CN105469112B - Image Feature Matching method and system based on local linear migration and rigid model - Google Patents

Image Feature Matching method and system based on local linear migration and rigid model Download PDF

Info

Publication number
CN105469112B
CN105469112B CN201510807463.3A CN201510807463A CN105469112B CN 105469112 B CN105469112 B CN 105469112B CN 201510807463 A CN201510807463 A CN 201510807463A CN 105469112 B CN105469112 B CN 105469112B
Authority
CN
China
Prior art keywords
image
parameter
matched
matrix
model
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201510807463.3A
Other languages
Chinese (zh)
Other versions
CN105469112A (en
Inventor
梅晓光
马泳
樊凡
黄珺
马佳义
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Wuhan University WHU
Original Assignee
Wuhan University WHU
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Wuhan University WHU filed Critical Wuhan University WHU
Priority to CN201510807463.3A priority Critical patent/CN105469112B/en
Publication of CN105469112A publication Critical patent/CN105469112A/en
Application granted granted Critical
Publication of CN105469112B publication Critical patent/CN105469112B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/74—Image or video pattern matching; Proximity measures in feature spaces
    • G06V10/75—Organisation of the matching processes, e.g. simultaneous or sequential comparisons of image or video features; Coarse-fine approaches, e.g. multi-scale approaches; using context analysis; Selection of dictionaries
    • G06V10/757—Matching configurations of points or features

Landscapes

  • Engineering & Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Theoretical Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Computing Systems (AREA)
  • Databases & Information Systems (AREA)
  • Artificial Intelligence (AREA)
  • General Health & Medical Sciences (AREA)
  • Medical Informatics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)

Abstract

The invention proposes a kind of Image Feature Matching method and system based on local linear migration and rigid transformation, correct matching is established for the matching of the mistake by removal initial matching point centering, the corresponding model of geometric transformation between image to be matched is established in rigid geometric transformation including being directed between image to be matched, and establishing matching double points is the correct matched corresponding model of posterior probability, is based on nearest-neighbors match point, least square method, optimal method solving model parameter；Initial matching point is calculated to for correctly matched posterior probability, and according to threshold decision initial matching point to correcting errors.This method to be modeled for there are the case where rigid transformation between image to be matched, matched error rate is significantly reduced, even if still keeping good robustness there are in the case where a large amount of erroneous matchings in preliminary matches.

