CN101567051B  Image matching method based on characteristic points  Google Patents
Image matching method based on characteristic points Download PDFInfo
 Publication number
 CN101567051B CN101567051B CN2009100524538A CN200910052453A CN101567051B CN 101567051 B CN101567051 B CN 101567051B CN 2009100524538 A CN2009100524538 A CN 2009100524538A CN 200910052453 A CN200910052453 A CN 200910052453A CN 101567051 B CN101567051 B CN 101567051B
 Authority
 CN
 China
 Prior art keywords
 sign indicating
 point
 corner sign
 ring type
 similarity
 Prior art date
Links
 238000005457 optimization Methods 0.000 claims abstract description 8
 238000004422 calculation algorithm Methods 0.000 claims description 45
 230000001131 transforming Effects 0.000 claims description 31
 238000006243 chemical reactions Methods 0.000 claims description 25
 230000000875 corresponding Effects 0.000 claims description 21
 238000000034 methods Methods 0.000 claims description 18
 230000001808 coupling Effects 0.000 claims description 17
 238000010168 coupling process Methods 0.000 claims description 17
 238000005859 coupling reactions Methods 0.000 claims description 17
 239000004744 fabrics Substances 0.000 claims description 8
 238000000605 extraction Methods 0.000 claims description 6
 238000010276 construction Methods 0.000 claims description 4
 230000023298 conjugation with cellular fusion Effects 0.000 claims description 3
 230000013011 mating Effects 0.000 claims description 3
 239000000203 mixtures Substances 0.000 claims description 3
 230000021037 unidirectional conjugation Effects 0.000 claims description 3
 125000004122 cyclic group Chemical group 0.000 claims description 2
 230000000717 retained Effects 0.000 claims description 2
 230000004927 fusion Effects 0.000 abstract description 2
 230000000052 comparative effects Effects 0.000 description 4
 238000010586 diagrams Methods 0.000 description 4
 238000004364 calculation methods Methods 0.000 description 3
 230000002950 deficient Effects 0.000 description 3
 230000001133 acceleration Effects 0.000 description 2
 238000004458 analytical methods Methods 0.000 description 2
 230000000694 effects Effects 0.000 description 2
 238000005516 engineering processes Methods 0.000 description 2
 238000002372 labelling Methods 0.000 description 2
 230000000007 visual effect Effects 0.000 description 2
 281000188848 Facebook Graph Search companies 0.000 description 1
 281000019761 Intel, Corp. companies 0.000 description 1
 239000012141 concentrates Substances 0.000 description 1
 238000003745 diagnosis Methods 0.000 description 1
 201000010099 diseases Diseases 0.000 description 1
 230000003760 hair shine Effects 0.000 description 1
 230000002401 inhibitory effects Effects 0.000 description 1
 239000011159 matrix materials Substances 0.000 description 1
 238000003909 pattern recognition Methods 0.000 description 1
 238000005070 sampling Methods 0.000 description 1
 238000002054 transplantation Methods 0.000 description 1
 238000005303 weighing Methods 0.000 description 1
Abstract
Description
Technical field
The invention belongs to patternrecognition and technical field of image processing, be specifically related to a kind of method for registering images based on unique point.This method can be used for the image registration based on unique point, image mosaic with inlay motion tracking, the information fusion of medical science single mode and multimodality images, and CBIR etc.
Background technology
Point Pattern Matching is appreciated that to be the images match based on characteristic, estimates from two points to be matched concentrated search corresponding relation and to the mapping between the point set.Point Pattern Matching is widely used in remote sensing (image mosaic, image mosaic), medical image (state of an illness diagnosis and tracking) and computer vision (target or scene Recognition, motion tracking).
Though a lot of in recent years outstanding Point Pattern Matching algorithms are suggested, almost there is not a kind of algorithm on computing time and precision, to satisfy application requirements simultaneously.Present disclosed algorithm is broadly divided into 5 types: clustering algorithm (document that sees reference [1] [2]); Shape contextual algorithms (document that sees reference [3]); Loose labeling algorithm (document that sees reference [4] [5]), progressively optimized Algorithm (document that sees reference [6]) is schemed matching algorithm (document that sees reference [7]).
Wherein,
Clustering algorithm hypothesis transformation model is similarity transformation; Only need two pairs of match points because estimate similarity transformation; Algorithm is at first concentrated from two points to be matched and selected two pairs of points to make up, and estimates conversion coefficient then, and is right with other points that possibly mate of conversion checking of estimating.Because also is correct with the point of correct match to the conversion of estimating, carry out cluster with this conversion, then can might gather such to institute by correct match point, and if the point of correct match many, class also can be very big accordingly; On the contrary, if with the conversion to estimating of the point of erroneous matching, then this conversion is wrong certainly, and it is exactly randomness that erroneous conversions has characteristics.Erroneous matching that is to say that making sorter with erroneous conversions carries out cluster, then only converges to seldom point mostly at random.Algorithm carries out cluster according to this character to two feature point sets, and in the cluster result, maximum kind will be as corresponding relation output to be used for subsequent treatment.The defective of this type algorithm is that computation complexity is high, and is consuming time, and if two feature point set difference in size very big, algorithm performance also can nonconstant.Hough transformation was used to preservice afterwards with speedup computation, yet coupling remains a kind of work very consuming time, and was and very many at noise spot, and perhaps under two very large situation of point set difference in size, algorithm not necessarily guarantees to find out best correspondence.
