CN102201119A - Method and system for image registering based on control point unbiased transformation - Google Patents

Abstract

The invention is suited for the field of image processing and provides an image registering method, comprising the following steps of constructing a present transformation function based on a present control point couple set in a digital image to be registered, determining the initial positions of a newly-added source control point and a newly-added target control point based on the transformation function, determining a target control point corresponding to the newly-added control point in a local area of the newly-added target control point based on the unbiasedness of the transformation and the comprehensive measurement of the local registering measurement, temporarily adding the newly-added control point and the newly-added target control point into the present control point couple set and calculating the registering measurement of the global image, determining the probability value of the newly-added control point couple entering the present control point couple set based on the variance of the image registering measurement, adding the newly-added control point couple into the present control point couple set if the probability value is larger than a randomly generated probability reference value, and repeating the above process till the image registering precision achieves the requirement or there is no newly-added control point. According to the invention, the precision symbolizing the corresponding relation between points is effectively improved and a transformation function with better unbiasedness is obtained.

Description

A kind of method for registering images and system that does not have inclined to one side conversion based on the reference mark

Technical field

The invention belongs to technical field of image processing, relate in particular to a kind of method for registering images and system that does not have inclined to one side conversion based on the reference mark.

Background technology

The purpose of image registration is the conversion of seeking between two width of cloth images, makes the position alignment of respective objects in the image.Method for registering based on the monumented point corresponding relation is a class important method in the image elastic registrating, but because its transforming function transformation function adopts parameter model, can not guarantee it is reversible, therefore is not suitable for the large deformation registration problems.

In order to make based on the transforming function transformation function in the method for registering of monumented point corresponding relation is reversible, and method commonly used is the structure objective function, by transforming function transformation function is increased constraint condition, can obtain reversible transforming function transformation function.Method commonly used retrains as the transformation function coefficients to B batten structure, and is reversible to guarantee transforming function transformation function.If the Jacobian determinant of transforming function transformation function is permanent in just, then transforming function transformation function can guarantee it is reversible.The transforming function transformation function of these method constructs only is reversible, can not guarantee it is no inclined to one side.If the Jacobian value of forward and reciprocal transformation function is symmetrically distributed, then be called no inclined to one side conversion near 1.It is extremely important for image registration not have inclined to one side conversion, and it means no matter image A is to the image B registration, or image B can make same target obtain registration to the image A registration.At present fewer about the research of not having inclined to one side conversion, the symmetrical cost function of Gholipour definition make the transforming function transformation function after source images and target image exchange reciprocal, but this method adopts the FFD deformation model, and calculated amount is bigger.Johnson has constructed the consistance method for registering of monumented point, and this method can make monumented point reach consistent in forward and reverse conversion, but can cause the monumented point corresponding relation to take place also to need to introduce consistance intensity registration and further adjust than mistake.Leow has defined the symmetrical KL distance between transforming function transformation function and the ideal transformation, and has defined symmetrical objective function and seek no inclined to one side transforming function transformation function, but the Leow employing is the viscous fluid registration model, and calculated amount is still very big.

Guaranteeing that structure is very attracting based on the no inclined to one side conversion of basis function expansion, because this class transforming function transformation function can guarantee the corresponding relation between the unique point under the monumented point corresponding relation prerequisite, has clear and definite analytic expression simultaneously, be easy to analyze, calculate simply, calculated amount is little.But have following difficulty: (1) seeks relatively difficulty of accurate monumented point corresponding relation.The monumented point corresponding relation often is subjected to the influence of local extremum and error occurs, registration results is produced have a strong impact on; (2) the no inclined to one side ratio of transformation of structure is difficult, because the transforming function transformation function of basis function expansion itself can not guarantee reversibility, is not suitable for the large deformation registration problems.Therefore seldom there is documents how to construct reversible transformation based on the monumented point corresponding relation, Miller has constructed the differomorphism function under the prerequisite that guarantees the monumented point corresponding relation, but this method has adopted the light stream model, need find the solution partial differential equation, calculation of complex, and research is based on the no inclined to one side conversion of monumented point corresponding relation just still less.

Structure need solve two problems based on the no inclined to one side conversion of monumented point corresponding relation: (1) seeks corresponding monumented point.The method of many definite monumented point corresponding relations has been proposed at present, the information-theoretical method that is based on representative row property, these methods all need be at the optimum match point of search, if initial searching position preferably can be provided, and will be very helpful to improving matching precision.(2) under the prerequisite that guarantees the monumented point corresponding relation, structure is based on the no inclined to one side conversion of basis function expansion.If guarantee the monumented point corresponding relation, the transforming function transformation function form of basis function extended architecture determines that its unbiasedness can't guarantee to have only by adjusting the position that monumented point distributes, or add monumented point in position, can adjust the unbiasedness of transforming function transformation function.

Summary of the invention

Technical matters to be solved by this invention is to provide a kind of method for registering images that does not have inclined to one side conversion based on the reference mark, be intended to solve the monumented point corresponding relation that exists when prior art is used image registration interpolation based on monumented point and be difficult to determine, simultaneously the relatively poor problem of transforming function transformation function unbiasedness.

