CN101315698A - Characteristic matching method based on straight line characteristic image registration - Google Patents

Abstract

The invention discloses a feature matching method in image registration based on straight-line features, aiming at solving the feature matching problem in image registration based on straight-line features; the technical proposal of the invention comprises the following steps of: firstly, extracting straight-line grouping from a reference image and an image to be registered as a registration primitive and defining a matching matrix M; then calculating reliability measure, similarity measure and consistency measure of spatial relationship between the straight-line grouping of the two images and obtaining a reliability measure matrix R, a similarity measure matrix B and a consistency measure matrix S of spatial relationship; then establishing an energy function which takes the M as argument and transforming the matching problem of straight-line grouping into the optimization problem of the energy function; finally adopting a two-step iterative optimization algorithm to carry out iterative optimization for the optimization problem of the energy function; the invention combines the reliability, the consistency of the spatial relationship and the similarity of the straight-line grouping organically in the process of feature matching, overcomes the defects that in the existing method, the calculation is complex, the wrong matching is easy to occur and the obtaining of corresponding image points is not accurate.

Description

Based on the feature matching method in the straight line characteristic image registration

Technical field

The invention belongs to the method for registering images of technical field of image processing, especially the method for mating based on the linear feature that extracts in two width of cloth images in the process of image registration.

Background technology

Image registration is the important module in various remote sensing image processing softwares, the Medical Image Processing software, it is one step of key during remote sensing image fusion, environmental surveillance, change-detection, image mosaic, weather forecasting and Medical Image Processing are used, purpose is that the same region that different times, different points of view or different sensors are obtained or two width of cloth or the multiple image of object carry out geometry correction, makes them reach consistance on the locus.Registration problems between the multiple image often can resolve into the registration problems between a plurality of two width of cloth images, therefore, the image registration problem mainly all is the geometry correction problem between research two width of cloth images, and wherein piece image is called reference picture, and another width of cloth image is called image subject to registration.Because the importance of image registration, in the last thirty years, people have carried out research extensively and profoundly to it, have obtained a lot of important achievements, have proposed a large amount of methods.

Existing method for registering images mainly is divided into two big classes: based on the method in zone with based on the method for feature.

Based on the method in zone is to get image block onesize in the image block of a wicket and the reference picture to add up the ground relevance ratio from image subject to registration, similarity measure (is generally adopted quadratic sum, related coefficient of gray scale difference etc., also may be some attributes of transform domain, as Fourier transform, wavelet transformation etc.) center of maximum two windows is as a pair of corresponding image points, utilize all corresponding image points and the least-squares algorithm that obtain to estimate the transformation model parameter vector then, finish registration.Because the similarity measure that adopts is general relevant with image radiation characteristic and transformation model, thereby be not suitable for registration between the two bigger width of cloth images of radiation characteristic and geometric deformation based on the method in zone.

Method based on feature generally includes three steps: extract feature at first respectively from reference picture and image subject to registration, these features generally are spatial features, for example point, edge, profile, straight-line segment, zone etc.; The attribute information that utilize to extract feature then carries out characteristic matching, by the feature of coupling to obtaining corresponding image points; Estimate the transformation model parameter vector according to these corresponding image points and least-squares algorithm at last, finish registration.Follow based on the method in zone and compare, not too responsive based on the method for feature to the image radiation distortion, can be adapted to the registration between the different sensor images, become the emphasis and the focus of image registration techniques research.

Key and difficult point based on the method for registering images of feature are the character pair of how seeking in reference picture and the image subject to registration, i.e. characteristic matching problem.Proposed multiple strategy in the existing document and solved this problem, mainly be divided into and utilize spatial relationship, utilize characteristic similarity, utilize sequential interblock space relation and these three class methods of characteristic similarity.The method of utilizing spatial relationship is at first to extract one or more reference mark (as the mid point of straight line, the center of gravity in zone, the projecting point on the profile etc.) from each feature that two width of cloth images extract, and seeks the optimal mapping model parameter vector of each reference mark when spatial relationship reaches global coherency in two width of cloth images then.The great advantage of these class methods is generally can both obtain registration results preferably, but its greatest problem is a calculation of complex, require the correct matching characteristic number must be simultaneously greater than the number of features of any class situation not in the know, could obtain reliable result, and along with the minimizing of correct matching characteristic quantity, the complexity of algorithm will significantly increase, and the success ratio of registration but greatly descends.The method of utilizing characteristic similarity is each feature that at first adopts in one group of parametric description, two width of cloth images, utilizes the similarity measure between these calculation of parameter different characteristics then, and it is right to seek matching characteristic according to the principle of similarity measure maximum.The feature description parameter generally will satisfy several conditions such as unchangeability, uniqueness, stability and independence, and feature description parameter commonly used has related coefficient, square invariant, form matrix, chain code etc.The characteristics of these class methods are to calculate simply, but when having the identical or close feature of a plurality of shapes in the image, cause the mistake coupling between the feature easily, make registration fail.Just because of utilize spatial relationship or characteristic similarity to carry out characteristic matching separately limitation is arranged all, people expect making up this two kinds of matching criterior naturally, Here it is utilizes the method for sequential interblock space relation and characteristic similarity, this method at first utilizes characteristic similarity to reject some unlikely characteristics combination of coupling, to reduce search volume, utilize spatial relationship to obtain final characteristic matching result then based on the spatial relationship method.These class methods combine the advantage of preceding two class methods, on the property taken into account of registration speed and precision, be better than preceding two class methods, but in fact this sequential array mode is still a kind of method based on spatial relationship, can not overcome the inherent shortcoming based on spatial relationship: promptly along with the minimizing of correct matching characteristic quantity in two width of cloth images, the complexity of algorithm will significantly increase, the success ratio of registration but greatly descends.

In addition, the existing overwhelming majority treats each feature to be matched based on the method for registering images of feature in the characteristic matching stage with being equal to, does not consider feature extraction result's reliability.Because aspects such as picture noise and feature extraction algorithms, more or less there are some uncertainties (inaccurate as the position extraction, feature extraction is imperfect etc.) in the feature extraction result.Like this,, will cause the registration results precision that obtains not high if some insecure features are estimated the transformation model parameter vector as matching characteristic, even the registration failure.

Key based on the method for registering images of feature is to obtain corresponding image points in the point, line, surface feature of extracting from reference picture and image subject to registration.Point feature itself can be used as corresponding image points, thereby it becomes the most frequently used a kind of feature in the method for registering images based on feature, but when image resolution ratio is low, the point feature is often located inaccurate, be not suitable for high-precision image registration, and the quantity of information that the some feature contains is limited, and coupling is also relatively more difficult.Comparatively speaking, the line feature is considered to be more suitable for being used for image registration, and it contains relatively abundanter quantity of information on the one hand, is beneficial to coupling; On the other hand, often contain a large amount of line features in the image of man-made target.Although the face feature also contains abundant quantity of information, be not always to contain the face feature in the image, and the face feature usually can be with a series of line character representation.Therefore, the line feature is based on the preferred features in the method for registering images of feature.Most existing often all is directly linear feature to be mated as the registration primitive based on the method for registering of linear feature, select for use then the coupling linear feature to the mid point of (in reference picture and the image subject to registration corresponding two rectilinear(-al)s) as corresponding image points, when extraction of straight line imperfect (endpoint location that is straight line extracts inaccurate), the corresponding image points of Huo Deing is incorrect often like this.