The shape contextual algorithms also is to study the Point Pattern Matching technology of comparative heat in the last few years.The shape contextdescriptive point concentrate other unique points space distribution information of certain unique point relatively.The unique point of mating in the same shape of two width of cloth images has the shapes similar context, and each unique point is corresponding direction histogram is to represent other unique points space constraint of oneself relatively.This algorithm can solve the Point Pattern Matching of nonrigid transformation, and most of similar algorithm use SPL estimates conversion, has dependence yet the problem of this type algorithm is sampling to unique point; Image background can not be too complicated, and the test of various algorithms is used for simple form fit also mostly based on generated data; Algorithm performance is relatively good; But the computation complexity of this type algorithm is than higher, and each unique point all need utilize every other point to calculate hereinafter in shape, therefore in complex scene; Under the situation that the object of identification relatively is difficult to again cut apart, the poor performance of algorithm.
Loose labeling algorithm defines a kind of probability distribution on transformation relation; Be optimized with discrete logarithm then, algorithm is actually the process of an iteration, since a thicker coupling; Progressively upgrade the matching probability matrix then, this algorithm is easy to converge on local extremum.
Progressively optimized Algorithm also is to utilize global characteristic point to make up a compatible function, and through upgrading compatible function to solve correspondence, this algorithm has identical defective with loose algorithm.Two types of algorithms are bigger at feature point set, and algorithm performance is all poor under the many again situation of noise, and can not guarantee that algorithm can restrain under various situation.
The figure matching algorithm is the Point Pattern Matching algorithm of the comparative heat of recent researches.Through two point sets to be matched are built into weighted graph, attributed relational graph etc. are searched for compatible part then in two figure, and the part of promptly mating is probably perhaps carried out the subgraph coupling, perhaps shines upon coupling fully, guarantees that majorized function is minimum.This type algorithm can be summed up as the graph search problem, is the difficult problem of a NP, can not guarantee various survey examples are all separated.A lot of algorithms also all are to utilize majorized function to estimate; Obtain more excellent separating rather than optimum solution; Another defective of figure coupling is to handle two point sets that all contain much noise point, except computing time is long, so this algorithm also is difficult to obtain rational corresponding relation.
The list of references relevant with the present invention has:
[1]A.Goshtasby.Description?and?discrimination?of?planar?shapes?using?shape?matrices.IEEE?Trans.Pattern?Analysis?and?Machine?Intelligence，7：738743，1985.
[2]S.H.Chang，F.H.Cheng，W.H.Hsu，and?G.Z.Wu.Fast?algorithm?for?point?pattern?matching：Invariant?to?translations，rotations?and?scale?changes.Pattern?Recognition，30：311320，1997.
[3]S.Belongie，J.Malik，and?J.Puzicha，“Shape?Matching?and?Object?Recognition?Using?ShapeContexts，”IEEE?Trans.Pattern?Analysis?and?Machine?Intelligence，vol.24，no.4，pp.509522，Apr.2002.
[4]W.J.Christmas，J.Kittler，and?M.Petrou，“Structural?Matching?in?Computer?Vision?UsingProbabilistic?Relaxation，”IEEE?Trans.Pattern?Analysis?and?Machine?Intelligence，vol.17，no.8，pp.749764，Aug.1994.
[5]R.C.Wilson?and?E.R.Hancock，“Structural?Matching?by?Discrete?Relaxation，”IEEE?Trans.Pattern?Analysis?and?Machine?Intelligence，vol.19，no.6，pp.634648，June?1997.
[6]S.Gold?and?A.Rangarajan，“A?Graduated?Assignment?Algorithm?for?Graph?Matching，”IEEE?Trans.Pattern?Analysis?and?Machine?Intelligence，vol.18，no.4，pp.377388，Apr.1996.
[7]Tibe’rio?S.Caetano，Terry?Caelli，Fellow，IEEE，Dale?Schuurmans，and?Dante?A.C.Barone.Graphical?Models?and?Point?Pattern?Matching.IEEE?Trans.Pattern?Analysis?and?MachineIntelligence，28：16461663，2006.
Summary of the invention
The present invention is intended to overcome the deficiency that exists in the prior art, proposes a kind of method for registering images based on unique point.