The present invention is achieved in that a kind ofly do not have the method for registering images of inclined to one side conversion based on the reference mark, and described method comprises the steps:

Steps A is constructed current transforming function transformation function according to the current reference mark pair set in the digital picture subject to registration; Described current reference mark pair set comprises set of initial source reference mark and the set of initial target reference mark;

Step B selects newly-increased reference mark, source on every side according to the current transforming function transformation function that structure in the steps A obtains, and will increase the initial position of the mapping position at reference mark, source as newly-increased target control point newly at reference mark, current source;

Step C if can find newly-increased reference mark, source among the step B, then according to the unbiasedness and the local registration amount of conversion, adjusts newly-increased target control point;

Step D will increase reference mark, source and newly-increased target control point newly and temporarily join in the pair set of current reference mark, and calculate the symmetrical KL distance of new transforming function transformation function, and image overall registration tolerance;

Step e, according to the variable quantity of image overall registration tolerance, and the symmetrical KL distance of new transforming function transformation function, determine that newly-increased reference mark is to joining the probable value in the pair set of current reference mark; Described newly-increased reference mark is to comprising corresponding reference mark, newly-increased source and newly-increased target control point;

Step F, if calculate probable value in the step e greater than the probability reference value that produces at random, then with newly-increased reference mark to adding in the pair set of current reference mark;

Repeating step A is to step F.

The present invention also provides a kind of does not have the figure registration system of inclined to one side conversion based on the reference mark, comprises;

Current transforming function transformation function tectonic element is used for constructing current transforming function transformation function according to the current reference mark pair set of digital picture subject to registration; Described current reference mark pair set comprises set of initial source reference mark and the set of initial target reference mark;

Newly-increased reference mark, source selected cell, be used for selecting newly-increased reference mark, source on every side at reference mark, current source, and will increase the initial position of the mapping position at reference mark, source newly as newly-increased target control point according to the current transforming function transformation function that described current transforming function transformation function tectonic element structure obtains;

Newly-increased target control point determining unit if reference mark, described newly-increased source selected cell can find newly-increased reference mark, source, then is used for unbiasedness and local registration amount according to conversion, adjusts newly-increased target control point;

The probability calculation unit is used for newly-increased reference mark, source and newly-increased target control point are temporarily joined current reference mark pair set, and calculates the symmetrical KL distance of new transforming function transformation function, and image overall registration tolerance; And according to the variable quantity of image overall registration tolerance, and the symmetrical KL distance of new transforming function transformation function, determine that newly-increased reference mark is to joining the probable value in the pair set of current reference mark; Described newly-increased reference mark comprises corresponding reference mark, newly-increased source and newly-increased target control point to the guarantor;

Current reference mark pair set updating block, if calculate probable value in the described probability calculation unit greater than the probability reference value that produces at random, then with newly-increased reference mark to adding in the pair set of current reference mark.

The present invention is directed to the monumented point corresponding relation, proposed a kind of gradual method for registering images based on no inclined to one side conversion, this method only needs several initial mark points seldom, constantly adds new monumented point by iteration and improves registration accuracy gradually.This method is calculated simple, not only can be used to solve little deformable registration problem, also is applicable to large deformation image elastic registrating problem.

Description of drawings

Fig. 1 is the process flow diagram of the method for registering images that provides of the embodiment of the invention;

Fig. 2 is the process flow diagram that obtains newly-increased impact point initial position that the embodiment of the invention provides;

Fig. 3 is the right process flow diagram in definite newly-increased corresponding reference mark that the embodiment of the invention provides;

Fig. 4 be the embodiment of the invention provide receive the right process flow diagram in newly-increased reference mark with the simulated annealing probability model;

Fig. 5 be the embodiment of the invention provide do not have the logical organization schematic diagram of the figure registration system of inclined to one side conversion based on the reference mark.

Embodiment

In order to make purpose of the present invention, technical scheme and advantage clearer,, the present invention is further elaborated below in conjunction with drawings and Examples.Should be appreciated that specific embodiment described herein only in order to explanation the present invention, and be not used in qualification the present invention.

In embodiments of the present invention, according to quantity initial source reference mark and the set of target control point seldom, structure initial transformation function.Near the initial source reference mark, determine newly-increased reference mark, source, obtain the initial position of newly-increased target control point, and determine the hunting zone of newly-increased target control point according to the initial transformation function.In the hunting zone of newly-increased target control point,, obtain the accurate position of newly-increased target control point according to the unbiasedness and the local registration amount of transforming function transformation function; Newly-increased reference mark, source and newly-increased target control point are temporarily added in the pair set of current reference mark, calculate the registering images under this expansion reference mark pair set, and the registration of computed image tolerance.Unbiasedness according to this registration tolerance and current transforming function transformation function is determined the simulated annealing probability, receives this monumented point to joining in the pair set of current reference mark with probability.Repeat said process, till image registration accuracy reaches requirement.

Fig. 1 shows the flow process of the image registration that the embodiment of the invention provides, and details are as follows:

In step S101,, construct current transforming function transformation function according to current reference mark pair set.