Summary of the invention

The technical problem to be solved in the present invention is: propose a kind of new linear feature matching process, overcome existing method calculation of complex, cause mistake coupling and corresponding image points to obtain inaccurate defective easily, solve based on the characteristic matching problem in the straight line characteristic image registration.

Technical scheme of the present invention is:

The first step is extracted the straight line marshalling respectively, as the registration primitive from reference picture and image subject to registration.

The general image that obtains from objective scenery all is two-dimentional, can represent with a two-dimensional array.Reference picture and image subject to registration are designated as I respectively ₁(x ₁, y ₁) and I ₂(x ₂, y ₂), they represent the gray-scale value of two width of cloth images, wherein (x respectively ₁, y ₁), (x ₂, y ₂) represent the pixel coordinate in two width of cloth images respectively.Adopt sane straight line extracting method to extract N from reference picture ₁The bar linear feature extracts N from image subject to registration ₂The bar linear feature is designated as respectively ${\mathrm{<figure-callout\; id="1"\; label="L"\; filenames="A20081003157500252.png,A20081003157500262.png"\; state="\{\{state\}\}">L</figure-callout>}}_1=\{{\mathrm{<figure-callout\; id="1"\; label="l"\; filenames="A20081003157500252.png,A20081003157500262.png"\; state="\{\{state\}\}">l</figure-callout>}}_1^1,{\mathrm{<figure-callout\; id="1"\; label="l"\; filenames="A20081003157500252.png,A20081003157500262.png"\; state="\{\{state\}\}">l</figure-callout>}}_1^2,\&CenterDot;\&CenterDot;\&CenterDot;,l_1^{N_1}\}$ With $L_2=\{l_2^1,l_2^2,\&CenterDot;\&CenterDot;\&CenterDot;,l_2^{N_2}\},$ N wherein ₁, N ₂(concrete straight line extracting method is referring to Wen Gongjian, Wang Runsheng to be positive integer.A kind of sane straight line extraction algorithm.The software journal, 12 (11): 1660-1667,2001.).

As described in the background art, directly utilize linear feature to mate, the corresponding image points that obtains when extraction of straight line is imperfect is often inaccurate.In order to overcome this shortcoming, the present invention utilizes the straight line marshalling (by L ₁Or L ₂In a straight line forming of any two straight lines to) mate as the registration primitive.If directly any two linear features that extract in reference picture and the image subject to registration are all formed a straight line marshalling, the straight line marshalling quantity of acquisition is too big, has increased the calculated amount of follow-up coupling.For this reason, only select those straight line that satisfies certain constraint condition marshallings as the registration primitive, constraint condition is: the extraction intersection point of straight line marshalling to the Hausdroff between two straight lines forming it apart from sum less than a default thresholding T, T gets 30 pixels of 10.

Extracted N ' in the hypothetical reference image ₁Individual straight line marshalling has been extracted N ' in the image subject to registration ₂Individual straight line marshalling is designated as respectively ${\mathrm{<figure-callout\; id="1"\; label="Z"\; filenames="A20081003157500252.png,A20081003157500262.png"\; state="\{\{state\}\}">Z</figure-callout>}}_1=\{{\mathrm{<figure-callout\; id="1"\; label="z"\; filenames="A20081003157500252.png,A20081003157500262.png"\; state="\{\{state\}\}">z</figure-callout>}}_1^1,{\mathrm{<figure-callout\; id="1"\; label="z"\; filenames="A20081003157500252.png,A20081003157500262.png"\; state="\{\{state\}\}">z</figure-callout>}}_1^2,\&CenterDot;\&CenterDot;\&CenterDot;,z_1^{{\mathrm{<figure-callout\; id="1"\; label="N"\; filenames="A20081003157500252.png,A20081003157500262.png"\; state="\{\{state\}\}">N</figure-callout>}}_1^{\&prime;}}\}$ With $Z_2=\{z_2^1,z_2^2,\&CenterDot;\&CenterDot;\&CenterDot;,z_2^{N_2^{\&prime;}}\},$ N ' wherein ₁, N ' ₂Be positive integer.Defining a size is N ' ₁* N ' ₂The coupling matrix M represent matching relationship between the marshalling of reference picture and image cathetus subject to registration, promptly

Because a straight line marshalling in the reality in the reference picture can only be complementary with a straight line marshalling in the image subject to registration, therefore mating matrix M should satisfy matching criterior one by one, promptly

$\left\{\begin{array}{c}\underset{j=1}{\overset{N_2^{\&prime;}}{\mathrm{\&Sigma;}}}{M}_{\mathrm{ij}}\&le;1(1\&le;i\&le;{\mathrm{<figure-callout\; id="1"\; label="N"\; filenames="A20081003157500252.png,A20081003157500262.png"\; state="\{\{state\}\}">N</figure-callout>}}_1^{\&prime;})\\ \underset{i=1}{\overset{{\mathrm{<figure-callout\; id="1"\; label="N"\; filenames="A20081003157500252.png,A20081003157500262.png"\; state="\{\{state\}\}">N</figure-callout>}}_1^{\&prime;}}{\mathrm{\&Sigma;}}}{M}_{\mathrm{ij}}\&le;1(1\&le;j\&le;N_2^{\&prime;})\end{array}\right.---\left(2\right)$

Therefore solve just become how to obtain mating matrix M based on the characteristic matching problem in the straight line characteristic image registration.Use parameter $V=[{\mathrm{<figure-callout\; id="1"\; label="l\; H\; k"\; filenames="A20081003157500252.png,A20081003157500262.png"\; state="\{\{state\}\}">l</figure-callout>}}_{Hk}^1,l_H^{k2},{\mathrm{<figure-callout\; id="1"\; label="length\; H\; k"\; filenames="A20081003157500252.png,A20081003157500262.png"\; state="\{\{state\}\}">length</figure-callout>}}_{Hk}^1,{\mathrm{length}}_H^{k2},{\mathrm{<figure-callout\; id="1"\; label="S\; H\; k"\; filenames="A20081003157500252.png,A20081003157500262.png"\; state="\{\{state\}\}">S</figure-callout>}}_{Hk}^1,S_H^{k2},{\mathrm{<figure-callout\; id="1"\; label="E\; H\; k"\; filenames="A20081003157500252.png,A20081003157500262.png"\; state="\{\{state\}\}">E</figure-callout>}}_{Hk}^1,E_H^{k2},{\mathrm{<figure-callout\; id="1"\; label="M\; H\; k"\; filenames="A20081003157500252.png,A20081003157500262.png"\; state="\{\{state\}\}">M</figure-callout>}}_{Hk}^1,M_H^{k2},O_H^k,x_H^k,y_H^k]$ Any one straight line marshalling z in reference picture and the image subject to registration is described _H ^k(H=1, k=1,2 ..., N ' ₁H=2, k=1,2 ..., N ' ₂) attribute, l wherein _H ^K1, l _H ^K2Straight line marshalling z is formed in expression _H ^kTwo linear features, length _H ^K1, length _H ^K2, S _H ^K1, S _H ^K2, E _H ^K1, E _H ^K2, M _H ^K1, M _H ^K2Represent their length, four end points and two mid points respectively, O _H ^kBe the extraction intersection point of straight line marshalling, x _H ^k, y _H ^kBe to extract intersecting point coordinate.