The present invention at first proposes a kind of ring type corner sign indicating number model based on unique point k nearest neighbor space structure, the similarity between the sign indicating number of ring type corner is explored and used it for the similarity of the local space of weighing two unique points.Described similar ring type corner sign indicating number has not only been described the similarity of the local space structure between two unique points, can also be used to estimate local map.The sign indicating number modelling of ring type corner is on the similarity transformation basis, and the present invention has carried out loose simultaneously to ring type corner sign indicating number simulated condition, makes matching algorithm have robustness to a certain degree affine and view transformation.Similarity between the unique point by they related maximal phase confirm like the similar length of ring type corner sign indicating number.The present invention at first carries out structure matching according to the similarity of the local space structure of unique point, utilize then match point related local map cluster be optimized coupling.The result of optimization of matching is used to estimate optimum mapping.Aspect the acceleration matching process, the present invention has avoided a large amount of unnecessary computations through set up index in the search volume, thereby guarantees that matching algorithm has also reached when guaranteeing precision near the requirement of application in real time.
The invention discloses a kind of algorithm of Point Pattern Matching fast and effectively, its process flow diagram is seen accompanying drawing 1.Comprise that the step is following rapid:
Step 1): feature extraction; Two images to be matched are carried out feature extraction, and matching algorithm is based on image characteristic point, and coupling extraction in early stage has convergentdivergent; The feature point set of rotation and translation invariance also can have affine and view transformation to a certain degree as input between the point set;
Step 2): similarity is calculated; The ring type corner sign indicating number of the k nearest neighbor structure of extract minutiae and construction feature point; Index is set up in the search volume; Utilize maximal phase between binary chop and the increment matching algorithm calculated characteristics point like ring type corner sign indicating number, and with maximal phase like the similar length of ring type corner as the similarity between two unique points, confirm the unique point that possibly mate when calculating similarity;
Step 3): structure matching; Similarity between the unique point has been described the similarity of the local space structure of unique point in the physical sense; The present invention is according to the similarity between unique point size, and the unique point that possibly mate is divided into three types: unique point class badClass that can not coupling, possibly mate but mate not unique unique point class unknownClass; The unique point class goodClass that possibly mate carries out local map to the match point among the goodClass simultaneously and estimates;
Step 4): optimization of matching, utilize match point among the goodClass related local map carry out cluster, in cluster result, what the definition element was maximum type be maximum kind, the class of element more than second be inferior class greatly.If simultaneously greater than inferior big dvielement number, perhaps maximum kind and the ratio of inferior big type element number are greater than certain threshold value greater than certain threshold value for the maximum kind element number, maximum kind will be estimated optimal transformation as matching result;
Step 5): optimal transformation estimates that the present invention supposes being transformed to except similarity transformation between the image, also has affine and view transformation to a certain degree, and the conversion estimation model is an affined transformation.
Particularly,
Two point sets in the step 1 of the present invention are defined as S and T, and size is respectively n _{1}And n _{2}, while defining point P _{i}∈ S, Q _{j}∈ T.
Ring type corner sign indicating number in the described step 2 is defined as:
Shown in accompanying drawing 2, the k nearest neighbor space structure of unique point can pass through a ring type corner representation.For defining ring type corner sign indicating number better, the present invention with the combination at the angle of every limit and its counterclockwise adjacency in the accompanying drawing 2 (a) be defined as a corner sign indicating number (EdgeAngle Code) EAC=(E, θ).Accompanying drawing 2 (a) has K corner sign indicating number (E _{i}, θ _{i}), i=0,1 ..., K1.Also successively they being linked to be a loop configuration then obtains unique point P by counterclockwise sequential encoding to this K corner sign indicating number _{c}Ring type corner sign indicating number, shown in accompanying drawing 2 (b), the present invention defines the length that K is a ring type corner sign indicating number.
The defined ring type of the present invention corner sign indicating number has following characteristic:
1) connectedness: sign indicating number begins to travel through forward from arbitrary corner, can all corner sign indicating numbers be traveled through in order to finish.For ease of expression, the present invention will travel through subscript and be defined as m _{t}, wherein t representes nonnegative integer, and m _{t}Represent that then t gets surplus result to K, do to guarantee that subscript can not crossed the border when the corner sign indicating number is traveled through like this.
2) space constraint property: the order of corner sign indicating number has been described the restriction relation that the neighbours of unique point are ordered.Corner sign indicating number order
3) equivalence: same unique point can have K different ring type corner sign indicating number, but they have described the same local space structure of unique point all of equal valuely.(a) can know by accompanying drawing 2, and during as first corner sign indicating number, the present invention will obtain different ring type corner sign indicating numbers with different neighbours point, but they are of equal value in the physical sense, have promptly all described identical space structure.The present invention is defined as equivalence to this specific character of ring type corner sign indicating number.