In embodiments of the present invention, the image that interpolation is handled is a digital picture, and initial control point seldom (generally is less than 10 pairs) to known and quantity.Suppose that the set of initial source reference mark is P=[p ₁, p ₂..., p _N] ^T, the set of initial target reference mark is Q=[q ₁, q ₂..., q _N] ^T, the initial function of structure is h (x),

$h\left(x\right)=x+\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{w}_{i}U\left(\right||x-p_i|\left|\right)---\left(1\right)$

X ∈ R wherein ²Be the arbitrfary point in the image, R is the real number set, p _i, i=1 ..., N is the reference mark, source, U (r) is a radial basis function, coefficient w _i, i=1 ..., N is found the solution by equation (2).

KW＝Q-P (2)

K is N * N matrix, K _Ij=U (|| p _i-p _j||), W=[w ₁, w ₂..., w _N] ^TInitial control point to quantity is seldom generally got less than 10 pairs.These reference mark preferably are evenly distributed on the entire image zone to setting by manual type, and its corresponding relation is general relatively accurately.

In step S102, select newly-increased reference mark, source on every side at reference mark, current source according to the current transforming function transformation function that structure among the step S101 obtains, determine the initial position of newly-increased target control point.

In the present embodiment, can construct current transforming function transformation function, near reference mark, current source, determine newly-increased reference mark, source, and determine the initial position of newly-increased target control point according to current transforming function transformation function based on radial basis function according to the initial control point pair set.

Current transforming function transformation function is h (x), and its Jacobian determinant is Dh (x).Computing method at the Jacobian determinant at an x place are as follows:

$\mathrm{Dh}\left(x\right)=\left|\begin{array}{cc}\frac{{\∂h}_x\left(x\right)}{\∂x}& \frac{{\∂h}_x\left(x\right)}{\∂y}\\ \frac{{\∂h}_y\left(x\right)}{\∂x}& \frac{{\∂h}_y\left(x\right)}{\∂y}\end{array}\right|=\frac{{\∂h}_x}{\∂x}\·\frac{{\∂h}_y}{\∂y}-\frac{\∂h_x}{\∂y}\·\frac{{\∂h}_y}{\∂x}---\left(3\right)$

Select newly-increased reference mark, source p _aFor

$p_a=\underset{x}{\arg }\max |\mathrm{Dh}\left(x\right)-1|,$ And distT _Min＜|| x-p _i||＜distT _Max, x ∈ E, $\∀p_i\∈P$ (4)

Satisfy the newly-increased source reference mark p of formula (3) _aHave following characteristics: the Jacobian determinant and 1 of (1) this point differs bigger, and promptly the unbiasedness of transforming function transformation function is relatively poor near this point; (2) distance of this point and current monumented point can not too far (distance be less than maximal distance threshold distT _Max), promptly near current monumented point, select, simultaneously can not hypertelorism (distance is greater than minimum threshold of distance distT _Min), in order to avoid the initial position error is excessive, the validity of influence search; (3) this point drops on the image border, and E is the set of image border point.

Target control point q _aInitial position be:

$q_a^0=h\left(p_a\right).$

In step S102, if can not find suitable newly-increased source reference mark p _a, then algorithm finishes.Otherwise, continue step S103.

In step S103,, adjust newly-increased target control point according to the unbiasedness and the local registration amount of conversion.The specific implementation flow process as shown in Figure 2.

In step S201, at the initial position with newly-increased target control point is in the regional area at center, choose the point that drops on the image border as the alternative target reference mark, add newly-increased reference mark, source and alternative target reference mark to the reference mark set temporarily, calculate interim transforming function transformation function, concrete steps such as S101 repeat no more here.

In step S202, calculate the symmetrical KL distance of interim transforming function transformation function.

Suppose that h ' will put (p temporarily _a, x) add set (P, the interim transforming function transformation function that obtains after Q) to.Make the forward of P → Q be mapped as h ', oppositely be mapped as h ' ^-1Note h ' and h ' ^-1The Jacobian determinant be respectively | Dh ' (x) | and | Dh ' ^-1(x) |.Define three probability density functions:

${\mathrm{pdf}}_h\left(\mathrm{\ξ}\right)=\left|{\mathrm{Dh}}^{\′}\left(\mathrm{\ξ}\right)\right|,{\mathrm{pdf}}_{h^{-1}}\left(\mathrm{\ξ}\right)=\left|{\mathrm{Dh}}^{\′-1}\left(\mathrm{\ξ}\right)\right|,{\mathrm{pdf}}_{\mathrm{id}}\left(\mathrm{\ξ}\right)=1---\left(5\right)$

Pdf wherein _IdIt is the unit mapping (probability density function of id (x)=x).The KL distance can be described two similaritys between the probability density function, KL (pdf _{H '}, pdf _Id) be the KL distance between positive-going transition and the unit transformation,

Be the KL distance between reciprocal transformation and the unit transformation, the symmetrical KL of definition conversion h ' apart from sKL (h ') is

$\mathrm{sKL}\left(h^{\′}\right)=\mathrm{KL}({\mathrm{pdf}}_{h^{\′}},{\mathrm{pdf}}_{\mathrm{id}})+\mathrm{KL}({\mathrm{pdf}}_{h^{\′-1}},{\mathrm{pdf}}_{\mathrm{id}})$

$=\∫\left(\right|{\mathrm{Dh}}^{\′}\left(x\right)-1\left|\right)\log \left|{\mathrm{Dh}}^{\′}\left(x\right)\right|\mathrm{dx}---\left(6\right)$

$=\∫\left(\right|{\mathrm{Dh}}^{\′-1}\left(x\right)-1)\log \left|{\mathrm{Dh}}^{\′-1}\left(x\right)\right|\mathrm{dx}$

In step S203, according to newly-increased reference mark, the source p of interim transforming function transformation function h ' calculating _aRegional area Loc (p of living in _a) and newly-increased target control point q _aRegional area Loc (q of living in _a) registration tolerance.

