back is except final step, and remaining all is to carry out at subgraph.

Step 2, extract the angle point with reference to subgraph and subgraph subject to registration.

Angle point extractive technique that the below uses the present invention, existing describes:

The Harris Corner Detection Algorithm only relates to the first order derivative of image, defines first matrix M:

$M=G\&CircleTimes;\left[\begin{array}{cc}{I_x}^2& I_x{I}_y\\ I_x{I}_y& {I_y}^2\end{array}\right]=\left[\begin{array}{cc}<{I_x}^2>& <I_x{I}_y>\\ <I_x{I}_y>& <{I_y}^2>\end{array}\right]$

I wherein _xGradient for the x direction of image I; I _yGradient for the y direction of image I; G is Gauss's template;＜expression Gauss template function convolution: $<I_x^2>=G\&CircleTimes;I_x^2,$ $<I_y^2>=G\&CircleTimes;I_y^2,<I_x{I}_y>=G\&CircleTimes;I_x{I}_y,$

The expression convolution.The angle point response function CRF definition of adopting Nobel to propose:

$\mathrm{CRF}=\frac{\mathrm{trace}\left(M\right)}{\det \left(M\right)}=\frac{<I_x^2>+<I_y^2>}{<I_x^2>\&CenterDot;<I_y^2>-{<I_x{I}_y>}^2}---\left(16\right)$

Wherein, det is determinant of a matrix; Trace is matrix trace; The Local modulus maxima of CRF is angle point.

More for the angle point number that existing method proposes, excessive, the slow-footed problem of the coupling computation complexity that causes, the method that the present invention takes the non-very big inhibition of neighborhood and total amount to suppress is screened the angle point that initial detecting goes out, and namely the threshold value by CRF is set and the method for ordering are got front topN the last angle point collection R of angle point composition; Then remove being had a few in the sub-image boundary 4* σ among the R (σ is the Gaussian smoothing factor), purpose is that the angle point of reservation is tried one's best at the middle position of entire image, avoid because rotation, translation, convergent-divergent etc. shift out the part angle point, improve the repetition rate of extracting angle point.

Step 3, the angle point that extracts on reference subgraph and the subgraph subject to registration is carried out feature describe, obtain the proper vector of each angle point;

The detailed process of this step is:

Utilize 12 of each angle point to tie up the proper vector that gradient vectors are used as angle point among the present invention, because it is 12 dimension data, therefore can greatly improve registration speed of the present invention.Concrete principle is described as follows:

The nervous physiology that Young (1987) carries out is tested and is shown, human retina and cerebral cortex receptive field section can be simulated with Gaussian derivative.Therefore the present invention's each angle point to trying to achieve is asked its 4 rank Gauss's partial derivatives

G＝[G _x，G _y，G _xx，G _xy，G _yy，G _xxx，G _xxy，G _xyy，G _yyy，G _xxxx，G _xxxy，G _xxyy，G _xyyy，G _yyyy]。

G wherein ₁=[G _x, G _y], ask according to G1

$\cos \left(\mathrm{\&theta;}\right)=\frac{G_x}{\left|G_1\right|}$

$\sin \left(\mathrm{\&theta;}\right)=\frac{G_y}{\left|G_1\right|}$

Rotation matrix:

$L=\left[\begin{array}{cc}\cos \mathrm{\&theta;}& \sin \mathrm{\&theta;}\\ -\sin \mathrm{\&theta;}& \cos \mathrm{\&theta;}\end{array}\right]$

G ₂＝[G _xx，G _xy，G _yy]

The symmetric tensor of G2: g ₂=[G _Xx, G _Xy, G _Xy, G _Yy]

The tensor transition matrix:

$M_2=\left[\begin{array}{cccc}{\cos }^2\mathrm{\&theta;}& \cos \mathrm{\&theta;}\sin \mathrm{\&theta;}& \cos \mathrm{\&theta;}\sin \mathrm{\&theta;}& {\sin }^2\mathrm{\&theta;}\\ -\cos \mathrm{\&theta;}\sin \mathrm{\&theta;}& -{\sin }^2\mathrm{\&theta;}& {\cos }^2\mathrm{\&theta;}& \cos \mathrm{\&theta;}\sin \mathrm{\&theta;}\\ {\sin }^2\mathrm{\&theta;}& -\cos \mathrm{\&theta;}\sin \mathrm{\&theta;}& -\cos \mathrm{\&theta;}\sin \mathrm{\&theta;}& {\cos }^2\mathrm{\&theta;}\end{array}\right]$