Second step, calculate the reliability measure between the marshalling of reference picture and image cathetus subject to registration, obtain reliability measure matrix R.

The reliability of straight line marshalling is by the reliability decision of two linear features forming it.As what mention in the background technology, owing to picture noise and feature extraction algorithm itself, the extracting position of linear feature and its physical location may be inconsistent in the reality, and this inconsistency is mainly reflected in two aspects: end points and direction.Under the certain situation of direction, the endpoint location of two linear features is to the not influence of extracting position of straight line marshalling intersection point, and therefore, direction skew that need only the calculated line feature is to the influence of straight line marshalling intersection point extracting position.

For L ₁Or L ₂In any straight line feature l _H ^m(H=1, m=1,2 ..., N ₁H=2, m=1,2 ..., N ₂), supposed l _H ^mMid point M _H ^mAnd l _H ^mAngle is Δ θ _H ^mTwo dotted lines are l _H ^mThe maximum direction that may depart from, Δ θ _H ^mBe called the direction skew.Like this, l _H ^mActual position only may drop between two dotted lines.Generally speaking, linear feature length is long more, and it is sane more that it extracts the result, i.e. direction shifted by delta θ _H ^mMore little, when linear feature reached certain-length, it extracted the result direction skew does not take place, promptly $\mathrm{\&Delta;}{\mathrm{\&theta;}}_H^m=0;$ On the other hand, the direction shifted by delta θ of linear feature _H ^mTend to less than certain scope.Therefore, can be with linear feature l _H ^mDirection shifted by delta θ _H ^mBe defined as:

$\mathrm{\&Delta;}{\mathrm{\&theta;}}_H^m=\mathrm{\&alpha;exp}[-\frac{{\left({\mathrm{length}}_H^m\right)}^2}{{{2\mathrm{\&sigma;}}_1}^2}]---\left(3\right)$

Wherein α is a default thresholding, is used for limiting direction shifted by delta θ _H ^mMaximum magnitude, generally get α=0.175 radian, i.e. the direction of linear feature skew generally is no more than 10 degree; Length _H ^mBe l _H ^mLength; σ ₁Be the variance of Gaussian function, the direction skew will not take place in its decision linear feature how long, generally get 20 40, specifically depend on the resolution of image, and image resolution ratio is high more, σ ₁The closer to 40.

For any one straight line marshalling z in reference picture or the image subject to registration _H ^k, z _H ^kBy linear feature l _H ^K1, l _H ^K2Form, cross l _H ^K1Mid point M _H ^K1And l _H ^K1Angle is Δ θ _H ^K1Two dotted lines are l _H ^K1L is crossed in contingent maximum direction skew _H ^K2Mid point M _H ^K2And l _H ^K2Angle is Δ θ _H ^K2Two dotted lines are l _H ^K2Contingent maximum direction skew, four dotted lines meet at A, B, C and D 4 points.Δ θ _H ^K1With Δ θ _H ^K2Be linear feature l _H ^K1And l _H ^K2Direction skew, their size is determined by formula (3).Like this, linear feature l _H ^K1And l _H ^K2Actual position can only drop on them separately between two dotted lines of mid point, so straight line marshalling z _H ^kTrue intersection point Q _H ^kWill drop on by A B, the quadrilateral r that C and D surround at 4 _ABCDIn, and may drop on r _ABCDOn interior any position.Therefore, z _H ^kTrue intersection point Q _H ^kBe a stochastic variable, the probability density function of remembering it is f (Q _H ^k).Use d _H ^kExpression z _H ^kTrue intersection point Q _H ^kWith extraction intersection point O _H ^kBetween distance, d then _H ^kAlso be a stochastic variable, note d _H ^kExpectation be expressed as d _H ^k, then

${\stackrel{\&OverBar;}{d}}_H^k=E\left(d_H^k\right)=E\left(\right||O_H^k-Q^H|\left|\right)={\&Integral;}_{r_{\mathrm{ABCD}}}\left|\right|O_H^k-Q_H^k\left|\right|f\left(Q_H^k\right)d{Q}_H^k---\left(4\right)$

Wherein E () expression is to asking expectation, and ‖ ‖ represents Euclidean distance, during H=1, and k=1,2 ..., N ' ₁, during H=2, k=1,2 ..., N ' ₂z _H ^kTrue intersection point Q _H ^kCan be Gaussian distribution, even distribution etc., the present invention adopts even distribution, promptly

$f\left(Q_H^k\right)=\frac{1}{S_{r_{\mathrm{ABCD}}}}---\left(5\right)$

Wherein

Expression r _ABCDArea, its big I is according to crossing l _H ^K1Mid point M _H ^K1With mistake l _H ^K2Mid point M _H ^K2Four dotted line equations obtain, four dotted line equations are by Δ θ _H ^K1With Δ θ _H ^K2And the linear feature l that extracts _H ^K1, l _H ^K2Obtain.Like this, (4) formula can be write as

${\stackrel{\&OverBar;}{d}}_H^k=\frac{1}{S_{r_{\mathrm{ABCD}}}}{\&Integral;}_{r_{\mathrm{ABCD}}}\left|\right|O_H^k-Q_H^k\left|\right|d{Q}_H^k---\left(6\right)$

Straight line is organized into groups z _H ^kReliability be designated as c _H ^k, it and d _H ^kRelevant, d _H ^kMore little, z _H ^kReliable more, i.e. c _H ^kBig more.c _H ^kThe employing following formula is tried to achieve:

$c_H^k=\exp [-\frac{{\left({\stackrel{\&OverBar;}{d}}_H^k\right)}^2}{{{2\mathrm{\&sigma;}}_2}^2}]---\left(7\right)$

σ wherein ₂Be the variance of Gaussian function, generally get 24 in the reality.

The reliability measure matrix R of straight line marshalling is that a size is N ' ₁* N ' ₂Matrix, it each R _IjThe marshalling of expression straight line is to LP _IjReliability measure, LP _IjBe Z ₁In any one straight line marshalling z ₁ ⁱWith Z ₂Any one straight line marshalling z ₂ ^jThe straight line marshalling of forming is right.Generally speaking, if form the straight line marshalling to LP _IjThe marshalling of two straight lines all be reliably, i.e. straight line marshalling z ₁ ⁱReliability c ₁ ⁱWith straight line marshalling z ₂ ^jReliability c ₂ ^jAll bigger, LP then _IjAlso be reliably, i.e. R _IjAlso bigger.Therefore, can adopt D-S evidence theory method, to c ₁ ⁱAnd c ₂ ^jMerge calculating, obtain R _Ij, promptly

$R_{\mathrm{ij}}=\frac{c_1^i\&times;c_2^j}{c_1^i\&times;c_2^j+(1-c_1^i)(1-c_2^j)}---\left(8\right)$

As seen, reliability measure matrix R can calculated in advance obtain, and is a constant matrices.