$\mathrm{MSD}(\mathrm{Loc}\left(q_a\right),\mathrm{Loc}\left(p_a\right))=\underset{(i^{\′},j^{\′})\∈\mathrm{Loc}\left(p_a\right)}{\underset{(i,j)\∈\mathrm{Loc}\left(q_a\right)}{\mathrm{\Σ}}}|f(i,j)-f(i^{\′},j^{\′})|---\left(7\right)$

Wherein, (i, the j) point of expression in the reference picture, the point in (i ', j ') expression image subject to registration, f (i, j) be reference picture point (i, the intensity level of j) locating, f (i ', j ') be image subject to registration point (i ', the intensity level of j) locating.

In step S204,, determine the corresponding point of optimal target reference mark as newly-increased reference mark, source according to symmetrical KL distance and the image local registration tolerance of interim transforming function transformation function h '.

At regional area

The middle p that seeks _aCorresponding point q _aOptimal objective reference mark q _aFor

Wherein h ' will put (p temporarily _a, (P, the transforming function transformation function that obtains after Q), S ο h ' are that MSD is a square error through the image after h ' conversion, and sKL is symmetrical KL distance x) to add set to.

In step S103, will increase reference mark, source and newly-increased target control point newly and temporarily join in the pair set of current reference mark, and calculate the symmetrical KL distance of new transforming function transformation function, and global registration tolerance.Idiographic flow as shown in Figure 3.

In step S301, add newly-increased reference mark, source and newly-increased target control point to the reference mark set temporarily, calculate interim transforming function transformation function, concrete steps such as S101 repeat no more here.

In step S302, calculate the symmetrical KL distance of interim transforming function transformation function, concrete steps such as S202 repeat no more here.

In step S303, calculate the global registration tolerance of registering images according to interim transforming function transformation function.

Suppose and to put (p _a, q _a) (P obtains transforming function transformation function h after Q) to add set to _a, mutual information MI (the S ο h after the calculating deformation between image and the reference picture _a, R) measure as global registration.

At step S105, according to the variable quantity of image registration tolerance, and the symmetrical KL distance of new transforming function transformation function, determine that newly-increased reference mark is to joining the probable value in the pair set of current reference mark.

Point is to (p _a, q _a) (P, Q) the simulated annealing probability P in is to add set to

(9)

$P\_\mathrm{sKL}(p_a,q_a)=\left\{\begin{array}{cc}\exp \left(\frac{\mathrm{sKL}\left(h\right)-\mathrm{sKL}\left(h_a\right)}{T}\right)& \mathrm{sKL}\left(h\right)<\mathrm{sKL}\left(h_a\right)\\ 1& \mathrm{otherwise}\end{array}\right.$

Wherein MI is image after the conversion and the mutual information between the reference picture, and h is to (p with point _a, q _a) (P, Q) transforming function transformation function before, S ο h are through the image after the h conversion, h to add set to _aBe to (p with point _a, q _a) add set (P, Q) transforming function transformation function afterwards, S ο h to _aBe through h _aImage after the conversion.With h ' (x) | compare, h ' (x) | be the interim transforming function transformation function when determining corresponding relation, and h _aIt then is the transforming function transformation function behind definite corresponding relation.

In step S106, receive this reference mark to (p with probability _a, q _a) join current reference mark pair set (P, Q) in, idiographic flow such as Fig. 4.

In step S401, utilize tandom number generator to produce one [0,1] interval random number r.

In step S402, if satisfy MI (S ο h _a, R)＞(S ο h, R), then the monumented point pair set is updated to P to MI ^New=[P, p _a], Q ^New=[Q, q _a], otherwise, if r＜P (p _a, q _a), will put (p _a, q _a) add to the original point pair set (P, Q) in.

Go to step S101, repeat above step.

One of ordinary skill in the art will appreciate that all or part of step that realizes in the method that the various embodiments described above provide can finish by programmed instruction and relevant hardware, described program can be stored in the computer read/write memory medium, and this storage medium can be ROM/RAM, disk, CD etc.

What Fig. 5 showed that the embodiment of the invention provides does not have the logical organization principle of the figure registration system of inclined to one side conversion based on the reference mark, for convenience of description, only shows the part relevant with present embodiment.This figure registration system can be for being built in the unit of software unit, hardware cell or software and hardware combining in the image processing equipment.

With reference to Fig. 5, the figure registration system that does not have an inclined to one side conversion based on the reference mark that the embodiment of the invention provides comprises current transforming function transformation function tectonic element 51, newly-increased reference mark, source selected cell 52, newly-increased target control point determining unit 53, probability calculation unit 54 and current reference mark pair set updating block 55.