Similar ring type corner sign indicating number definition in the described step 2 is specially:
Shown in accompanying drawing 3; If it is proportional that two ring type corner sign indicating numbers begin to satisfy corresponding sides from a certain opposite side; Corresponding angles are proportional and equal 1; The present invention just says that these two ring type corner sign indicating numbers are similar, and the length of one of them ring type corner sign indicating number is defined as the similar length of similar ring type corner sign indicating number, and the space structure that similar ring type corner sign indicating number is corresponding is defined as similar space structure.Similar ring type corner sign indicating number shown in the accompanying drawing 3 is from (E _{0}, E ' _{0}) begin to satisfy following formula:
Be calculation procedure two described similar ring type corner sign indicating numbers, the present invention has also defined the add operation of two adjacent corner sign indicating numbers, is specially:
The result of two adjacent corner sign indicating number add operations still is a corner sign indicating number, and the limit of new corner sign indicating number is the limit of first corner sign indicating number, and the angle of new corner sign indicating number then is the angle sum of two addition corner sign indicating numbers.In fact the physical significance of phase add operation has been represented two adjacent neighbours' points of counter clockwise direction are merged, and is about to second neighbours' point deletion, then the angle of the counterclockwise adjacency in article one limit is upgraded.
Be calculation procedure two described similar ring type corner sign indicating numbers, the present invention has also defined the compare operation of two corner sign indicating numbers, if two corner sign indicating number (E _{1}, θ _{1}), (E ' _{1}, θ ' _{1}) satisfy
Wherein ε is the error control factor, approaches 0 positive number, (E _{2}, θ _{2}), (E ' _{2}, θ ' _{2}) be respectively (E _{1}, θ _{1}), (E ' _{1}, θ ' _{1}) the corner sign indicating number of next adjacency, the present invention is regulation (E just _{1}, θ _{1}), (E ' _{1}, θ ' _{1}) compare operation do not need add operation, the physical significance that does not need add operation be the expression by E _{1}, θ _{1}, E _{2}The triangle that constitutes with by E ' _{1}, θ ' _{1}, E ' _{2}The triangle that constitutes is similar.
Described step 2 similarity is calculated and is specially:
(1) set up the search volume, the ring type corner sign indicating number of the k nearest neighbor of extract minutiae and construction feature point, the search volume is that impact point is concentrated the n on the ring type corner sign indicating number of all unique points _{2}K corner sign indicating number set up index with acceleration search to these corner sign indicating numbers by angle.Synoptic diagram is seen in the middle of the accompanying drawing 4 and the right two table.
(2) confirm object search, the coupling purpose is to find match point.Therefore the single object search is P _{i}Ring type corner sign indicating number related each corner sign indicating number, with regard to P _{i}, total K object search (E _{k}, θ _{k}), shown in accompanying drawing 4 left side forms.
(3) confirm search strategy, the present invention is at first to (E _{k}, θ _{k}) utilize the binary search corner sign indicating number (E that angle is identical in the seek error allowed band in the search volume _{n}, θ _{n}), establishing characteristic of correspondence point is Q _{j}If find, the present invention also can be in finding corner other identical (E of linear search angle in the error allowed band before and after the sign indicating number _{n}, θ _{n}).If (E _{k}, θ _{k}) and (E _{n}, θ _{n}) comparative result do not need add operation, then respectively with (E _{k}, θ _{k}) and (E _{n}, θ _{n}) be the ring type corner sign indicating number L of two equivalences of dot generation _{P}={ A _{0}, A _{1}..., A _{(K1)}And L _{Q}={ B _{0}, B _{1}..., B _{(K1)}, A wherein _{k}, B _{n}Expression corner sign indicating number, note E _{k} ^{A}, θ _{k} ^{A}, E _{n} ^{B}, θ _{n} ^{B}Represent A respectively _{k}, B _{n}Bian Hejiao.Press their maximal phase of increment matching algorithm calculating then like ring type corner sign indicating number.The increment matching algorithm is following: at first to A _{0}, B _{0}Compare operation, if θ _{0} ^{A}, θ _{0} ^{B}Unequal, the next corner sign indicating number that the corner sign indicating number that then angle is little is adjacent carries out add operation and generates new corner sign indicating number; If θ _{0} ^{A}, θ _{0} ^{B}Equate, then verify θ _{0} ^{A}, θ _{0} ^{B}Adjacent corresponding sides (E _{0} ^{A}, E _{0} ^{B}) and (E _{1} ^{A}, E _{1} ^{B}) whether proportional, if proportional, then with A _{0}, B _{0}Be retained in respectively in the similar ring type corner sign indicating number of having tried to achieve (just having begun to be sky), if disproportionate, then with this A _{0}, B _{0}Carry out add operation with adjacent separately next corner sign indicating number and generate new corner sign indicating number, next compare operation is carried out in circulation then, and cyclic process is till one of them ring type corner sign indicating number has traveled through.False code is represented as follows:
Suppose t _{1}And t _{2}Be two loop variables, t _{1}=0, t _{2}=0.Make temp _{A}=A _{k}, temp _{B}=B _{n}, rate=E _{n} ^{B}/ E _{k} ^{A}, then the increment matching algorithm is following:
Only keep the maximum similar ring type corner sign indicating number of similar length in the search procedure,, all keep when the maximum ring type corner sign indicating number of similar length has when a plurality of.With maximal phase like the similar length of ring type corner sign indicating number as the similarity between two unique points, notice that the similarity of all two unique points not calculated in the search procedure is 0.