The 3rd step, calculate the similarity measure between the marshalling of reference picture and image cathetus subject to registration, obtain the similarity measure matrix B.

The similarity measure matrix B is that a size is N ' ₁* N ' ₂Matrix, it each B _IjAny two straight lines marshalling z in expression reference picture and the image subject to registration ₁ ⁱAnd z ₂ ^jSimilarity measure.Similarity measure is represented two similarity degrees between the straight line marshalling, and the general features marshalling is similar more, and they just may mate more.Similarity measure B _IjBe defined as

$B_{\mathrm{ij}}=\left|\right|\frac{{\mathrm{length}}_1^{i1}-{\mathrm{length}}_1^{i2}}{{\mathrm{length}}_1^{i1}+{\mathrm{length}}_1^{i2}}|-|\frac{{\mathrm{length}}_2^{j1}-{\mathrm{length}}_2^{j2}}{{\mathrm{length}}_2^{\mathrm{<figure-callout\; id="1"\; label="j"\; filenames="A20081003157500252.png,A20081003157500262.png"\; state="\{\{state\}\}">j</figure-callout>}1}+{\mathrm{length}}_2^{j2}}\left|\right|---\left(9\right)$

By (9) formula as can be known, B _IjMore little expression straight line marshalling z ₁ ⁱAnd z ₂ ^jSimilar more, promptly the possibility of their couplings is also big more.For and straight line marshalling between reliability measure be consistent, i.e. similarity measure between the straight line marshalling is big more, two straight lines marshallings may be mated more, and require its value to drop between the 0-1, therefore need carry out conversion to (9) formula, promptly

B _ij＝F ₁(B _ij) (10)

F wherein ₁() is a monotonic decreasing function, adopts Gaussian function, promptly

$F_1\left(x\right)=\exp (-\frac{x^2}{{2\mathrm{\&sigma;}}_3^2})---\left(11\right)$

σ wherein ₃Be the variance of Gaussian function, generally get σ in practice ₃=1.With B _IjSubstitution (11) formula,

$B_{\mathrm{ij}}=\exp (-\frac{{B_{\mathrm{ij}}}^2}{2{\mathrm{\&sigma;}}_3^2})---\left(12\right)$

As seen, the similarity measure matrix B also can calculated in advance obtain, and also is a constant matrices.

In the 4th step, the spatial relationship consistance of calculating between the marshalling of reference picture and image cathetus subject to registration is estimated, and obtains the spatial relationship consistance and estimates matrix S.

Geometry between the straight line that spatial relationship consistance between the straight line marshalling is come with conversion from image subject to registration with regard to the straight line marshalling that is meant in the reference picture is organized into groups closes on degree.The Euclidean distance of extraction intersection point under the transformation model constraint that the present invention adopts straight line to organize into groups measured the spatial relationship consistance between them.

It also is big or small N ' that the spatial relationship consistance is estimated matrix S ₁* N ' ₂Matrix, it each S _IjAny two straight lines marshalling z in expression reference picture and the image subject to registration ₁ ⁱAnd z ₂ ^jBetween the spatial relationship consistance estimate.S _IjCalculate by following formula

$S_{\mathrm{ij}}=\sqrt{{(x_1^i-T_x\left(x_2^j\right))}^2+{(y_1^i-T_y\left(y_2^j\right))}^2}---\left(13\right)$

T wherein _x(), T _y() represents the transforming function transformation function of transformation model on x direction and y direction between reference picture and the image subject to registration respectively.Transformation model between reference picture and the image subject to registration can be described with one group of transformation model parameter vector P usually, and P is relevant with the transformation model type of employing, and for example, when adopting the similarity transformation model, P is 4 dimensional vectors; When adopting the affined transformation model, P is 6 dimensional vectors; When adopting the projective transformation model, P is 8 dimensional vectors.

For and straight line marshalling between reliability measure, similarity measure be consistent, promptly the spatial relationship consistance is estimated greatly more, two straight lines marshallings may be mated more, need do a conversion to (13) formula, the present invention also adopts the Gaussian function of (11) formula, promptly

S _ij＝F ₁(S _ij) (14)

With S _IjSubstitution (14) formula,

$S_{\mathrm{ij}}=\exp (-\frac{{S_{\mathrm{ij}}}^2}{2{\mathrm{\&sigma;}}_3^2})---\left(15\right)$

The spatial relationship consistance is estimated matrix S and T _x(), T _y() is relevant, and be therefore relevant with transformation model parameter vector P.

In the 5th step, set up energy function

Any two straight lines marshalling in reference picture and the image subject to registration is if correct coupling, and according to top computing method, then estimate all should be bigger for the reliability measure between them, similarity measure and spatial relationship consistance.For this reason, be defined as follows energy function:

$E(M,P)=\underset{i=1}{\overset{{\mathrm{<figure-callout\; id="1"\; label="N"\; filenames="A20081003157500252.png,A20081003157500262.png"\; state="\{\{state\}\}">N</figure-callout>}}_1^{\&prime;}}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N_2^{\&prime;}}{\mathrm{\&Sigma;}}}{M}_{\mathrm{ij}}{R}_{\mathrm{ij}}{B}_{\mathrm{ij}}{S}_{\mathrm{ij}}---\left(16\right)$

If it is right to comprise the coupling straight line marshalling of mistake in the coupling matrix M, it will be inaccurate or even wrong then utilizing the corresponding image points of M acquisition and the transformation model parameter vector P of least-squares algorithm calculating, at this moment the straight line of most of actual match marshalling between the spatial relationship consistance estimate and will significantly descend, cause the E () that calculates by this M littler than the E () that calculates by the right M of coupling straight line marshalling that does not comprise mistake.On the other hand, if it is right to add correct coupling straight line marshalling in M, then utilize the corresponding image points of M acquisition and the transformation model parameter vector P of least-squares algorithm calculating to change very little, therefore, although the right spatial relationship consistance of some correct coupling straight line marshallings is estimated and may be reduced to some extent, but owing to increased new item in E (), the value of E () still can increase.Therefore, straight lines marshalling that has comprised all correct couplings when the coupling matrix M to and to have rejected the straight line marshalling of all erroneous matching right, then the value of E () will reach a global maximum, at this moment coupling matrix M is optimum, transformation model parameter vector P by the Optimum Matching matrix computations also is optimum simultaneously, promptly

<M _opt，P _opt>＝arg?max(E(·)) (17)

M wherein _Opt, P _OptRepresent optimum coupling matrix and optimum transformation model parameter vector respectively, arg max () represents maximizing.Like this, the matching problem of straight line marshalling just is transformed into an energy function optimization problem.