Wherein, current transforming function transformation function tectonic element 51 is constructed current transforming function transformation function according to the current reference mark pair set in the digital picture subject to registration, and described current reference mark pair set comprises set of initial source reference mark and the set of initial target reference mark.Newly-increased reference mark, source selected cell 52 is selected newly-increased reference mark, source on every side according to the current transforming function transformation function that current transforming function transformation function tectonic element 51 structures obtain at reference mark, current source, and will increase the initial position of the mapping position at reference mark, source as newly-increased target control point newly.If newly-increased reference mark, source selected cell 52 can find newly-increased reference mark, source, then, adjust newly-increased target control point by the unbiasedness and the local registration amount of newly-increased target control point determining unit 53 according to conversion.Then, probability calculation unit 54 will increase the reference mark, source newly and newly-increased target control point temporarily joins in the pair set of current reference mark, and calculates the symmetrical KL distance of new transforming function transformation function, and image overall registration tolerance; Probability calculation unit 54 and the variable quantity of measuring according to the image overall registration, and the symmetrical KL distance of new transforming function transformation function determine that newly-increased reference mark is to joining the probable value in the pair set of current reference mark; Described newly-increased reference mark comprises corresponding reference mark, newly-increased source and newly-increased target control point to the guarantor.At last then by current reference mark pair set updating block 55 according to calculating probable value in the probability calculation unit 54, with newly-increased reference mark to adding in the pair set of current reference mark.Above-mentioned each module repeats said process, till image registration accuracy reaches requirement.

Further, above-mentioned current transforming function transformation function tectonic element 51 is according to the current transforming function transformation function h of following formula construction (x):

$h\left(x\right)=x+\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{w}_{i}U\left(\right||x-p_i|\left|\right)$

Wherein x is the arbitrfary point in the image, x ∈ R ², R is the real number set, p _i, i=1 ..., N is the set of initial source reference mark, U (r) is a radial basis function, coefficient w _i, i=1 ..., N is found the solution by equation KW=Q-P and obtains, and wherein K is N * N matrix, K _Ij=U (|| p _i-p _j||), W=[w ₁, w ₂..., w _N] ^T, Q is the set of initial target reference mark.

Further, the reference mark, newly-increased source selected according to following formula of reference mark, above-mentioned newly-increased source selected cell 52:

$p_a=\underset{x}{\arg }\max |\mathrm{Dh}\left(x\right)-1|,$ And distT _Min＜|| x-p _i||＜distT _Max, x ∈ E, $\&ForAll;p_i\&Element;P$

Wherein, x ∈ R ², R is the real number set, h (x) is current transforming function transformation function, P={p _i, i=1 ..., N} is the set of image border point for initial source reference mark set, E, Dh (x) is the Jacobian determinant of current transforming function transformation function h (x), distT _MinAnd distT _MaxBe respectively minimum threshold of distance and maximal distance threshold;

Above-mentioned newly-increased target control point determining unit 53 is in the regional area at center at the initial position with newly-increased target control point at first, choose the point that drops on the image border as the alternative target reference mark, add newly-increased reference mark, source and alternative target reference mark to the reference mark set, the symmetrical KL that constructs interim transforming function transformation function h ' and calculate this interim transforming function transformation function h ' is apart from sKL (h ') temporarily;

sKL(h′)＝∫(|Dh′(x)-1|)log|Dh′(x)|dx

Wherein, x ∈ R ², R is real number set, h ' (x) | be interim transforming function transformation function, Dh ' (x) be interim transforming function transformation function h ' (x) | the Jacobian determinant;

Above-mentioned newly-increased target control point determining unit 53 is measured MSD according to the registration that regional area of living in is put in newly-increased reference mark, the source regional area of living in of interim transforming function transformation function h ' calculatings and newly-increased target control again:

$\mathrm{MSD}(\mathrm{Loc}\left(q_a\right),\mathrm{Loc}\left(p_a\right))=\underset{(i^{\&prime;},j^{\&prime;})\&Element;\mathrm{Loc}\left(p_a\right)}{\underset{(i,j)\&Element;\mathrm{Loc}\left(q_a\right)}{\mathrm{\&Sigma;}}}|f(i,j)-f(i^{\&prime;},j^{\&prime;})|$

Wherein, p _aAnd Loc (p _a) be respectively newly-increased reference mark, source and regional area of living in thereof, q _aAnd Loc (q _a) be respectively newly-increased target control point and regional area of living in thereof, (i j) represents point in the reference picture, (i, j ') point in the expression image subject to registration, (i is that reference picture is at point (i, the intensity level of j) locating j) to f, f (i ', the j ') intensity level that to be image subject to registration locate at point (i ', j ');

Above-mentioned newly-increased target control point determining unit 53 is again according to the symmetrical KL distance and the described registration tolerance MSD of interim transforming function transformation function, determines that according to following formula optimum newly-increased target control point is with as the corresponding point that increase the reference mark, source newly:

Wherein h ' will put (p temporarily _a, (P, the transforming function transformation function that obtains after Q), S ο h ' are that MSD is a square error through the image after h ' conversion, and sKL is symmetrical KL distance x) to add current reference mark pair set to.