Description

Image Feature Matching method and system based on local linear migration and rigid model
Technical field
The present invention relates to Image Feature Matching technical fields, specifically, the present invention relates to based on local linear migration and The Image Feature Matching technical solution of rigid model.
Background technique
The elementary object of images match be will use different sensors in different times with obtained under visual angle it is same The same area of the two images of scene is corresponded to.
In in the past few decades, scholars have studied many methods to solve remote sensing images matching problem.These methods It can be roughly divided into two types: matching process and feature-based matching method based on region.The former passes through search two images The similarity degree of original gray value finds match information in middle certain area；The latter then utilizes description of local feature similar Property or space geometry constraint to find matching double points.In the case where there is a small amount of significant details, the information ratio of gray value offer Local shape and structure are more, so that the matching effect of the method based on region is more preferable.But the method calculation amount based on region Greatly, it and in the case where image fault and luminosity change and is not suitable for.On the contrary, characterization method has better robustness, it can be with The image for handling babble, is widely used.
How to find corresponding match point in two images, form matching double points, and ensures that the correctness of matching double points is The key of image matching method.
Matching process based on region mainly has correlation method, Fourier methods and three kinds of mutual information method.Correlation method it is main Thought is to calculate the similitude that window is corresponded in two images, then will have a pair of maximum similarity degree as matching double points. But correlation method can not be applicable in the inapparent texture-free region of similitude, and be calculated complicated.Fourier methods are utilized image and exist The Fourier Representations of frequency domain.Compared with traditional correlation technique, this method is higher in computational efficiency and makes an uproar to frequency class Sound has good robustness.However, this method has certain restrictions with the image of different spectrum structures in processing.Mutual trust For breath method although its matching effect is good, it is unable to get the global maximum in entire search space, therefore inevitable Meeting lower its robustness.
In feature-based matching method, a kind of strategy for being divided into two steps is generallyd use.The first step is retouched by feature The similarity degree for stating son determines one group of initial matching double points, and wherein the overwhelming majority is correctly to match, but inevitably contain A large amount of erroneous matching.Second step finally can be matched correctly by removing the matching of mistake by geometrical constraint Point is to the geometric parameter converted between two images.The typical example of this strategy includes RANSAC method (M.A.Fischler and R.C.Bolles,“Random sample consensus:A paradigm for model fitting with application to image analysis and automated cartog-raphy,” Commun.ACM, vol.24, no.6, pp.381-395, Jun.1981), ARHV method (P.H.S.Torr and A.Zisserman,“MLESAC:A new robust estimator with application to estimating image geometry,”Comput.Vis.Image Under-stand.,vol.78,no.1,pp.138–156, ) etc. Apr.2000 the method and VFC method (J.Ma, J.Zhao, J.Tian, A.L.Yuille, and of parameter model are depended on Z.Tu,“Robust point matching via vector field consensus,”IEEE Trans.Image Process., vol.23, no.4, pp.1706-1721, Apr.2014.), GS method (H.Liu and S.Yan, " Common visual pattern discovery via spatially coher-ent correspondence,”in Proc.IEEE Conf.Comput.Vis.Pattern Recog., Jun.2010, pp.1609-1616.) and ICF method (X.Li and Z.Hu,“Rejecting misma tches by correspondence function,”Int.J.Comput.Vis., Vol.89, no.1, pp.1-17, Aug.2010) method based on nonparametric model such as.
Although these methods obtain success in many fields, in the picture containing largely because of visual angle change caused by The initial matching point pair that many mistakes can be obtained when localized distortion and more complex picture material, after preliminary matches, works as error rate When more than certain proportion, these methods cannot effectively remove mistake.Therefore needing a kind of pair of initial matching error rate has The matching process of higher robustness.
Summary of the invention
For prior art defect, the invention proposes a kind of characteristics of image based on local linear migration and rigid model Matching technique scheme.
In order to achieve the above objectives, the technical solution adopted by the present invention provides a kind of based on local linear migration and rigid model Image Feature Matching method, include the following steps,
Step 1, establishing the corresponding model of geometric transformation and matching double points between image to be matched is that correct matched posteriority is general The corresponding model of rate, realization is as follows,
For the rigid geometric transformation between image to be matched, foundation transformation mathematical model is as follows,
Y=t (x)=sRx+o
Wherein, if two width image to be matched are image a and image b, x and y are the seat of pixel on image a and image b respectively Vector is marked, t (x) indicates that rigid geometric transform relation, s indicate the dimension scale between image to be matched, and R is one 2 × 2 square Battle array indicates the rotation between image to be matched, and o is one 2 × 1 matrix, indicates the translation between image to be matched；