In the 6th step, the optimization problem that adopts two step iteration optimization algorithms that (17) formula is described is carried out iteration optimization, obtains M _OptAnd P _OptComprise following step:

Step 1: initialization transformation model parameter vector P

Adopt a sequential assemblage characteristic similarity and the conforming method of spatial relationship to come initialization transformation model parameter vector P, method is: at first according to characteristic similarity estimate the minimized number 3-5 straight line marshalling doubly of selecting to find the solution the transformation model needs to as the marshalling of candidate matches straight line to (as for the affined transformation model, it is right to find the solution the marshalling of 3 pairs of couplings of the minimum needs of transformation model parameter vector straight line, so just select the marshalling of the right straight line of 9-15 to), it is right to estimate the minimized number Optimum Matching straight line marshalling that utilizes the method for exhaustion or RANSAC algorithm to obtain to find the solution transformation model parameter vector P needs according to the spatial relationship consistance then, utilizes their extraction intersection point to obtain the initial value P of transformation model parameter vector P ₀

Step 2: initialization controlled variable β=β ₀

β ₀, β _rAnd β _MaxBe respectively initial value, rate of growth and the maximal value of controlled variable β, generally get β ₀=0.5, β _rGet 1.05 1.1, β _Max=10.

Step 3: known transform model parameter vector P, adopt soft assigning method to calculate the coupling matrix M

Known transform model parameter vector P, the spatial relationship consistance is estimated matrix S and is obtained in advance by formula (15), and formula (16) just becomes like this

$E\left(M\right)=\underset{i=1}{\overset{{\mathrm{<figure-callout\; id="1"\; label="N"\; filenames="A20081003157500252.png,A20081003157500262.png"\; state="\{\{state\}\}">N</figure-callout>}}_1^{\&prime;}}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N_2^{\&prime;}}{\mathrm{\&Sigma;}}}{M}_{\mathrm{ij}}{R}_{\mathrm{ij}}{B}_{\mathrm{ij}}{S}_{\mathrm{ij}}=\underset{i=1}{\overset{{\mathrm{<figure-callout\; id="1"\; label="N"\; filenames="A20081003157500252.png,A20081003157500262.png"\; state="\{\{state\}\}">N</figure-callout>}}_1^{\&prime;}}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N_2^{\&prime;}}{\mathrm{\&Sigma;}}}{M}_{\mathrm{ij}}{W}_{\mathrm{ij}}---\left(18\right)$

W wherein _Ij=R _IjB _IjS _IjIt is a constant.

The greatest difficulty of finding the solution (17) formula is exactly the two-way constraint that it must satisfy (1), (2) formula.In order not wait constraint to become equated constraint, will mate matrix M increases delegation and row, i.e. the marshalling of straight line in reference picture z ₁ ⁱWhen getting along well any one straight line marshalling coupling in the image subject to registration, $M_{i(N_2^{\&prime;}+1)}=1;$ Straight line marshalling z in image subject to registration ₂ ^jWhen any one the straight line marshalling in the discord reference picture is mated, $M_{({\mathrm{<figure-callout\; id="1"\; label="N"\; filenames="A20081003157500252.png,A20081003157500262.png"\; state="\{\{state\}\}">N</figure-callout>}}_1^{\&prime;}+1)j}=1.$ Like this, (2) formula can be changed into

$\left\{\begin{array}{c}\underset{j=1}{\overset{N_2^{\&prime;}+1}{\mathrm{\&Sigma;}}}{M}_{\mathrm{ij}}=1(1\&le;i\&le;{\mathrm{<figure-callout\; id="1"\; label="N"\; filenames="A20081003157500252.png,A20081003157500262.png"\; state="\{\{state\}\}">N</figure-callout>}}_1^{\&prime;})\\ \underset{i=1}{\overset{{\mathrm{<figure-callout\; id="1"\; label="N"\; filenames="A20081003157500252.png,A20081003157500262.png"\; state="\{\{state\}\}">N</figure-callout>}}_1^{\&prime;}+1}{\mathrm{\&Sigma;}}}{M}_{\mathrm{ij}}=1(1\&le;j\&le;N_2^{\&prime;})\end{array}\right.---\left(19\right)$

Obviously, (17) formula by the constraint of (19) formula is exactly a typical assignment problem.In nearest decades, scholars have proposed a large amount of algorithms and have solved this optimization problem, as lax labeling algorithm, linear programming algorithm, simulated annealing, genetic algorithm, neural network algorithm and progressive soft assignment algorithm etc.The present invention adopts the progressive soft assigning method of Steven proposition (specifically referring to S.Gold, A.Rangarajan.A Graduated Assignment Algorithm for Graph Matching, IEEETrans.on Pattern Analysis and Machine Intelligence, vol.18 (4), pp:377-387,1996.) obtaining to mate matrix M, concrete steps are:

(1)M _ij＝βW _ij；

(2) go normalization coupling matrix M, promptly $M_{\mathrm{ij}}=M_{\mathrm{ij}}/\underset{j=1}{\overset{N_2^{\&prime;}}{\mathrm{\&Sigma;}}}{M}_{\mathrm{ij}};$

(3) row normalization coupling matrix M, promptly $M_{\mathrm{ij}}=M_{\mathrm{ij}}/\underset{i=1}{\overset{N_i^{\&prime;}}{\mathrm{\&Sigma;}}}{M}_{\mathrm{ij}};$

(4) when convergence, algorithm finishes; Otherwise change (2).

The convergence criterion of coupling matrix M is

$\underset{i=1}{\overset{{\mathrm{<figure-callout\; id="1"\; label="N"\; filenames="A20081003157500252.png,A20081003157500262.png"\; state="\{\{state\}\}">N</figure-callout>}}_1^{\&prime;}+1}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N_2^{\&prime;}+1}{\mathrm{\&Sigma;}}}|M_{\mathrm{ij}}^0-{\mathrm{<figure-callout\; id="1"\; label="M\; ij"\; filenames="A20081003157500252.png,A20081003157500262.png"\; state="\{\{state\}\}">M</figure-callout>}}_{\mathrm{ij}}^{}|<{\mathrm{\&epsiv;}}_0---\left(20\right)$

M wherein _Ij ⁰, M _Ij ¹Respectively before the expression once with the coupling matrix M of current acquisition in element M _IjValue, ε ₀Be a default thresholding, generally get 0.01-0.05.

Step 4: known coupling matrix M, adopt polytomy variable variable metric algorithm computational transformation model parameter vector P, allow β=β then _rβ.