D (1: 3)-M ₂* g ₂(Matrix Multiplication)

G ₃＝[G _xxx，G _xxy，G _xyy，G _yyy]

The symmetric tensor of G3: g ₃=[G _Xxx, G _Xxy, G _Xxy, G _Xyy, G _Xxy, G _Xyy, G _Xyy, G _Yyy]

The tensor transition matrix:

$M_3=\left[\begin{array}{cccccccc}{\cos }^3\mathrm{\&theta;}& -{\cos }^2\mathrm{\&theta;}\sin \mathrm{\&theta;}& {\cos }^2\mathrm{\&theta;}\sin \mathrm{\&theta;}& \cos \mathrm{\&theta;}{\sin }^2\mathrm{\&theta;}& {\cos }^2\mathrm{\&theta;}\sin \mathrm{\&theta;}& \cos \mathrm{\&theta;}{\sin }^2\mathrm{\&theta;}& \cos \mathrm{\&theta;}{\sin }^2\mathrm{\&theta;}& -{\sin }^3\mathrm{\&theta;}\\ -{\cos }^2\mathrm{\&theta;}\sin \mathrm{\&theta;}& -\cos \mathrm{\&theta;}{\sin }^2\mathrm{\&theta;}& -\cos \mathrm{\&theta;}{\sin }^2\mathrm{\&theta;}& {\sin }^3\mathrm{\&theta;}& -{\cos }^3\mathrm{\&theta;}& {\cos }^2\mathrm{\&theta;}\sin \mathrm{\&theta;}& {\cos }^2\mathrm{\&theta;}\sin \mathrm{\&theta;}& \cos \mathrm{\&theta;}{\sin }^2\mathrm{\&theta;}\\ \cos \mathrm{\&theta;}{\sin }^2\mathrm{\&theta;}& -{\sin }^3\mathrm{\&theta;}& -{\cos }^2\mathrm{\&theta;}\sin \mathrm{\&theta;}& -\cos \mathrm{\&theta;}{\sin }^2\mathrm{\&theta;}& {\cos }^2\mathrm{\&theta;}\sin \mathrm{\&theta;}& -{\cos }^2\mathrm{\&theta;}\sin \mathrm{\&theta;}& {\cos }^3\mathrm{\&theta;}& -{\cos }^2\mathrm{\&theta;}\sin \mathrm{\&theta;}\\ -{\sin }^3\mathrm{\&theta;}& \cos \mathrm{\&theta;}{\sin }^2\mathrm{\&theta;}& \cos \mathrm{\&theta;}{\sin }^2\mathrm{\&theta;}& {\cos }^2\mathrm{\&theta;}\sin \mathrm{\&theta;}& \cos \mathrm{\&theta;}{\sin }^2\mathrm{\&theta;}& {\cos }^2\mathrm{\&theta;}\sin \mathrm{\&theta;}& {\cos }^2\mathrm{\&theta;}\sin \mathrm{\&theta;}& {\cos }^3\mathrm{\&theta;}\end{array}\right]$

D(4:7)-M ₃*g ₃

The like, obtain D (8:12), with the descriptor of D (1:12) as angle point.Comparing, is 128 dimensions based on its descriptor of method of SIFT, and improved its descriptor of PCA-SIFT method is 36 dimensions.Description of the invention is 12 dimensions, therefore at algorithm efficiency and calculated amount larger minimizing is arranged.

Step 4: carry out similarity measurement and characteristic matching with subgraph subject to registration with reference to the proper vector of angle point on the subgraph, obtain K matching double points;

1) for each angle point p that extracts _i, seek and p _iP contiguous point consists of p _iApart from neighborhood, i=1,2 ... N, N by on two width of cloth subgraphs total number of extraction angle point; Carried out step 1) afterwards, respectively formed altogether N apart from neighborhood at two width of cloth subimages.

2) calculate successively the mahalanobis distance of all the Corner Feature vectors in the reference subgraph of each Corner Feature vector in the subgraph subject to registration, with mahalanobis distance less than setting threshold d _Mth1Two angle points be defined as matching double points, a plurality of matching double points consist of coupling formation 1.