Further, above-mentioned probability calculation unit 54 determines that according to following formula newly-increased reference mark is to joining the probable value in the pair set of current reference mark:

$P\_\mathrm{sKL}(p_a,q_a)=\left\{\begin{array}{cc}\exp \left(\frac{\mathrm{sKL}\left(h\right)-\mathrm{sKL}\left(h_a\right)}{T}\right)& \mathrm{sKL}\left(h\right)<\mathrm{sKL}\left(h_a\right)\\ 1& \mathrm{otherwise}\end{array}\right.$

Wherein MI is image after the conversion and the mutual information between the reference picture, and h is to (p with newly-increased reference mark _a, q _a) (P, Q) transforming function transformation function before, S ο h are through the image after the current transforming function transformation function h conversion to add current reference mark pair set to; h _aBe to (p with point _a, q _a) add set (P, Q) transforming function transformation function afterwards, S ο h to _aBe through h _aImage after the conversion;

Above-mentioned current reference mark pair set updating block 55 at first utilizes tandom number generator to produce one [0,1] interval random number r; If satisfy MI (S ο h _a, R)＞(S ο h R), then is updated to P with current reference mark pair set to MI ^New=[P, p _a], Q ^New=[Q, q _a]; Otherwise, if r＜P (p _a, q _a), then will put (p _a, q _a) add to the original point pair set (P, Q) in;

Wherein, P is the set of reference mark, current source, and Q is the set of current goal reference mark.

It is identical that above-mentioned figure registration system carries out the method for registering images that the principle of image registration above describes, and gives unnecessary details no longer one by one.

In embodiments of the present invention, under the right prerequisite in known a small amount of reference mark,, determine limited, a more accurate regional area, be beneficial to the search of corresponding point for searching corresponding reference mark according to the transforming function transformation function of current reference mark to providing.In the process of search corresponding point, utilized the comprehensive measurement of conversion unbiasedness and registration tolerance, can construct unbiasedness transforming function transformation function preferably.Right by the reference mark that continuous increase is new, progressively improve image registration accuracy.This method algorithm is simple, only needs a spot of initial control point, is a kind of easy, effective method for registering images based on the reference mark corresponding relation.

The above only is preferred embodiment of the present invention, not in order to restriction the present invention, all any modifications of being done within the spirit and principles in the present invention, is equal to and replaces and improvement etc., all should be included within protection scope of the present invention.

Claims (10)

1. one kind is not had the method for registering images of inclined to one side conversion based on the reference mark, it is characterized in that described method comprises the steps:

Steps A is constructed current transforming function transformation function according to the current reference mark pair set in the digital picture subject to registration; Described current reference mark pair set comprises set of initial source reference mark and the set of initial target reference mark;

Step B selects newly-increased reference mark, source on every side according to the current transforming function transformation function that structure in the steps A obtains, and will increase the initial position of the mapping position at reference mark, source as newly-increased target control point newly at reference mark, current source;

Step C if can find newly-increased reference mark, source among the step B, then according to the unbiasedness and the local registration amount of conversion, adjusts newly-increased target control point;

Step D will increase reference mark, source and newly-increased target control point newly and temporarily join in the pair set of current reference mark, and calculate the symmetrical KL distance of new transforming function transformation function, and image overall registration tolerance;

Step e, according to the variable quantity of image overall registration tolerance, and the symmetrical KL distance of new transforming function transformation function, determine that newly-increased reference mark is to joining the probable value in the pair set of current reference mark; Described newly-increased reference mark is to comprising corresponding reference mark, newly-increased source and newly-increased target control point;

Step F, if calculate probable value in the step e greater than the probability reference value that produces at random, then with newly-increased reference mark to adding in the pair set of current reference mark;

Repeating step A is to step F.

2. method for registering images as claimed in claim 1 is characterized in that, described steps A is according to the current transforming function transformation function h of following formula construction (x):

$h\left(x\right)=x+\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{w}_{i}U\left(\right||x-p_i|\left|\right)$

Wherein x is the arbitrfary point in the image, x ∈ R ², R is the real number set, p _i, i=1 ..., N is the set of initial source reference mark, U (r) is a radial basis function, coefficient w _i, i=1 ..., N is found the solution by equation KW=Q-P and obtains, and wherein K is N * N matrix, K _Ij=U (|| p _i-p _j), W=[w ₁, w ₂..., w _N] ^T, Q is the set of initial target reference mark.

3. method for registering images as claimed in claim 1 is characterized in that, the reference mark of selecting according to following formula among the described step B, newly-increased source:

$p_a=\underset{x}{\arg }\max |\mathrm{Dh}\left(x\right)-1|,$ And distT _Min＜|| x-p _i||＜distT _Max, x ∈ E, $\&ForAll;p_i\&Element;P$

Wherein, x ∈ R ², R is the real number set, h (x) is current transforming function transformation function, P={p _i, i=1 ..., N} is the set of image border point for initial source reference mark set, E, Dh (x) is the Jacobian determinant of current transforming function transformation function h (x), distT _MinAnd distT _MaxBe respectively minimum threshold of distance and maximal distance threshold.