Known coupling matrix M, spatial relationship consistance are estimated each element S of matrix S _IjJust can be expressed as the function of transformation model parameter vector P, formula (16) just becomes like this

$E\left(P\right)=\underset{i=1}{\overset{{\mathrm{<figure-callout\; id="1"\; label="N"\; filenames="A20081003157500252.png,A20081003157500262.png"\; state="\{\{state\}\}">N</figure-callout>}}_1^{\&prime;}}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N_2^{\&prime;}}{\mathrm{\&Sigma;}}}{M}_{\mathrm{ij}}{R}_{\mathrm{ij}}{B}_{\mathrm{ij}}{S}_{\mathrm{ij}}=\underset{i=1}{\overset{{\mathrm{<figure-callout\; id="1"\; label="N"\; filenames="A20081003157500252.png,A20081003157500262.png"\; state="\{\{state\}\}">N</figure-callout>}}_1^{\&prime;}}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N_2^{\&prime;}}{\mathrm{\&Sigma;}}}{S}_{\mathrm{ij}}\left(P\right)U_{\mathrm{ij}}---\left(21\right)$

U wherein _Ij=M _IjR _IjB _IjIt is a constant.Obviously, (17) optimization problem of Miao Shuing is exactly the maximization problems of a multidimensional variable, also there are a lot of algorithms can solve it at present, the present invention adopts the polytomy variable variable metric algorithm of better performances (specifically referring to W.Press, S.Teukolsky, W.Vetterling and B.Flannery.Numerical Recipes in C, the Art of Scientific Computing, Second Edition, Cambridge University Press, 1992.) obtain transformation model parameter vector P.

Step 5: if controlled variable β＞β _Max, judge whether transformation model parameter vector P restrains, if convergence, algorithm finishes, if do not restrain, returns step 2; If controlled variable β≤β _Max, then change step 3.

The convergence criterion of transformation model parameter vector P is

‖P ⁰-P ¹‖＜ε ₁ (22)

P wherein ⁰, P ¹Respectively before the expression once with the transformation model parameter vector of current acquisition, ε ₁Be a default thresholding, generally get 0.01-0.05.

Adopt the present invention can reach following technique effect:

(1) the present invention with straight line marshalling as the registration primitive, overcome existing method directly with linear feature as the registration primitive, when feature extraction is imperfect, cause the inaccurate deficiency of corresponding image points of acquisition;

(2) the present invention is in the linear feature matching process, considered feature extraction result's reliability, guarantee that reliable feature preferentially is used for coupling, overcome existing method and treated each feature on an equal basis, when causing the shortcoming that the registration results precision is not high even registration is failed when insecure feature is used for registration as matching characteristic extracting;

(3) organic assembling of the present invention reliability, spatial relationship consistance and the similarity of straight line marshalling set up energy function, overcome the defective that causes mistake coupling, calculation of complex easily that exists based on the spatial relationship method, based on characteristic similarity method and sequential interblock space relation and similarity method.

Description of drawings

Fig. 1 is an overview flow chart of the present invention;

Fig. 2 is the present invention's second step cathetus feature and reliability synoptic diagram thereof;

Fig. 3 is the present invention's second step cathetus marshalling and reliability synoptic diagram thereof;

Fig. 4 is two step iteration optimization algorithms process flow diagrams in the 6th step of the present invention;

Fig. 5 is a soft assignment algorithm process flow diagram in the 6th step of the present invention.

Embodiment

Fig. 1 is an overview flow chart of the present invention, may further comprise the steps:

1, from reference picture and image subject to registration, extracts the straight line marshalling respectively, as the registration primitive.

2, straight line marshalling in the calculating reference picture and the reliability measure between the straight line marshalling in the image subject to registration obtain reliability measure matrix R.

3, straight line marshalling in the calculating reference picture and the similarity measure between the straight line marshalling in the image subject to registration obtain the similarity measure matrix B.

4, straight line marshalling in the calculating reference picture and the spatial relationship consistance between the straight line marshalling in the image subject to registration are estimated, and obtain the spatial relationship consistance and estimate matrix S.

5, set up the energy function that (16) formula is described, the linear feature matching problem is converted into the optimization problem that (17) formula is described.

6, the optimization problem that adopts two step iteration optimization algorithms that (17) formula is described is carried out iteration optimization, obtains optimum coupling matrix M _OptWith transformation model parameter vector P _Opt

Fig. 2 is the present invention's second step cathetus feature and reliability synoptic diagram thereof.For any straight line feature l in reference picture or the image subject to registration _H ^m(H=1, m=1,2 ..., N ₁H=2, m=1,2 ..., N ₂), as shown in the figure, supposed l _H ^mMid point M _H ^mAnd l _H ^mAngle is Δ θ _H ^mTwo dotted lines are l _H ^mThe maximum direction that may depart from, l _H ^mActual position only may drop between two dotted lines.

Fig. 3 is the present invention's second step cathetus marshalling and reliability synoptic diagram thereof.For any one straight line marshalling z in reference picture or the image subject to registration _H ^k, z _H ^kBy linear feature l _H ^K1, l _H ^K2Form, as shown in the figure, cross l _H ^K1Mid point M _H ^K1And l _H ^K1Angle is Δ θ _H ^K1Two dotted lines are l _H ^K1L is crossed in contingent maximum direction skew _H ^K2Mid point M _H ^K2And l _H ^K2Angle is Δ θ _H ^K2Two dotted lines are l _H ^K2Contingent maximum direction skew, four dotted lines meet at A, B, C and D 4 points.Like this, linear feature l _H ^K1And l _H ^K2Actual position can only drop on them separately between two dotted lines of mid point, so straight line marshalling z _H ^kTrue intersection point Q _H ^kWill drop on by A B, the quadrilateral r that C and D surround at 4 _ABCDIn, and may drop on r _ABCDOn interior any position.

Fig. 4 is the process flow diagram of two step iteration optimization algorithms in the 6th step of the present invention, comprises five steps:

1, initialization transformation model parameter vector P=P ₀

2, initialization controlled variable β=β ₀

3, known transform model parameter vector P adopts soft assigning method to calculate the coupling matrix M.

4, known coupling matrix M adopts polytomy variable variable metric algorithm computational transformation model parameter vector P, β=β _rβ.

5, if controlled variable β＞β _Max, judge whether transformation model parameter vector P restrains, if convergence, algorithm finishes, if do not restrain, returns step 2; If controlled variable β≤β _Max, then change step 3.

Fig. 5 is the soft assignment algorithm process flow diagram in the 6th step of the present invention, comprises four steps:

1、M _ij＝βW _ij。

2, go normalization coupling matrix M, promptly $M_{\mathrm{ij}}=M_{\mathrm{ij}}/\underset{j=1}{\overset{N_2^{\&prime;}}{\mathrm{\&Sigma;}}}{M}_{\mathrm{ij}}.$

3, row normalization coupling matrix M, promptly $M_{\mathrm{ij}}=M_{\mathrm{ij}}/\underset{i=1}{\overset{N_i^{\&prime;}}{\mathrm{\&Sigma;}}}{M}_{\mathrm{ij}}.$

4, when the convergence of coupling matrix M, algorithm finishes; Otherwise change 2.