3) in the coupling formation, in 1, reject not at the matching double points in respective distances field, obtain K matching double points;

Detailed process is: find out successively matching double points separately apart from neighborhood no and nm, and form in twos matching double points matrix M (p*p element); In M, utilize coupling formation 1, find out apart from field nm with can a worthy angle point apart among the no of field, and statistics satisfies the angle point number m of matching relationship among the nm, if m 〉=p (p is parameter preset), then define apart from neighborhood no and nm as corresponding apart from the field, and definition is final matching double points apart from the matching double points on neighborhood no and the nm, otherwise reject not matching double points in the respective distances field (namely from coupling formation 1 deletion apart from the matching double points on neighborhood no and the nm), the like, must mate formation 2; K matching double points before in coupling formation 2, getting at last, the K that an obtains matching double points is the higher matching double points of similarity on the whole, K=4 is to matching double points for the better reservation of the present invention, can when guaranteeing matching precision, improve matching speed like this.

The present invention utilizes the affine unchangeability of mahalanobis distance to carry out angle point invariant features similarity measurement.

Mahalanobis distance be solved to prior art, existing it is carried out simple declaration: for putting the sample space (being feature space) that consists of here by n

(t represents transposition, the m representation dimension), wherein any sample point To another sample space $X^2=\{{(x_{11}^2,\&CenterDot;\&CenterDot;\&CenterDot;x_{1m}^2)}^t,\&CenterDot;\&CenterDot;\&CenterDot;,{(x_{\mathrm{nm}}^2,\&CenterDot;\&CenterDot;\&CenterDot;x_{\mathrm{nm}}^2)}^t$ In arbitrary sample point $X_j^2={(x_{j1}^2,\&CenterDot;\&CenterDot;\&CenterDot;x_{\mathrm{jm}}^2)}^t$ Mahalanobis distance be:

$d_M(i,j)=\sqrt{{(X_i^1-X_j^2)}^t{C}^{-1}(X_i^1-X_j^2)}$

Wherein, C represents covariance matrix; C ^-1The inverse matrix of expression C.

Suppose

With

Be respectively subgraph subject to registration and with reference to the proper vector of one group of matching double points between subgraph, be 12 dimension invariant features vectors here.

Because covariance matrix is a real symmetric positive definite matrix, the present invention carries out following decomposition: C to mahalanobis distance ^-1=P ^tDP, P is orthogonal matrix here, D is diagonal matrix, then:

$\begin{array}{c}{d}_M(i,j)=\sqrt{{(X_i^1-X_j^2)}^t{P}^t\sqrt{D}\&CenterDot;\sqrt{D}P(X_i^1-X_j^2)}\\ =d_E(\sqrt{D}P{X}_i^1,\sqrt{D}P{X}_j^2)\end{array}---\left(18\right)$

Calculate two angle points according to following formula

With

Corresponding mahalanobis distance, the method is with mahalanobis distance d _MBe converted to simple Euclidean distance d _ECalculating, when hardware is realized, can avoid redundant inversion operation, reduce the space requirement to internal memory, improve the real-time of algorithm.

Different from the method for traditional search matching strategy, the algorithm of this Binding distance neighborhood and threshold value has been realized the second degree matches process of " by thick to smart ", so that search volume and calculated amount dynamically dwindle with the variation of topN threshold value, so that matching rate is higher.

Step 5, determined subgraph subject to registration and with reference to subgraph in behind the matching relationship between the angle point, based on K matching double points, adopt least-squares algorithm, transformation matrix H between direct computational transformation image subject to registration and the reference picture, based on transformation matrix H, adopt bilinear interpolation to treat registering images and rebuild afterwards.

For example: establish (x _i, y _i) and (x _i', y _i') (i=1,2,3,4) be respectively upper final coupling point set A and the B of image A subject to registration and template image B; Utilize the method for least square method estimation affine transformation parameter to be:

A*H＝B

$A=\left[\begin{array}{cccccc}{x}_1& y_1& 0& 0& 1& 0\\ x_2& y_2& 0& 0& 1& 0\\ x_3& y_3& 0& 0& 1& 0\\ x_4& y_4& 0& 0& 1& 0\\ 0& 0& x_1& y_1& 0& 1\\ 0& 0& x_2& y_2& 0& 1\\ 0& 0& x_3& y_3& 0& 1\\ 0& 0& x_4& y_4& 0& 1\end{array}\right]$

B＝[x ₁′；x ₂′；x ₃′；x ₄′；y ₁′；y ₂′；y ₃′y ₄′；]

Therefore, H=pinv (A) * B

As seen, the present invention is by reasonably choosing subgraph, and in feature extraction, invariant features is described, and the feature-based registration method is improved in the aspects such as similarity measurement, and effective subgraph that these methods are applied to extract, thus the faster registration that carries out image more accurately.