(4) the definite unique point that possibly mate, the present invention defines S (A _{k}, B _{n})={ L _{k}, L _{n}Represent with E _{k} ^{A}, E _{n} ^{B}Be the similar ring type corner sign indicating number of initial corresponding sides, L _{k}, L _{n}Represent similar ring type corner sign indicating number.Make S (A _{k}, B _{n}) similar length be len ^{Ij}(k, n), unique point P then _{i}, Q _{j}Similarity be:
If Q _{j} ^{*}Satisfy
The present invention just thinks Q _{j} ^{*}Be P _{i}Possible match point, P in the computation process _{i}Have a plurality of Q that possibly mate _{j} ^{*}, all keep.
Step 3 structure matching of the present invention is specially:
If P _{i}The Q that possibly mate _{j} ^{*}There is NO, then to possibility match point (P _{i}, Q ^{*} _{j}) be classified as follows:
Class condition is explained as follows: (the P among the badClass _{i}, Q ^{*} _{j}) similarity all less than 3, when K is bigger,, (P is described like K=15 _{i}, Q ^{*} _{j}) the local space textural difference very big, P _{i}Or Q ^{*} _{j}Possibly be noise spot, the possibility of coupling is very little; (P among the unknownClass _{i}, Q ^{*} _{j}) similarity all more than or equal to 3, when K is bigger,, (P is described like K=15 _{i}, Q ^{*} _{j}) the local space texture ratio more similar, but P _{i}The Q that possibly mate ^{*} _{j}More than one, this explanation only can't be confirmed P through the similarity of local space structure _{i}With which Q ^{*} _{j}Coupling needs subsequent authentication to confirm; (P among the goodClass _{i}, Q ^{*} _{j}) similarity all more than or equal to 3, when K is bigger,, (P is described like K=15 _{i}, Q ^{*} _{j}) the local space texture ratio more similar, and P _{i}The Q that possibly mate ^{*} _{j}Has only one, (P _{i}, Q ^{*} _{j}) possibility of correct match is very big, and similarity is big more, explains that the local space structure is similar more, the possibility of coupling is big more.
According to abovementioned analysis, the present invention is with (the P among the goodClass _{i}, Q ^{*} _{j}) as the result of structure matching, the present invention simultaneously is also to the (P among the goodClass _{i}, Q ^{*} _{j}) carry out the local similar conversion and estimate,, note in local similar conversion estimation procedure (P _{i}, Q ^{*} _{j}) the one group respective neighbours point related with them make up and then estimate one group of local similar conversion, if (P _{i}, Q ^{*} _{j}) similarity be ls, then can estimate the conversion of ls group local similar, matching algorithm is got the average of ls group local similar conversion as the final local similar conversion of estimating.
Described step 4, optimization of matching is specially:
(1) to (P among the goodClass _{i}, Q ^{*} _{j}) sort by similarity is descending;
(2) to arranging (the P of preface _{i}, Q ^{*} _{j}) scan, if (P _{i}, Q ^{*} _{j}) not by the cluster mistake, then with (P _{i}, Q ^{*} _{j}) type of being center C _{i}, (P _{i}, Q ^{*} _{j}) related local map T _{i}As sorter, do not verified by the possible match point of cluster and related respective neighbours point among goodClass and the unknownClass.Cluster process is following: at first with (P _{i}, Q ^{*} _{j}) related respective neighbours point is included into such, because sorter T _{i}Come out by these respective neighbours point estimation; Utilize among other goodClass of sorter checking and the unknownClass the not possible match point of cluster and related respective neighbours point then; If the point of being verified is to satisfying sorter; Then it is carried out mark, no longer carry out subsequent treatment and it is included into such, otherwise it is ignored.If a pair of point that wherein need verify is for (P Q), establishes P through T _{i}Point after the mapping is T _{i}(P), if T _{i}(P) can accept in error range with the Euclidean distance of Q, the present invention thinks that then (P Q) meets T _{i}, with (P Q) is classified as C _{i}Be formulated as:
T _{i}(P)，Q＜ε _{T}
ε _{T}Be the error control factor.Attention is every couple of (P in cluster process _{i}, Q ^{*} _{j}) only by cluster once, i.e. (P _{i}, Q ^{*} _{j}) belong to and can only belong to a class.Cluster process must to all not cluster cross other maybe match points and related respective neighbours point verify.The beginning condition of cluster is also to exist among the goodClass not have cluster to cross (P next time _{i}, Q ^{*} _{j}).