If known one group of initial matching point centering, point set is X={ x on image a1,…,xN}T, corresponding point set on image b For Y={ y1,…,yN}T, calculating wherein n-th pair of match point is correct matched posterior probability pnThere is following posterior probability mathematical modulo Type,
Wherein, γ and σ is model parameter, and e is mathematics constant, and b is preset coefficient；
Step 2, according to point set X={ x1,…,xN}TWith Y={ y1,…,yN}TSolving model parameter, including following sub-step Suddenly,
It step 2.1, is each match point xn, n=1 ..., N search for K nearest neighbours' match point respectively, and K takes pre- If value；
Step 2.2, according to the search result of step 2.1, least square method is used to solve dimension as the weight matrix of N × N W；
Step 2.3, by optimal method solving model parameter s, R, o, γ, σ, including following sub-step,
Step 2.3.1, initialization, including enable γ=γ0, R=I2×2, o=0, P=IN×N, γ0For the default initial of γ Value, enables current iteration number k=1, calculates σ using following model parameter formula,
Wherein, matrix T=(t (x1),…,t(xN))T, tr () indicates to seek the mark of matrix；
Step 2.3.2, updates matrix P, including using gained posterior probability mathematical model in step 1, be calculated N to With point to respectively correct matched posterior probability p1,…,pN, enable P=diag (p1,…,pN), diag indicates diagonal matrix；
Step 2.3.3, calculating parameter s, R, o are as follows,
Using following formula calculating parameter s,
Wherein, matrix By IN×NOmission is denoted as I, matrix Q=(I-W)TP (I-W), λ are preset parameter；
Using following formula calculating parameter R,
R=UDVT
Wherein, D=diag (1, det (UVT)), the determinant of det () representing matrix, matrix U and V pass through singular value decomposition It obtains；Using following formula calculating parameter o,
O=μy-sRμx
Step 2.3.4, parameter s, R, the o being calculated according to step 2.3.3, it is as follows to recalculate parameter γ, σ,
Using following formula calculating parameter γ,
σ is calculated using model parameter formula in step 2.3.1；
Step 2.3.5 differentiates the condition of convergence, including calculating current parameter L, when meeting k=kmaxOr (L-Lold)/ Lold≤ ε terminates iteration, kmaxFor maximum number of iterations, ε is convergence threshold；Otherwise, k=k+1, return step 2.3.2；It is described The calculation formula of parameter L is as follows,
Wherein, LoldIndicate the last L being calculated；
Step 3, initial matching point is calculated to for correctly matched posterior probability, and according to threshold decision initial matching point pair Correct errors, realize it is as follows,
The model parameter solved in the step 2.3 is substituted into posterior probability mathematical model described in step 1, is calculated N-th pair of matching double points is correct matched posterior probability；Work as pnWhen >=threshold, then it is assumed that n-th pair of match point is correct Match；Work as pnWhen < threshold, then it is assumed that n-th pair of match point is the matching of mistake, and wherein threshold is preset judgement threshold Value.
Moreover, step 2.2 includes following sub-step,
Step 2.2.1, as match point xjIt is not the match point x searched in the step 2.1iK neighbours' match point it For the moment, then by WijIt is set as 0, wherein match point xjWith match point xiFor the match point in point set X, and i, j=1 ..., N, WijTable Show the element of weight matrix W the i-th row jth column；
Step 2.2.2,Constraint condition under, solve to obtain using least square method so that following generations Valence function E (W) gets the weight matrix W of minimum value, and cost function E (W) is as follows,
The present invention correspondingly provides a kind of Image Feature Matching system based on local linear migration and rigid model, including With lower module,
Model construction module is correct for establishing the corresponding model of geometric transformation and matching double points between image to be matched The corresponding model of the posterior probability matched, realization is as follows,
For the rigid geometric transformation between image to be matched, foundation transformation mathematical model is as follows,
Y=t (x)=sRx+o
Wherein, if two width image to be matched are image a and image b, x and y are the seat of pixel on image a and image b respectively Vector is marked, t (x) indicates that rigid geometric transform relation, s indicate the dimension scale between image to be matched, and R is one 2 × 2 square Battle array indicates the rotation between image to be matched, and o is one 2 × 1 matrix, indicates the translation between image to be matched；
If known one group of initial matching point centering, point set is X={ x on image a1,…,xN}T, corresponding point set on image b For Y={ y1,…,yN}T, calculating wherein n-th pair of match point is correct matched posterior probability pnThere is following posterior probability mathematical modulo Type,
Wherein, γ and σ is model parameter, and e is mathematics constant, and b is preset coefficient；
Parametric solution module, for according to point set X={ x1,…,xN}TWith Y={ y1,…,yN}TSolving model parameter, packet Following submodule is included,
Neighbours' match point submodule, for being each match point xn, n=1 ..., N search for K nearest neighbours respectively Match point, K take preset value；
Weight matrix submodule is solved for the search result according to neighbours' match point submodule using least square method Dimension is the weight matrix W of N × N；
Optimization submodule is used for by optimal method solving model parameter s, R, o, γ, σ, including to place an order Member,