In sum, above is preferred embodiment of the present invention only, is not for limiting protection scope of the present invention.Within the spirit and principles in the present invention all, any modification of doing, be equal to replacement, improvement etc., all should be included within protection scope of the present invention.In sum, above is preferred embodiment of the present invention only, is not for limiting protection scope of the present invention.Within the spirit and principles in the present invention all, any modification of doing, be equal to replacement, improvement etc., all should be included within protection scope of the present invention.

Claims (9)

1. the fast image registration method based on the subimage Corner Feature is characterized in that, concrete steps comprise:

Step 1, choose with reference to subgraph and subgraph subject to registration;

From reference picture, choose a subgraph as the reference subgraph, from image subject to registration, choose a coordinate space with reference to the identical subgraph of subgraph as subgraph subject to registration;

Step 2, extract the angle point with reference to subgraph and subgraph subject to registration;

Step 3, the angle point that extracts on reference subgraph and the subgraph subject to registration is carried out feature describe, obtain the proper vector of each angle point;

Step 4, carry out similarity measurement and characteristic matching with subgraph to be matched with reference to the proper vector of angle point on the subgraph, finally obtain K matching double points;

The detailed process of this step is:

1) for each angle point p that extracts _i, seek and p _iP contiguous point consists of p _iApart from neighborhood, i=1,2 ... N, N by on two width of cloth subgraphs total number of extraction angle point;

2) calculate successively the mahalanobis distance of all the Corner Feature vectors in the reference subgraph of each Corner Feature vector in the subgraph subject to registration, with mahalanobis distance less than setting threshold d _Mth1Two angle points be defined as matching double points, a plurality of matching double points consist of coupling formation 1;

3) in coupling formation 1, reject not at the matching double points in respective distances field, obtain K matching double points;

Step 5, based on K matching double points, adopt least square method to calculate transformation matrix H between image subject to registration and the reference picture, utilize described transformation matrix H with image registration subject to registration to reference picture.

2. described fast image registration method based on the subimage Corner Feature according to claim 1 is characterized in that, described is strong, the obvious width of cloth subgraph of architectural feature of contrast on the reference picture with reference to subgraph.

3. described fast image registration method based on the subimage Corner Feature according to claim 2 is characterized in that, the described process of choosing with reference to subgraph is:

At first become n the subgraph that size is identical with reference to image segmentation, next calculates entropy and the average gradient of each subgraph, then selects the subgraph of entropy and average gradient sum maximum as the reference subgraph.

4. described fast image registration method based on the subimage Corner Feature according to claim 3 is characterized in that described n 〉=4.

5. described fast image registration method based on the subimage Corner Feature according to claim 1 and 2 is characterized in that, extracts identically with the method for subgraph angle point subject to registration with reference to subgraph, and detailed process is:

At first, based on the Harris angular-point detection method, detect the subgraph angle point; Secondly, the method for taking the non-very big inhibition of neighborhood and total amount to suppress is screened the angle point that initial detecting goes out, and extracts the top n angle point; Then, remove the angle point that is positioned on the subgraph borderline region, thereby extract required angle point.

6. described fast image registration method based on the subimage Corner Feature according to claim 5 is characterized in that, described subgraph borderline region is for being the rectangular area of σ apart from border width, and wherein σ is the Gaussian smoothing factor.

7. described fast image registration method based on the subimage Corner Feature according to claim 1 is characterized in that, described proper vector is 12 dimension gradient vectors of angle point.

8. described fast image registration method based on the subimage Corner Feature according to claim 1 is characterized in that mahalanobis distance d _M(i, j) utilizes orthogonal matrix P and diagonal matrix D to represent the inverse matrix C of covariance matrix C in the process of calculating ^-1, convert the calculating of mahalanobis distance to Euclidean distance d _FCalculating;

$d_M(i,j)=\sqrt{{(X_i^1-X_j^2)}^t{C}^{-1}(X_i^1-X_j^2)}$

$=\sqrt{{(X_i^1-X_j^2)}^t{P}^t\sqrt{D}\&CenterDot;\sqrt{D}P(X_i^1-X_j^2)}$

$=d_E(\sqrt{D}{\mathrm{PX}}_i^1,\sqrt{D}P{X}_j^2)$

Wherein,

The proper vector that represents i angle point on the subgraph subject to registration, Expression is with reference to the proper vector of j angle point on the subgraph.

9. described fast image registration method based on the subimage Corner Feature according to claim 1 is characterized in that the number K=4 of described matching double points.

Images