(3) cluster result is analyzed,, must it be deleted because generally there is erroneous matching among the structure matching result.In cluster result, the class that the definition element is maximum is a maximum kind, and the class of element more than second is inferior big type, a pair of match point of each element representation in the class.Generally, the unique point of correct match related local map also be correct, all correct match points all can gather in the error allowed band into one type, the point of respective class is very many to meeting; Otherwise, the point of erroneous matching related local map randomness very big, element is considerably less in corresponding type.Therefore the present invention is optimized coupling as follows: if the ratio of maximum kind and the element number of inferior big class is greater than certain threshold value; Perhaps the maximum kind element number is greater than inferior big type; The maximum kind element number is greater than certain threshold value simultaneously, and the present invention will export maximum kind as corresponding relation, carries out optimal transformation and estimates; Otherwise the present invention thinks that then two point set correlativitys of coupling are too poor, can't mate.
Description of drawings
Fig. 1: process flow diagram of the present invention.
Fig. 2: unique point P _{c}KNN space structure figure and corresponding ring type corner sign indicating number.
Fig. 3: similar ring type corner sign indicating number and corresponding similar space structure.
Fig. 4: similarity is calculated synoptic diagram.
Fig. 5: two images to be matched.
Fig. 6: the matching result of two width of cloth images shown in Figure 5.
Embodiment
Below provide instantiation to further specify application of the present invention.
Embodiment 1
Match objects is seen two width of cloth images shown in the accompanying drawing 5.Two width of cloth image takings are in different time, and different angles have only partial content identical, and the image size is 440 * 330.
Moving equipment of the present invention is Gateway T6307c notebook, Intel 2core 1.6G, 1G internal memory.The present invention has platform transplantation property, and concrete implementation platform is GCC environment and the Visual C++2005 environment under the Windows XP/Server 2003 under the Centos.Operational efficiency GCC is better than Visual C++2005.
The present invention stipulates the Euclidean distance error ε between equal error ε=0.1, two pixel of two scalars _{T}=10, define two kinds of errors and be intended to calculate similar ring type corner sign indicating number and cluster.Neighbours' number of the unique point of the present invention simultaneously is set to K=15.
The practical implementation situation is following:
Step 1: from two images to be matched, obtain two feature point sets as input, unique point only need comprise that twodimensional coordinate information gets final product.The present invention has adopted single scale Harris Corner Detection Algorithm, Harris response coefficient α=0.06, difference yardstick σ _{D}=3, level and smooth yardstick σ _{I}=7, it is 3 * 3 that local nonmaximum value suppresses window, and inhibiting value is 0.01 times of overall maximum angle point response.Two width of cloth images extract 189 and 208 unique points respectively.Corresponding point set is designated as S and T, n respectively _{1}=189 and n _{2}=208.
Step 2: similarity is calculated, and 15 neighbours that extract each unique point are according to the ring type corner sign indicating number of the definition calculated characteristics point of ring type corner sign indicating number.In computation process, avoid removing Zoperation.When scalar is done division, as required, add or deduct a very little positive number, in order to avoid program crashing.
Based on the definition of similar ring type corner sign indicating number, calculate the similarity of two characteristic points to be matched, based on classifying rules, confirm the element among goodClass and the unknownClass, calculate simultaneously element among the goodClass related local map.Concrete steps are following:
(1) sets up the search volume.208 * 15 corner sign indicating number (E to the ring type corner sign indicating number of 208 unique points among the T _{n}, θ _{n}) sort to set up index according to angular dimension.
(2) confirm object search.For obtaining P _{i}The point Q that possibly mate ^{*} _{j}, need be to P _{i}Ring type corner sign indicating number on 15 corner sign indicating number (E _{k}, θ _{k}) (k=0,1 ..., 15) search for respectively.
(3) confirm search strategy.Select each (E successively _{k}, θ _{k}), search for by angle in the search volume, if search for successfully, the corner sign indicating number (E that is then searching _{n}, θ _{n}) front and back continuation linear search other corners sign indicating number (E that angle equates in error range _{n}, θ _{n}), if (E _{k}, θ _{k}) and (E _{n}, θ _{n}) comparative result do not need add operation, then respectively with (E _{k}, θ _{k}) and (E _{n}, θ _{n}) be the ring type corner sign indicating number of two equivalences of dot generation and utilize maximal phase that the increment matching algorithm calculates them like ring type corner sign indicating number.
(4) the definite unique point that possibly mate, P _{i}15 (E _{k}, θ _{k}) after all search finishes, keep the maximum similar ring type corner sign indicating number of similar length in the computation process, with the maximal phase that keeps like related another unique point Q of ring type corner sign indicating number _{j}As P _{i}The point Q that possibly mate ^{*} _{j}, then all preserve if any a plurality of.
Step 3: structure matching, 189 unique points among the S are classified according to similarity and the unique point number that possibly mate.Keep goodClass and unknownClass.And according to the local map of elements correlation among the similarity transformation Model Calculation goodClass.The similarity transformation model is following:
X＝sxcosθsysinθ+t _{x}
Y＝sxsinθ+sycosθ+t _{y}
S wherein, θ is respectively the zoom factor and the anglec of rotation, t _{x}, t _{y}It is respectively the translational movement of x and y direction.(x is y) with (X Y) is the unique point coordinate of conversion front and back respectively.Note in the local similar transformation calculations process, just the possible match point one group respective neighbours point related with them can be estimated one group of conversion coefficient, if so P _{i}, Q ^{*} _{j}Similarity is ls, then can estimate the conversion of ls group local similar, and the present invention organizes ls the average of local similar conversion as the final local similar conversion of estimating.
Step 4: optimization of matching, sort by similar length to the element among the goodClass in the step 3 from big to small.Because similar length is big more, the local space structure of two unique points of expression is similar more, and the possibility of coupling is also big more, from the maximum P of similar length _{i}, Q ^{*} _{j}The beginning cluster just can guarantee that also the cluster effect is good more, and the matching effect that obtains is also good more.So the maximum a pair of unique point of similarity that cluster is at first selected not gather class begins.Carry out cluster by clustering rule then.To the splicing and the recognition application of large scale scene image, establish s _{1}, s _{2}Be maximum kind and time element number of big class, if s _{1}>=2s _{2}Perhaps s _{1}＞s _{2}, s _{1}＞23, algorithm will be used for optimal transformation and estimate the output of the element in the maximum kind as the optimization of matching result.Otherwise algorithm will think that coupling is unsuccessful.To the instance of accompanying drawing 5, cluster result one has 88 types, and wherein the maximum kind number of elements is 33, and inferior big dvielement number is 6.Through observing, the maximum kind element all is correct match in error range, and inferior big dvielement all is erroneous matching in error range.In accompanying drawing 5 example illustrated, maximum kind is 5.5 with inferior big dvielement number ratio, satisfies condition, and the match is successful.The element of maximum kind will carry out optimal transformation and estimate as corresponding relation.
Step 5: conversion estimates that because generally, the conversion between the scene image often has affine or view transformation to a certain degree, so the present invention adopts the affined transformation model to estimate the conversion between the scene image.The affined transformation model is following:
X＝ax+by+t _{x}
Y＝cx+dy+t _{y}
A wherein, b, c, d are respectively rotations, convergentdivergent and stretching factor, t
_{x}, t
_{y}It is respectively the translational movement of x and y direction.(x is y) with (X Y) is the unique point coordinate of conversion front and back respectively.To all match points,, estimate to obtain optimal transformation thereby carry out least square fitting by the linear system of equations of affined transformation model group.To the instance of accompanying drawing 5, optimal transformation does
Claims (2)
Priority Applications (1)
Application Number  Priority Date  Filing Date  Title 

CN2009100524538A CN101567051B (en)  20090603  20090603  Image matching method based on characteristic points 
Applications Claiming Priority (1)
Application Number  Priority Date  Filing Date  Title 

CN2009100524538A CN101567051B (en)  20090603  20090603  Image matching method based on characteristic points 
Publications (2)
Publication Number  Publication Date 

CN101567051A CN101567051A (en)  20091028 
CN101567051B true CN101567051B (en)  20120822 
Family
ID=41283196
Family Applications (1)
Application Number  Title  Priority Date  Filing Date 

CN2009100524538A CN101567051B (en)  20090603  20090603  Image matching method based on characteristic points 
Country Status (1)
Country  Link 

CN (1)  CN101567051B (en) 
Cited By (1)
Publication number  Priority date  Publication date  Assignee  Title 

US8965134B2 (en)  20110405  20150224  HewlettPackard Development Company, L.P.  Document registration 
Families Citing this family (22)
Publication number  Priority date  Publication date  Assignee  Title 

CN101916445A (en) *  20100825  20101215  天津大学  Affine parameter estimationbased image registration method 
CN102088569B (en) *  20101013  20130619  首都师范大学  Sequence image splicing method and system of lowaltitude unmanned vehicle 
CN102004786B (en) *  20101202  20121128  上海交通大学  Acceleration method in image retrieval system 
CN102567724B (en) *  20101211  20160622  罗普特（厦门）科技集团有限公司  Image correction system and method 
GB2487377B (en) *  20110118  20180214  Aptina Imaging Corp  Matching interest points 
CN102184560B (en) *  20110325  20130130  南昌航空大学  Templatebased CCDDR (charge coupled devicedigital radiography) image splicing method 
CN102129477B (en) *  20110423  20130109  山东大学  Multimodecombined image reordering method 
CN103814384B (en) *  20110609  20170818  香港科技大学  Tracking based on image 
CN102231191B (en) *  20110717  20121226  西安电子科技大学  Multimodal image feature extraction and matching method based on ASIFT (affine scale invariant feature transform) 
KR20130015146A (en) *  20110802  20130213  삼성전자주식회사  Method and apparatus for processing medical image, robotic surgery system using image guidance 
CN102857704B (en) *  20120912  20150819  天津大学  With the multisource video joining method of timedomain synchronous calibration technology 
CN104268140B (en) *  20140731  20170623  浙江大学  Image search method based on weight self study hypergraph and multivariate information fusion 
CN104616300B (en) *  20150203  20170728  清华大学  The image matching method and device separated based on sampling configuration 
CN105472272A (en) *  20151125  20160406  浙江工业大学  Multichannel video splicing method based on FPGA and apparatus thereof 
CN105869145B (en) *  20160322  20181214  武汉工程大学  A kind of nuclear magnetic resonance image multistep method for registering accelerated based on kt 
CN106446923B (en) *  20160525  20190806  哈尔滨工程大学  Classification method of medical image based on corners Matching 
CN106530341B (en) *  20161101  20191231  成都理工大学  Point registration algorithm for keeping local topology invariance 
CN106548493A (en) *  20161103  20170329  亮风台（上海）信息科技有限公司  A kind of method and system of figure matching 
CN108458655A (en) *  20170222  20180828  上海理工大学  Support the data configurableization monitoring system and method for vision measurement 
CN107240127A (en) *  20170419  20171010  中国航空无线电电子研究所  The image registration appraisal procedure of distinguished point based mapping 
CN108427927B (en) *  20180316  20201127  深圳市商汤科技有限公司  Object rerecognition method and apparatus, electronic device, program, and storage medium 
CN110162656A (en) *  20190505  20190823  南京师范大学  A kind of method and system enhancing image characteristic point information 
Citations (2)
Publication number  Priority date  Publication date  Assignee  Title 

CN101140624A (en) *  20071018  20080312  清华大学  Image matching method 
CN101350101A (en) *  20080909  20090121  北京航空航天大学  Method for autoregistration of multiamplitude deepness image 

2009
 20090603 CN CN2009100524538A patent/CN101567051B/en active IP Right Grant
Patent Citations (2)
Publication number  Priority date  Publication date  Assignee  Title 

CN101140624A (en) *  20071018  20080312  清华大学  Image matching method 
CN101350101A (en) *  20080909  20090121  北京航空航天大学  Method for autoregistration of multiamplitude deepness image 
Cited By (1)
Publication number  Priority date  Publication date  Assignee  Title 

US8965134B2 (en)  20110405  20150224  HewlettPackard Development Company, L.P.  Document registration 
Also Published As
Publication number  Publication date 

CN101567051A (en)  20091028 
Similar Documents
Publication  Publication Date  Title 

Wang et al.  Dynamic graph cnn for learning on point clouds  
Zhu et al.  Deep learning in remote sensing: A comprehensive review and list of resources  
Srivastava et al.  Functional and shape data analysis  
Zhou et al.  Flowweb: Joint image set alignment by weaving consistent, pixelwise correspondences  
Zamir et al.  Image geolocalization based on multiplenearest neighbor feature matching usinggeneralized graphs  
EP3074918B1 (en)  Method and system for face image recognition  
Zhou et al.  Deformable graph matching  
Ma et al.  Robust $ L_ {2} E $ estimation of transformation for nonrigid registration  
Zia et al.  Detailed 3d representations for object recognition and modeling  
Cho et al.  Finding matches in a haystack: A maxpooling strategy for graph matching in the presence of outliers  
Ma et al.  Nonrigid point set registration with robust transformation estimation under manifold regularization  
Zhou et al.  Factorized graph matching  
Sarlin et al.  Superglue: Learning feature matching with graph neural networks  
Barinova et al.  On detection of multiple object instances using hough transforms  
US8798357B2 (en)  Imagebased localization  
Wang et al.  Spectral clustering on multiple manifolds  
Zhou et al.  Multiimage matching via fast alternating minimization  
Mateus et al.  Articulated shape matching using laplacian eigenfunctions and unsupervised point registration  
Trivedi et al.  Lowlevel segmentation of aerial images with fuzzy clustering  
Zhang et al.  Image categorization by learning a propagated graphlet path  
Torresani et al.  A dual decomposition approach to feature correspondence  
Rodolà et al.  A scale independent selection process for 3d object recognition in cluttered scenes  
EP2720171B1 (en)  Recognition and pose determination of 3D objects in multimodal scenes  
Schmid et al.  Local grayvalue invariants for image retrieval  
Liu et al.  Robust subspace segmentation by lowrank representation. 
Legal Events
Date  Code  Title  Description 

PB01  Publication  
C06  Publication  
SE01  Entry into force of request for substantive examination  
C10  Entry into substantive examination  
GR01  Patent grant  
C14  Grant of patent or utility model 