Initialization unit, for enabling γ=γ0, R=I2×2, o=0, P=IN×N, γ0For the preset initial value of γ, order is worked as Preceding the number of iterations k=1 calculates σ using following model parameter formula,
Wherein, matrix T=(t (x1),…,t(xN))T, tr () indicates to seek the mark of matrix；
Updating unit, for updating gained posterior probability mathematical model, meter in matrix P, including use model construction module It is respectively correct matched posterior probability p to matching double points that calculation, which obtains N,1,…,pN, enable P=diag (p1,…,pN), diag is indicated Diagonal matrix；
First parameter calculation unit, it is as follows for calculating parameter s, R, o,
Using following formula calculating parameter s,
Wherein, matrix By IN×NOmission is denoted as I, matrix Q=(I-W)TP (I-W), λ are preset parameter；
Using following formula calculating parameter R,
R=UDVT
Wherein, D=diag (1, det (UVT)), the determinant of det () representing matrix, matrix U and V pass through singular value decomposition It obtains；Using following formula calculating parameter o,
O=μy-sRμx
Second parameter calculation unit, parameter s, R, o for being calculated according to the first parameter calculation unit, is recalculated Parameter γ, σ is as follows,
Using following formula calculating parameter γ,
σ is calculated using model parameter formula in initialization unit；
Iteration judging unit, for differentiating the condition of convergence, including calculating current parameter L, when meeting k=kmaxOr (L- Lold)/Lold≤ ε terminates iteration, kmaxFor maximum number of iterations, ε is convergence threshold；Otherwise, k=k+1, order updating unit work Make；The calculation formula of the parameter L is as follows,
Wherein, LoldIndicate the last L being calculated；
Judging result module, for calculating initial matching point to for correct matched posterior probability, and according to threshold decision For initial matching point to correcting errors, realization is as follows,
The model parameter solved in the optimization submodule is substituted into posterior probability described in model construction module Mathematical model, it is correct matched posterior probability that n-th pair of matching double points, which is calculated,；Work as pnWhen >=threshold, then it is assumed that the N is correctly to match to match point；Work as pnWhen < threshold, then it is assumed that n-th pair of match point is the matching of mistake, wherein Threshold is preset judgment threshold.
Moreover, weight matrix submodule include with lower unit,
Initial assignment unit, for working as match point xjIt is not the match point x searched in neighbours' match point submodulei K neighbours' match point for the moment, then by WijIt is set as 0, wherein match point xjWith match point xiFor the match point in point set X, And i, j=1 ..., N, WijIndicate the element of weight matrix W the i-th row jth column；
Unit is solved, is used forConstraint condition under, solve to obtain using least square method so that following Cost function E (W) gets the weight matrix W of minimum value, and cost function E (W) is as follows,
The invention has the following advantages that
1, the present invention proposes a kind of unified maximum likelihood frame for Image Feature Matching.With the parameter generally used Model is compared, and this frame can handle nonparametric model situation, and application range is wider.
2, the present invention constructs a kind of constraint of local geometric and optimizes to transformation model.This constraint saves adjacent spy Partial structurtes between sign point, therefore can be used to the case where handling containing a large amount of erroneous matchings, there is stronger robustness.
Specific embodiment
Technical solution of the present invention is further described below by embodiment.
Method proposed by the present invention carries out mathematical modeling to image to be matched rigid transformation first, then preliminary by removal Correct matching is established in the matching for the mistake established in a series of matching double points.This method utilizes the Bayes with hidden variable Maximal possibility estimation carries out mathematical modeling.Meanwhile a kind of geometrical constraint for retaining partial structurtes between adjacent characteristic point has also been constructed, Even if still keeping good robustness there are in the case where a large amount of erroneous matchings in preliminary matches.
Institute of embodiment of the present invention providing method mainly includes 3 steps:
Step 1, establishing the corresponding model of geometric transformation and matching double points between image to be matched is that correct matched posteriority is general The corresponding model of rate；
For the rigid geometric transformation between image to be matched, it is as follows to establish transformation mathematical model:
Y=t (x)=sRx+o
Wherein, if two width image to be matched are image a and image b, x and y are the seat of pixel on image a and image b respectively Vector is marked, t (x) indicates that rigid geometric transform relation, s indicate the dimension scale between image to be matched, and R is one 2 × 2 square Battle array indicates the rotation between image to be matched, and o is one 2 × 1 matrix, indicates the translation between image to be matched；
Due to inevitably wrong, the present invention using matching double points obtained by images match in the prior art progress initial matching Purpose be to remove to wherein mistake, then for point set on known one group of initial matching point centering image a be X= {x1,…,xN}T, corresponding point set is Y={ y on image b1,…,yN}T, midpoint x1Corresponding matching point y1, and so on, point xNIt is right Answer match point yN, comprising N to initial matching point pair, wherein n-th pair of matching double points is correct matched posterior probability pnIt can be by such as Lower posterior probability mathematical model is calculated:
Wherein, γ and σ is model parameter, and e is mathematics constant, and coefficient b can be preassigned by those skilled in the art, or Person uses variable element, and embodiment is using 2/5, i.e.,
Step 2, according to known initial matching point pair, i.e. point set X={ x1,…,xN}T, Y={ y1,…,yN}T, solve institute The parameter of two models in step 1 is stated, includes following sub-step:
It step 2.1, is each match point xn, the search of k-d tree association search method and its is respectively adopted in n=1 ..., N K nearest neighbours' match point；When it is implemented, those skilled in the art can voluntarily preset the value of K, K is taken in the present embodiment 15；
Step 2.2, according to the search result of the step 2.1, least square method is used to solve dimension as the weight of N × N Matrix W；
In embodiment, weight matrix W solution is as follows,
Step 2.2.1, as match point xjIt is not the match point x searched in the step 2.1iK neighbours' match point it For the moment, then by WijIt is set as 0, wherein match point xjWith match point xiFor the match point in point set X, and i, j=1 ..., N, WijTable Show the element of weight matrix W the i-th row jth column；
Step 2.2.2,Constraint condition under, solve to obtain using least square method so that following generations Valence function E (W) gets the weight matrix W of minimum value, and cost function E (W) is as follows:
Step 2.3, unknown mathematical model parameter: s, R, o, γ, σ is solved by optimal method；
The solution procedure of unknown mathematical model parameter s, R, o, γ, σ are as follows in embodiment:
Step 2.3.1, initiation parameter；Enable parameter γ=γ0(when it is implemented, those skilled in the art can be voluntarily Set the initial value γ of γ0, embodiment takes γ0=0.9), matrix R=I2×2(i.e. dimension is 2 × 2, square of element all 1 Battle array), parameter o=0, matrix P=IN×N(i.e. dimension be N × N, element all 1 matrix), current iteration number k=1, use Following model parameter formula calculate σ:
Wherein, matrix T=(t (x1),…,t(xN))T, tr () indicates to seek the mark of matrix；
Step 2.3.2 updates matrix P:
Using posterior probability mathematical model described in step 1, it is calculated the 1st ..., N is respectively correct to matching double points Matched posterior probability p1,…,pN, enable P=diag (p1,…,pN), diag indicates diagonal matrix；
Step 2.3.3, calculating parameter s, R, o；
Using following formula calculating parameter s:
Wherein, each matrix IN×1I.e. dimension be N × 1, element all 1 matrix, by IN×NOmission is denoted as I, matrix Q=(I-W)TP (I-W), wherein λ is One preset parameter, those skilled in the art can preset value when specific implementation, take 100 in the present embodiment；
Using following formula calculating parameter R:
R=UDVT
Wherein, D=diag (1, det (UVT)), the determinant of det () representing matrix, matrix U and V pass through singular value decomposition It obtains, formula is as follows:
Wherein, svd () indicates to carry out matrix singular value decomposition, and S is the result of singular value decomposition；
Using following formula calculating parameter o:
O=μy-sRμx
Step 2.3.4, parameter s, R, the o being calculated according to step 2.3.3 can recalculate ginseng based on current t (x) Number γ, σ；Using following formula calculating parameter γ:
σ is calculated using model parameter formula described in step 2.3.1；
Step 2.3.5 differentiates the condition of convergence:
Current parameter L is calculated, when meeting k=kmaxOr (L-Lold)/Lold≤ ε terminates iteration；Otherwise, i.e. k < kmax And (L-Lold)/Lold> ε, k=k+1 return to step 2.3.2, when it is implemented, those skilled in the art can free default maximum The number of iterations kmaxValue, the value of the present embodiment is 50；
Wherein, ε is convergence threshold, and those skilled in the art can preset value when specific implementation, such as take 0.0001, Lold Indicate that the last L for executing step 2.3.5 and being calculated can enable L when executing step 2.3.5 for the first timeoldIt is biggish for one Initial value, such as 10 6 powers, execute can be used the calculated result that last time execution executes step 2.3.5 later.The calculating of L Formula is as follows:
Step 3, initial matching point is calculated to for correctly matched posterior probability, and according to threshold decision initial matching point pair Correct errors；
The parameter solved in the step 2.3 is substituted into posterior probability mathematical model described in step 1, is calculated n-th It is correct matched posterior probability to matching double points；Work as pnWhen >=threshold, then it is assumed that n-th pair of match point is correctly to match； Work as pnWhen < threshold, then it is assumed that n-th pair of match point is the matching of mistake, and wherein threshold is judgment threshold, specific real Shi Shi, those skilled in the art can free default judgment threshold value, take 0.5 in the present embodiment.
When it is implemented, software mode, which can be used, in the above process realizes automatic running.It can also be mentioned using modular mode For corresponding system, the embodiment of the present invention correspondingly provides a kind of Image Feature Matching based on local linear migration and rigid model System comprises the following modules,
Model construction module is correct for establishing the corresponding model of geometric transformation and matching double points between image to be matched The corresponding model of the posterior probability matched, realization is as follows,