4. method for registering images as claimed in claim 1 is characterized in that, described step C specifically comprises:

Step C1, at the initial position with newly-increased target control point is in the regional area at center, choose the point that drops on the image border as the alternative target reference mark, add newly-increased reference mark, source and alternative target reference mark to the reference mark set, the symmetrical KL that constructs interim transforming function transformation function h ' and calculate this interim transforming function transformation function h ' is apart from sKL (h ') temporarily;

sKL(h′)＝∫(|Dh′(x)-1|)log|Dh′(x)|dx

Wherein, x ∈ R ², R is real number set, h ' (x) | be interim transforming function transformation function, Dh ' (x) be interim transforming function transformation function h ' (x) | the Jacobian determinant;

Step C2, put the registration tolerance MSD of regional area of living in according to newly-increased reference mark, the source regional area of living in of interim transforming function transformation function h ' calculating and newly-increased target control:

$\mathrm{MSD}(\mathrm{Loc}\left(q_a\right),\mathrm{Loc}\left(p_a\right))=\underset{(i^{\&prime;},j^{\&prime;})\&Element;\mathrm{Loc}\left(p_a\right)}{\underset{(i,j)\&Element;\mathrm{Loc}\left(q_a\right)}{\mathrm{\&Sigma;}}}|f(i,j)-f(i^{\&prime;},j^{\&prime;})|$

Wherein, p _aAnd Loc (p _a) be respectively newly-increased reference mark, source and regional area of living in thereof, q _aAnd Loc (q _a) be respectively newly-increased target control point and regional area of living in thereof, (i j) represents point in the reference picture, (i ', j ') point in the expression image subject to registration, (i is that reference picture is at point (i, the intensity level of j) locating j) to f, f (i ', the intensity level that j) to be image subject to registration locate at point (i ', j ');

Step C3, according to the symmetrical KL distance of interim transforming function transformation function and described registration tolerance MSD, determine that according to following formula optimum newly-increased target control point is with as the corresponding point that increase the reference mark, source newly:

Wherein h ' will put (p temporarily _a, (P, the transforming function transformation function that obtains after Q), S ο h ' are that MSD is a square error through the image after h ' conversion, and sKL is symmetrical KL distance x) to add current reference mark pair set to.

5. method for registering images as claimed in claim 1 is characterized in that, described step e determines that according to following formula newly-increased reference mark is to joining the probable value in the pair set of current reference mark:

$P\_\mathrm{sKL}(p_a,q_a)=\left\{\begin{array}{cc}\exp \left(\frac{\mathrm{sKL}\left(h\right)-\mathrm{sKL}\left(h_a\right)}{T}\right)& \mathrm{sKL}\left(h\right)<\mathrm{sKL}\left(h_a\right)\\ 1& \mathrm{otherwise}\end{array}\right.$

Wherein, MI is image after the conversion and the mutual information between the reference picture, and h is to (p with newly-increased reference mark _a, q _a) (P, Q) transforming function transformation function before, S ο h are through the image after the current transforming function transformation function h conversion to add current reference mark pair set to; h _aBe to (p with point _a, q _a) add set (P, Q) transforming function transformation function afterwards, S ο h to _aBe through h _aImage after the conversion.

6. method for registering images as claimed in claim 5 is characterized in that, described step F comprises:

Step F 1 utilizes tandom number generator to produce one [0,1] interval random number r;

Step F 2 is if satisfy MI (S ο h _a, R)＞(S ο h, R), then current reference mark pair set is updated to P to MI ^New=[P, p _a], Q ^New=[Q, q _a]; Otherwise, if r＜P (p _a, q _a), then will put (p _a, q _a) add to current reference mark pair set (P, Q) in;

Wherein, P is the set of reference mark, current source, and Q is the set of current goal reference mark.

7. one kind is not had the figure registration system of inclined to one side conversion based on the reference mark, it is characterized in that, comprising:

Current transforming function transformation function tectonic element is used for constructing current transforming function transformation function according to the current reference mark pair set of digital picture subject to registration; Described current reference mark pair set comprises set of initial source reference mark and the set of initial target reference mark;

Newly-increased reference mark, source selected cell, be used for selecting newly-increased reference mark, source on every side at reference mark, current source, and will increase the initial position of the mapping position at reference mark, source newly as newly-increased target control point according to the current transforming function transformation function that described current transforming function transformation function tectonic element structure obtains;

Newly-increased target control point determining unit if reference mark, described newly-increased source selected cell can find newly-increased reference mark, source, then is used for unbiasedness and local registration amount according to conversion, adjusts newly-increased target control point;

The probability calculation unit is used for newly-increased reference mark, source and newly-increased target control point are temporarily joined current reference mark pair set, and calculates the symmetrical KL distance of new transforming function transformation function, and image overall registration tolerance; And according to the variable quantity of image overall registration tolerance, and the symmetrical KL distance of new transforming function transformation function, determine that newly-increased reference mark is to joining the probable value in the pair set of current reference mark; Described newly-increased reference mark comprises corresponding reference mark, newly-increased source and newly-increased target control point to the guarantor;

Current reference mark pair set updating block, if calculate probable value in the described probability calculation unit greater than the probability reference value that produces at random, then with newly-increased reference mark to adding in the pair set of current reference mark.