Claims (3)

1. one kind based on the feature matching method in the straight line characteristic image registration, it is characterized in that may further comprise the steps:

The first step is extracted the straight line marshalling respectively as the registration primitive from reference picture and image subject to registration, method is: adopt sane straight line extracting method from reference picture I ₁(x ₁, y ₁) extraction N ₁The bar linear feature is from image I subject to registration ₂(x ₂, y ₂) the middle N that extracts ₂The bar linear feature is designated as respectively $L_1=\{l_1^1,l_1^2,...,l_1^{N_1}\}$ With $L_2=\{l_2^1,l_2^2,...,l_2^{N_2}\},$ N ₁, N ₂Be positive integer, (x ₁, y ₁), (x ₂, y ₂) represent the pixel coordinate in reference picture and the image subject to registration respectively, by L ₁Or L ₂In a straight line forming of any two straight lines organize into groups being straight line, use parameter $V=[l_H^{k1},l_H^{k2},\mathrm{lengt}{h}_H^{k1},\mathrm{lengt}{h}_H^{k2},S_H^{k1},S_H^{k2},E_H^{k1},E_H^{k2},M_H^{k1},M_H^{k2},O_H^k,x_H^k,y_H^k]$ Any one straight line marshalling z is described _H ^k(H=1, k=1,2 ..., N ' ₁H=2, k=1,2 ..., N ' ₂) attribute, l wherein _H ^K1, l _H ^K2Straight line marshalling z is formed in expression _H ^kTwo linear features, length _H ^K1, length _H ^K2, S _H ^K1, S _H ^K2, E _H ^K1, E _H ^K2, M _H ^K1, M _H ^K2Represent their length, four end points and two mid points respectively, O _H ^kBe the extraction intersection point of straight line marshalling, x _H ^k, y _H ^kIt is intersecting point coordinate; Defining a size is N ' ₁* N ' ₂The coupling matrix M represent matching relationship between the marshalling of reference picture and image cathetus subject to registration, promptly

N ' wherein ₁, N ' ₂Be positive integer, N ' ₁Be the straight line marshalling number of extracting in the reference picture, N ' ₂Be the straight line marshalling number of extracting in the image subject to registration, the straight line that extracts in reference picture marshalling $Z_1=\{z_1^1,z_1^2,...,z_1^{N_1^{\&prime;}}\},$ The straight line marshalling of extracting in the image subject to registration $Z_2=\{z_2^1,z_2^2,...,z_2^{N_2^{\&prime;}}\},$ 1≤i≤N ', 1≤j≤N '; The coupling matrix M satisfies matching criterior one by one, promptly

$\left\{\begin{array}{c}\underset{j=1}{\overset{N_2^{\&prime;}}{\mathrm{\&Sigma;}}}{M}_{\mathrm{ij}}\&le;1(1\&le;i\&le;N_1^{\&prime;})\\ \underset{i=1}{\overset{N_1^{\&prime;}}{\mathrm{\&Sigma;}}}{M}_{\mathrm{ij}}\&le;1(1\&le;j\&le;N_2^{\&prime;})\end{array}\right.---\left(2\right)$

Second step, calculate the reliability measure between the marshalling of reference picture and image cathetus subject to registration, obtain reliability measure matrix R, R is that a size is N ' ₁* N ' ₂Matrix, it each R _IjThe marshalling of expression straight line is to LP _IjReliability measure, LP _IjBe Z ₁In any one straight line marshalling z ₁ ⁱWith Z ₂Any one straight line marshalling z ₂ ^jThe straight line marshalling of forming is right, and method is: adopt D-S evidence theory method, to c ₁ ⁱAnd c ₂ ^jMerge calculating, obtain R _Ij, promptly

$R_{\mathrm{ij}}=\frac{c_1^i\&times;c_2^j}{c_1^i\&times;c_2^j+(1-c_1^i)(1-c_2^j)}---\left(8\right)$

C wherein ₁ ⁱBe straight line marshalling z ₁ ⁱReliability, c ₂ ^jBe straight line marshalling z ₂ ^jReliability, they calculate with following formula

$c_H^k=\exp [-\frac{{\left({\stackrel{\&OverBar;}{d}}_H^k\right)}^2}{2{{\mathrm{\&sigma;}}_2}^2}]---\left(7\right)$

σ wherein ₂Be the variance of Gaussian function, get 24; During H=1, k=1,2 ..., N ' ₁, during H=2, k=1,2 ... N ' ₂, d _H ^kd _H ^kBe d _H ^kExpectation, d _H ^kExpression z _H ^kTrue intersection point Q _H ^kWith extraction intersection point Q _H ^kBetween distance, have

${\stackrel{\&OverBar;}{d}}_H^k=\frac{1}{S_{r_{\mathrm{ABCD}}}}{\&Integral;}_{r_{\mathrm{ABCD}}}\left|\right|O_H^k-Q_H^k\left|\right|d{Q}_H^k---\left(6\right)$

Wherein, z _H ^kBy linear feature l _H ^K1, l _H ^K2Form, cross l _H ^K1Mid point M _H ^K1And l _H ^K1Angle is Δ θ _H ^K1Two dotted lines are l _H ^K1L is crossed in contingent maximum direction skew _H ^K2Mid point M _H ^K2And l _H ^K2Angle is Δ θ _H ^K2Two dotted lines are l _H ^K2Contingent maximum direction skew, four dotted lines meet at A, B, C and D 4 points, Δ θ _H ^K1With Δ θ _H ^K2Be linear feature l _H ^K1And l _H ^K2Direction skew, their size is determined by formula (3), r _ABCDBe A, B, 4 quadrilaterals that surround of C and D,

Expression r _ABCDArea, its size is according to crossing l _H ^K1Mid point M _H ^K1With mistake l _H ^K2Mid point M _H ^K2Four dotted line equations obtain, four dotted line equations are by Δ θ _H ^K1With Δ θ _H ^K2And the linear feature l that extracts _H ^K1, l _H ^K2Obtain,

$\mathrm{\&Delta;}{\mathrm{\&theta;}}_H^m=\mathrm{\&alpha;exp}[-\frac{{\left({\mathrm{length}}_H^m\right)}^2}{2{{\mathrm{\&sigma;}}_1}^2}]---\left(3\right)$

Wherein α is a default thresholding, is used for limiting direction shifted by delta θ _H ^mMaximum magnitude, get α=0.175 radian, i.e. the direction of linear feature skew is no more than 10 degree; Length _H ^mBe l _H ^mLength; σ ₁Be the variance of Gaussian function, the direction skew will not take place in its decision linear feature how long, get 20 40, specifically depend on the resolution of image, and image resolution ratio is high more, σ ₁The closer to 40;

The 3rd step, calculate the similarity measure between the marshalling of reference picture and image cathetus subject to registration, obtain the similarity measure matrix B, B is that a size is N ' ₁* N ' ₂Matrix, it each B _IjAny two straight lines marshalling z in expression reference picture and the image subject to registration ₁ ⁱAnd z ₂ ^jSimilarity measure, similarity measure is represented the similarity degrees between the marshalling of two straight lines, similarity measure B _IjBe defined as

$B_{\mathrm{ij}}=\left|\right|\frac{{\mathrm{length}}_1^{i1}-{\mathrm{length}}_1^{i2}}{{\mathrm{length}}_1^{i1}+{\mathrm{length}}_1^{i2}}|-|\frac{{\mathrm{length}}_2^{j1}-{\mathrm{length}}_2^{j2}}{{\mathrm{length}}_2^{j1}+{\mathrm{length}}_2^{j2}}\left|\right|---\left(9\right)$

(9) formula is carried out conversion, promptly

B _ij＝F ₁(B _ij)(10)

F wherein ₁() is a monotonic decreasing function, adopts Gaussian function, promptly

$F_1\left(x\right)=\exp (-\frac{x^2}{{2\mathrm{\&sigma;}}_3^2})---\left(11\right)$

σ wherein ₃Be the variance of Gaussian function, generally get σ in practice ₃=1, with B _IjSubstitution (11) formula,

$B_{\mathrm{ij}}=\exp (-\frac{{B_{\mathrm{ij}}}^2}{2{\mathrm{\&sigma;}}_3^2})---\left(12\right)$

In the 4th step, the spatial relationship consistance of calculating between the marshalling of reference picture and image cathetus subject to registration is estimated, and obtains the spatial relationship consistance and estimates matrix S, and S also is big or small N ' ₁* N ' ₂Matrix, it each S _IjRepresent any two straight lines marshalling z in two width of cloth images ₁ ⁱAnd z ₂ ^jBetween the spatial relationship consistance estimate S _IjCalculate by following formula

$S_{\mathrm{ij}}=\sqrt{{(x_1^i-T_x\left(x_2^j\right))}^2+{(y_1^i-T_y\left(y_2^j\right))}^2}---\left(13\right)$

T wherein _x(), T _y() represents the transforming function transformation function of transformation model on x direction and y direction between reference picture and the image subject to registration respectively, and the transformation model between reference picture and the image subject to registration is described with transformation model parameter vector P, and P is relevant with the transformation model type of employing; The Gaussian function of (11) formula of employing is done conversion to (13) formula, promptly

S _ij＝F ₁(S _ij)(14)

With S _IjSubstitution (14) formula,

$S_{\mathrm{ij}}=\exp (-\frac{{S_{\mathrm{ij}}}^2}{2{\mathrm{\&sigma;}}_3^2})---\left(15\right);$

In the 5th step, set up energy function:

$E(M,P)=\underset{i=1}{\overset{N_1^{\&prime;}}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N_2^{\&prime;}}{\mathrm{\&Sigma;}}}{M}_{\mathrm{ij}}{R}_{\mathrm{ij}}{B}_{\mathrm{ij}}{S}_{\mathrm{ij}}---\left(16\right)$

Straight lines marshalling that has comprised all correct couplings when the coupling matrix M to and to have rejected the straight line marshalling of all erroneous matching right, the value of E () will reach a global maximum, at this moment coupling matrix M is optimum, transformation model parameter vector P by the Optimum Matching matrix computations also is optimum simultaneously, promptly

<M _opt，P _opt>＝argmax(E(·))(17)

M wherein _Opt, P _OptRepresent optimum coupling matrix and optimum transformation model parameter vector respectively, arg max () represents that to maximizing like this, the matching problem of straight line marshalling is transformed into an energy function optimization problem;

In the 6th step, the optimization problem that adopts two step iteration optimization algorithms that (17) formula is described is carried out iteration optimization, obtains M _OptAnd P _Opt:

Step 1: adopt a sequential assemblage characteristic similarity and the conforming method of spatial relationship to come initialization transformation model parameter vector P: at first estimate and select to find the solution the minimized number 3-5 straight line marshalling doubly of transformation model needs to right as the marshalling of candidate matches straight line according to characteristic similarity, it is right to estimate the minimized number Optimum Matching straight line marshalling that utilizes the method for exhaustion or RANSAC algorithm to obtain to find the solution transformation model parameter vector P needs according to the spatial relationship consistance then, utilizes their extraction intersection point to obtain the initial value P of transformation model parameter vector P ₀

Step 2: initialization controlled variable β=β ₀, β ₀, β _rAnd β _MaxBe respectively initial value, rate of growth and the maximal value of controlled variable β, get β ₀=0.5, β _rGet 1.05-1.1, β _Max=10;

Step 3: known transform model parameter vector P, adopt soft assigning method to calculate the coupling matrix M;

Step 4: known coupling matrix M, spatial relationship consistance are estimated each element S of matrix S _IjJust can be expressed as the function of transformation model parameter vector P, formula (16) becomes

$E\left(P\right)=\underset{i=1}{\overset{N_1^{\&prime;}}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N_2^{\&prime;}}{\mathrm{\&Sigma;}}}{M}_{\mathrm{ij}}{R}_{\mathrm{ij}}{B}_{\mathrm{ij}}{S}_{\mathrm{ij}}=\underset{i=1}{\overset{N_1^{\&prime;}}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N_2^{\&prime;}}{\mathrm{\&Sigma;}}}{S}_{\mathrm{ij}}\left(P\right)U_{\mathrm{ij}}---\left(21\right)$

U wherein _Ij=M _IjR _IjB _IjBe a constant, adopt polytomy variable variable metric algorithm computational transformation model parameter vector P, then β=β _rβ;

Step 5: if controlled variable β＞β _Max, judge whether transformation model parameter vector P restrains, if convergence, algorithm finishes, if do not restrain, returns step 2; If controlled variable β≤β _Max, then change step 3; The convergence criterion of transformation model parameter vector P is

||P ⁰-P ¹||＜ε ₁(22)

P wherein ⁰, P ¹Respectively before the expression once with the transformation model parameter vector of current acquisition, ε ₁Be a default thresholding, get 0.01-0.05.

2. as claimed in claim 1 based on the feature matching method in the straight line characteristic image registration, only select those straight line that satisfies constraint condition marshallings as the registration primitive when it is characterized in that from reference picture and image subject to registration, extracting the straight line marshalling, constraint condition is: the extraction intersection point of straight line marshalling to the Hausdroff between two straight lines forming it apart from sum less than a default thresholding T, T gets 10-30 pixel.

3. as claimed in claim 1 based on the feature matching method in the straight line characteristic image registration, it is characterized in that described soft assigning method is:

(1)M _ij＝βW _ij；

(2) go normalization coupling matrix M, promptly $M_{\mathrm{ij}}=M_{\mathrm{ij}}/\underset{j=1}{\overset{N_2^{\&prime;}}{\mathrm{\&Sigma;}}}{M}_{\mathrm{ij}};$

(3) row normalization coupling matrix M, promptly $M_{\mathrm{ij}}=M_{\mathrm{ij}}/\underset{i=1}{\overset{N_i^{\&prime;}}{\mathrm{\&Sigma;}}}{M}_{\mathrm{ij}};$

(4) when convergence, algorithm finishes, otherwise changes (2), and the convergence criterion of coupling matrix M is

$\underset{i=1}{\overset{N_1^{\&prime;}+1}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N_2^{\&prime;}+1}{\mathrm{\&Sigma;}}}|M_{\mathrm{ij}}^0-M_{\mathrm{ij}}^1|<{\mathrm{\&epsiv;}}_0---\left(20\right)$

M wherein _Ij ⁰, M _Ij ¹Respectively before the expression once with the coupling matrix M of current acquisition in element M _IjValue, ε ₀Be a default thresholding, get 0.01-0.05.

Images