For the rigid geometric transformation between image to be matched, foundation transformation mathematical model is as follows,
Y=t (x)=sRx+o
Wherein, if two width image to be matched are image a and image b, x and y are the seat of pixel on image a and image b respectively Vector is marked, t (x) indicates that rigid geometric transform relation, s indicate the dimension scale between image to be matched, and R is one 2 × 2 square Battle array indicates the rotation between image to be matched, and o is one 2 × 1 matrix, indicates the translation between image to be matched；
If known one group of initial matching point centering, point set is X={ x on image a1,…,xN}T, corresponding point set on image b For Y={ y1,…,yN}T, calculating wherein n-th pair of match point is correct matched posterior probability pnThere is following posterior probability mathematical modulo Type,
Wherein, γ and σ is model parameter, and e is mathematics constant, and b is preset coefficient；
Parametric solution module, for according to point set X={ x1,…,xN}TWith Y={ y1,…,yN}TSolving model parameter, packet Following submodule is included,
Neighbours' match point submodule, for being each match point xn, n=1 ..., N search for K nearest neighbours respectively Match point, K take preset value；
Weight matrix submodule is solved for the search result according to neighbours' match point submodule using least square method Dimension is the weight matrix W of N × N；
Optimization submodule is used for by optimal method solving model parameter s, R, o, γ, σ, including to place an order Member,
Initialization unit, for enabling γ=γ0, R=I2×2, o=0, P=IN×N, γ0For the preset initial value of γ, order is worked as Preceding the number of iterations k=1 calculates σ using following model parameter formula,
Wherein, matrix T=(t (x1),…,t(xN))T, tr () indicates to seek the mark of matrix；
Updating unit, for updating gained posterior probability mathematical model, meter in matrix P, including use model construction module It is respectively correct matched posterior probability p to matching double points that calculation, which obtains N,1,…,pN, enable P=diag (p1,…,pN), diag is indicated Diagonal matrix；
First parameter calculation unit, it is as follows for calculating parameter s, R, o,
Using following formula calculating parameter s,
Wherein, matrix By IN×NOmission is denoted as I, matrix Q=(I-W)TP (I-W), λ are preset parameter；
Using following formula calculating parameter R,
R=UDVT
Wherein, D=diag (1, det (UVT)), the determinant of det () representing matrix, matrix U and V pass through singular value decomposition It obtains；Using following formula calculating parameter o,
O=μy-sRμx
Second parameter calculation unit, parameter s, R, o for being calculated according to the first parameter calculation unit, is recalculated Parameter γ, σ is as follows,
Using following formula calculating parameter γ,
σ is calculated using model parameter formula in initialization unit；
Iteration judging unit, for differentiating the condition of convergence, including calculating current parameter L, when meeting k=kmaxOr (L- Lold)/Lold≤ ε terminates iteration, kmaxFor maximum number of iterations, ε is convergence threshold；Otherwise, k=k+1, order updating unit work Make；The calculation formula of the parameter L is as follows,
Wherein, LoldIndicate the last L being calculated；
Judging result module, for calculating initial matching point to for correct matched posterior probability, and according to threshold decision For initial matching point to correcting errors, realization is as follows,
The model parameter solved in the optimization submodule is substituted into posterior probability described in model construction module Mathematical model, it is correct matched posterior probability that n-th pair of matching double points, which is calculated,；Work as pnWhen >=threshold, then it is assumed that the N is correctly to match to match point；Work as pnWhen < threshold, then it is assumed that n-th pair of match point is the matching of mistake, wherein Threshold is preset judgment threshold.
Further, weight matrix submodule include with lower unit,
Initial assignment unit, for working as match point xjIt is not the match point x searched in neighbours' match point submodulei K neighbours' match point for the moment, then by WijIt is set as 0, wherein match point xjWith match point xiFor the match point in point set X, And i, j=1 ..., N, WijIndicate the element of weight matrix W the i-th row jth column；
Unit is solved, is used forConstraint condition under, solve to obtain using least square method so that following Cost function E (W) gets the weight matrix W of minimum value, and cost function E (W) is as follows,
RANSAC, ICF, GS method of selection and the present invention carry out the comparison of images match.Comparing result such as following table, wherein just True rate refer to be in matching double points that method finally provides correct matching double points ratio；Missing rate refers to method in screening process The middle ratio for correct matching double points being judged as error matching points pair.It can be seen that this method used time is most short, accuracy rate highest is lost Leak rate is minimum.
Method Contrast on effect table
Method Time (second) Accuracy Missing rate
RANSAC 16.32 81.38% 1.36%
ICF 2.13 91.27% 18.39%
GS 0.98 70.95% 20.97%
The present invention 0.34 98.7% 0.87%
Specific embodiment described herein is only an example for the spirit of the invention.The neck of technology belonging to the present invention The technical staff in domain can make various modifications or additions to the described embodiments or be substituted using similar fashion, However, it does not deviate from the spirit of the invention or beyond the scope of the appended claims.