8. as claimed in claim 7 do not have the figure registration system of inclined to one side conversion based on the reference mark, it is characterized in that, described current transforming function transformation function tectonic element is according to the current transforming function transformation function h of following formula construction (x):

$h\left(x\right)=x+\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{w}_{i}U\left(\right||x-p_i|\left|\right)$

Wherein x is the arbitrfary point in the image, x ∈ R ², R is the real number set, p _i, i=1 ..., N is the set of initial source reference mark, U (r) is a radial basis function, coefficient w _i, i=1 ..., N is found the solution by equation KW=Q-P and obtains, and wherein K is N * N matrix, K _Ij=U (|| p _i-p _j), W=[w ₁, w ₂..., w _N] ^T, Q is the set of initial target reference mark.

9. as claimed in claim 7 do not have the figure registration system of inclined to one side conversion based on the reference mark, it is characterized in that the reference mark, newly-increased source that reference mark, described newly-increased source selected cell is selected according to following formula:

$p_a=\underset{x}{\arg }\max |\mathrm{Dh}\left(x\right)-1|,$ And distT _Min＜|| x-p _i||＜distT _Max, x ∈ E, $\&ForAll;p_i\&Element;P$

Wherein, x ∈ R ², R is the real number set, h (x) is current transforming function transformation function, P={p _i, i=1 ..., N} is the set of image border point for initial source reference mark set, E, Dh (x) is the Jacobian determinant of current transforming function transformation function h (x), distT _MinAnd distT _MaxBe respectively minimum threshold of distance and maximal distance threshold;

Described newly-increased target control point determining unit, be in the regional area at center at first at initial position with newly-increased target control point, choose the point that drops on the image border as the alternative target reference mark, add newly-increased reference mark, source and alternative target reference mark to the reference mark set, the symmetrical KL that constructs interim transforming function transformation function h ' and calculate this interim transforming function transformation function h ' is apart from sKL (h ') temporarily;

sKL(h′)＝∫(Dh′(x)-1|)log|Dh′(x)|dx

Wherein, x ∈ R ², R is real number set, h ' (x) | be interim transforming function transformation function, Dh ' (x) be interim transforming function transformation function h ' (x) | the Jacobian determinant;

Described newly-increased target control point determining unit is measured MSD according to the registration that regional area of living in is put in newly-increased reference mark, the source regional area of living in of interim transforming function transformation function h ' calculatings and newly-increased target control again:

$\mathrm{MSD}(\mathrm{Loc}\left(q_a\right),\mathrm{Loc}\left(p_a\right))=\underset{(i^{\&prime;},j^{\&prime;})\&Element;\mathrm{Loc}\left(p_a\right)}{\underset{(i,j)\&Element;\mathrm{Loc}\left(q_a\right)}{\mathrm{\&Sigma;}}}|f(i,j)-f(i^{\&prime;},j^{\&prime;})|$

Wherein, p _aAnd Loc (p _a) be respectively newly-increased reference mark, source and regional area of living in thereof, q _aAnd Loc (q _a) be respectively newly-increased target control point and regional area of living in thereof, (i j) represents point in the reference picture, (i ', j ') point in the expression image subject to registration, (i is that reference picture is at point (i, the intensity level of j) locating j) to f, f (i ', be that image subject to registration is at point (i, the intensity level of j) locating j);

Described newly-increased target control point determining unit is again according to the symmetrical KL distance and the described registration tolerance MSD of interim transforming function transformation function, determines that according to following formula optimum newly-increased target control point is with as the corresponding point that increase the reference mark, source newly:

Wherein h ' will put (p temporarily _a, (P, the transforming function transformation function that obtains after Q), S ο h ' are that MSD is a square error through the image after h ' conversion, and sKL is symmetrical KL distance x) to add current reference mark pair set to.

10. as claimed in claim 7 do not have the figure registration system of inclined to one side conversion based on the reference mark, it is characterized in that:

Described probability calculation unit determines that according to following formula newly-increased reference mark is to joining the probable value in the pair set of current reference mark:

$P\_\mathrm{sKL}(p_a,q_a)=\left\{\begin{array}{cc}\exp \left(\frac{\mathrm{sKL}\left(h\right)-\mathrm{sKL}\left(h_a\right)}{T}\right)& \mathrm{sKL}\left(h\right)<\mathrm{sKL}\left(h_a\right)\\ 1& \mathrm{otherwise}\end{array}\right.$

Wherein, MI is image after the conversion and the mutual information between the reference picture, and h is to (p with newly-increased reference mark _a, q _a) (P, Q) transforming function transformation function before, S ο h are through the image after the current transforming function transformation function h conversion to add current reference mark pair set to; h _aBe to (p with point _a, q _a) add set (P, Q) transforming function transformation function afterwards, S ο h to _aBe through h _aImage after the conversion;

Described current reference mark pair set updating block at first utilizes tandom number generator to produce one [0,1] interval random number r; If satisfy MI (S ο h _a, R)＞(S ο h R), then is updated to P with current reference mark pair set to MI ^New=[P, p _a], Q ^New=[Q, q _a]; Otherwise, if r＜P (p _a, q _a), then will put (p _a, q _a) add to the original point pair set (P, Q) in;

Wherein, P is the set of reference mark, current source, and Q is the set of current goal reference mark.

Images