Claims (4)

1. a kind of Image Feature Matching method based on local linear migration and rigid model, it is characterised in that: including following step Suddenly,
Step 1, establishing the corresponding model of geometric transformation and matching double points between image to be matched is correct matched posterior probability phase The model answered, realization is as follows,
For the rigid geometric transformation between image to be matched, foundation transformation mathematical model is as follows,
Y=t (x)=sRx+o
Wherein, if two width image to be matched be image a and image b, x and y be respectively pixel on image a and image b coordinate to Amount, t (x) indicate that rigid geometric transform relation, s indicate the dimension scale between image to be matched, and R is one 2 × 2 matrix, table Show the rotation between image to be matched, o is one 2 × 1 matrix, indicates the translation between image to be matched；
If known one group of initial matching point centering, point set is X={ x on image a1,…,xN}T, corresponding point set is Y on image b ={ y1,…,yN}T, calculating wherein n-th pair of match point is correct matched posterior probability pnThere is following posterior probability mathematical model,
Wherein, γ and σ is model parameter, and e is mathematics constant, and b is preset coefficient；
Step 2, according to point set X={ x1,…,xN}TWith Y={ y1,…,yN}TSolving model parameter, including following sub-step,
It step 2.1, is each match point xn, n=1 ..., N search for K nearest neighbours' match point respectively, and K takes preset value；
Step 2.2, according to the search result of step 2.1, least square method is used to solve dimension as the weight matrix W of N × N；
Step 2.3, by optimal method solving model parameter s, R, o, γ, σ, including following sub-step,
Step 2.3.1, initialization, including enable γ=γ0, R=I2×2, o=0, P=IN×N, γ0For the preset initial value of γ, enable Current iteration number k=1 calculates σ using following model parameter formula,
Wherein, matrix T=(t (x1),…,t(xN))T, tr () indicates to seek the mark of matrix；
Step 2.3.2 updates matrix P, including using gained posterior probability mathematical model in step 1, N is calculated to match point To respectively correct matched posterior probability p1,…,pN, enable P=diag (p1,…,pN), diag indicates diagonal matrix；
Step 2.3.3, calculating parameter s, R, o are as follows,
Using following formula calculating parameter s,
Wherein, matrix By IN×NOmission is denoted as I, matrix Q=(I-W)TP (I-W), λ are preset parameter；
Using following formula calculating parameter R,
R=UDVT
Wherein, D=diag (1, det (UVT)), the determinant of det () representing matrix, matrix U and V are obtained by singular value decomposition ?；Using following formula calculating parameter o,
O=μy-sRμx
Step 2.3.4, parameter s, R, the o being calculated according to step 2.3.3, recalculate that parameter γ, σ is as follows, and use is following Formula calculating parameter γ,
σ is calculated using model parameter formula in step 2.3.1；
Step 2.3.5 differentiates the condition of convergence, including calculating current parameter L, when meeting k=kmaxOr (L-Lold)/Lold≤ ε terminates iteration, kmaxFor maximum number of iterations, ε is convergence threshold；Otherwise, k=k+1, return step 2.3.2；The parameter L Calculation formula it is as follows,
Wherein, LoldIndicate the last L being calculated；
Step 3, initial matching point is calculated to for correctly matched posterior probability, and just according to threshold decision initial matching point pair Accidentally, realization is as follows,
The model parameter solved in the step 2.3 is substituted into posterior probability mathematical model described in step 1, is calculated n-th It is correct matched posterior probability to matching double points；Work as pnWhen >=threshold, then it is assumed that n-th pair of match point is correctly to match； Work as pnWhen < threshold, then it is assumed that n-th pair of match point is the matching of mistake, and wherein threshold is preset judgment threshold.
2. the Image Feature Matching method according to claim 1 based on local linear migration and rigid model, feature exist In: step 2.2 includes following sub-step,
Step 2.2.1, as match point xjIt is not the match point x searched in the step 2.1iOne of K neighbours' match point When, then by WijIt is set as 0, wherein match point xjWith match point xiFor the match point in point set X, and i, j=1 ..., N, WijIt indicates The element of weight matrix W the i-th row jth column；
Step 2.2.2,Constraint condition under, solve to obtain using least square method so that following cost letters Number E (W) gets the weight matrix W of minimum value, and cost function E (W) is as follows,
3. a kind of Image Feature Matching system based on local linear migration and rigid model, it is characterised in that: including with lower die Block,
Model construction module is correct matched for establishing the corresponding model of geometric transformation and matching double points between image to be matched The corresponding model of posterior probability, realization is as follows,
For the rigid geometric transformation between image to be matched, foundation transformation mathematical model is as follows,
Y=t (x)=sRx+o
Wherein, if two width image to be matched be image a and image b, x and y be respectively pixel on image a and image b coordinate to Amount, t (x) indicate that rigid geometric transform relation, s indicate the dimension scale between image to be matched, and R is one 2 × 2 matrix, table Show the rotation between image to be matched, o is one 2 × 1 matrix, indicates the translation between image to be matched；
If known one group of initial matching point centering, point set is X={ x on image a1,…,xN}T, corresponding point set is Y on image b ={ y1,…,yN}T, calculating wherein n-th pair of match point is correct matched posterior probability pnThere is following posterior probability mathematical model,
Wherein, γ and σ is model parameter, and e is mathematics constant, and b is preset coefficient；
Parametric solution module, for according to point set X={ x1,…,xN}TWith Y={ y1,…,yN}TSolving model parameter, including with Lower submodule,
Neighbours' match point submodule, for being each match point xn, n=1 ..., N search for nearest K neighbours matching respectively Point, K take preset value；
Weight matrix submodule solves dimension using least square method for the search result according to neighbours' match point submodule For the weight matrix W of N × N；
Optimization submodule, for by optimal method solving model parameter s, R, o, γ, σ, including with lower unit,
Initialization unit, for enabling γ=γ0, R=I2×2, o=0, P=IN×N, γ0For the preset initial value of γ, current change is enabled Generation number k=1 calculates σ using following model parameter formula,
Wherein, matrix T=(t (x1),…,t(xN))T, tr () indicates to seek the mark of matrix；
Updating unit is calculated for updating gained posterior probability mathematical model in matrix P, including use model construction module It is respectively correct matched posterior probability p to matching double points to N1,…,pN, enable P=diag (p1,…,pN), diag indicates diagonal Matrix；
First parameter calculation unit, it is as follows for calculating parameter s, R, o,
Using following formula calculating parameter s,
Wherein, matrix By IN×NOmission is denoted as I, matrix Q=(I-W)TP (I-W), λ are preset parameter；
Using following formula calculating parameter R,
R=UDVT
Wherein, D=diag (1, det (UVT)), the determinant of det () representing matrix, matrix U and V are obtained by singular value decomposition ?；Using following formula calculating parameter o,
O=μy-sRμx
Second parameter calculation unit, parameter s, R, o for being calculated according to the first parameter calculation unit, recalculates parameter γ, σ are as follows,
Using following formula calculating parameter γ,
σ is calculated using model parameter formula in initialization unit；
Iteration judging unit, for differentiating the condition of convergence, including calculating current parameter L, when meeting k=kmaxOr (L- Lold)/Lold≤ ε terminates iteration, kmaxFor maximum number of iterations, ε is convergence threshold；Otherwise, k=k+1, order updating unit work Make；The calculation formula of the parameter L is as follows,
Wherein, LoldIndicate the last L being calculated；
Judging result module, for calculating initial matching point to for correct matched posterior probability, and it is initial according to threshold decision Matching double points are corrected errors, and realization is as follows,
The model parameter solved in the optimization submodule is substituted into posterior probability mathematics described in model construction module Model, it is correct matched posterior probability that n-th pair of matching double points, which is calculated,；Work as pnWhen >=threshold, then it is assumed that n-th pair Match point is correctly to match；Work as pnWhen < threshold, then it is assumed that n-th pair of match point is the matching of mistake, wherein threshold For preset judgment threshold.
4. the Image Feature Matching system according to claim 3 based on local linear migration and rigid model, feature exist In: weight matrix submodule include with lower unit,
Initial assignment unit, for working as match point xjIt is not the match point x searched in neighbours' match point submoduleiK Neighbours' match point a period of time, then by WijIt is set as 0, wherein match point xjWith match point xiFor the match point in point set X, and i, j =1 ..., N, WijIndicate the element of weight matrix W the i-th row jth column；
Unit is solved, is used forConstraint condition under, solve to obtain using least square method so that following costs Function E (W) gets the weight matrix W of minimum value, and cost function E (W) is as follows,
CN201510807463.3A 2015-11-19 2015-11-19 Image Feature Matching method and system based on local linear migration and rigid model Active CN105469112B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201510807463.3A CN105469112B (en) 2015-11-19 2015-11-19 Image Feature Matching method and system based on local linear migration and rigid model

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201510807463.3A CN105469112B (en) 2015-11-19 2015-11-19 Image Feature Matching method and system based on local linear migration and rigid model

Publications (2)

Publication Number Publication Date
CN105469112A CN105469112A (en) 2016-04-06
CN105469112B true CN105469112B (en) 2018-12-14

Family

ID=55606783

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201510807463.3A Active CN105469112B (en) 2015-11-19 2015-11-19 Image Feature Matching method and system based on local linear migration and rigid model

Country Status (1)

Country Link
CN (1) CN105469112B (en)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105488754B (en) * 2015-11-19 2018-09-21 武汉大学 Image Feature Matching method and system based on local linear migration and affine transformation
CN109448031B (en) * 2018-09-28 2021-04-16 武汉大学 Image registration method and system based on Gaussian field constraint and manifold regularization

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103778433A (en) * 2014-01-15 2014-05-07 东莞华中科技大学制造工程研究院 Generalized-point-set matching method based on distances from points to lines
CN104599277A (en) * 2015-01-27 2015-05-06 中国科学院空间科学与应用研究中心 Image registration method for area-preserving affine transformation
CN104732546A (en) * 2015-04-02 2015-06-24 西安电子科技大学 Non-rigid SAR image registration method based on region similarity and local spatial constraint

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103778433A (en) * 2014-01-15 2014-05-07 东莞华中科技大学制造工程研究院 Generalized-point-set matching method based on distances from points to lines
CN104599277A (en) * 2015-01-27 2015-05-06 中国科学院空间科学与应用研究中心 Image registration method for area-preserving affine transformation
CN104732546A (en) * 2015-04-02 2015-06-24 西安电子科技大学 Non-rigid SAR image registration method based on region similarity and local spatial constraint

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
Robust Point Matching via Vector Field Consensus;Jiayi Ma ect.;《IEEE TRANSACTIONS ON IMAGE PROCESSING》;20140831;第23卷(第4期);第1706-1721页 *
基于t分布混合模型的点集非刚性配准算法;周志勇 等;《光学 精密工程》;20130930;第21卷(第9期);第2405-2420页 *

Also Published As

Publication number Publication date
CN105469112A (en) 2016-04-06

Similar Documents

Publication Publication Date Title
CN105469110B (en) Non-rigid transformation Image Feature Matching method and system based on local linear migration
CN109285110B (en) Infrared visible light image registration method and system based on robust matching and transformation
Min et al. A two-stream mutual attention network for semi-supervised biomedical segmentation with noisy labels
CN105488754B (en) Image Feature Matching method and system based on local linear migration and affine transformation
CN112581515B (en) Outdoor scene point cloud registration method based on graph neural network
CN109872305B (en) No-reference stereo image quality evaluation method based on quality map generation network
CN113822284B (en) RGBD image semantic segmentation method based on boundary attention
CN104732546B (en) The non-rigid SAR image registration method of region similitude and local space constraint
CN101650828B (en) Method for reducing random error of round object location in camera calibration
CN103440659B (en) Based on starry sky image Distortion Detect and the method for estimation of star pattern matching
CN106709943B (en) A kind of point cloud registration method based on optimal transmission
CN108335328B (en) Camera attitude estimation method and camera attitude estimation device
CN105469112B (en) Image Feature Matching method and system based on local linear migration and rigid model
CN111488498A (en) Node-graph cross-layer graph matching method and system based on graph neural network
CN109448031B (en) Image registration method and system based on Gaussian field constraint and manifold regularization
CN106778925B (en) Face recognition pose over-complete face automatic registration method and device
CN117253062A (en) Relay contact image characteristic quick matching method under any gesture
CN110163865B (en) Sampling method for unbalanced data in model fitting
CN114399547B (en) Monocular SLAM robust initialization method based on multiframe
CN105938615B (en) Feature based is oriented to the method for registering images and system of GMM and edge image
CN116188249A (en) Remote sensing image registration method based on image block three-stage matching
CN108205805B (en) Automatic establishment method for dense correspondence of cone beam CT image intermediate pixels
CN115063890A (en) Human body posture estimation method based on two-stage weighted mean square loss function
CN113781559B (en) Robust abnormal matching point eliminating method and image indoor positioning method
CN112304311B (en) Method for solving BA problem based on differential evolution algorithm for SLAM process

Legal Events

Date Code Title Description
C06 Publication
PB01 Publication
C10 Entry into substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant