CN102200999B - Method for retrieving similarity shape - Google Patents

Method for retrieving similarity shape Download PDF

Info

Publication number
CN102200999B
CN102200999B CN201110106315A CN201110106315A CN102200999B CN 102200999 B CN102200999 B CN 102200999B CN 201110106315 A CN201110106315 A CN 201110106315A CN 201110106315 A CN201110106315 A CN 201110106315A CN 102200999 B CN102200999 B CN 102200999B
Authority
CN
China
Prior art keywords
shape
distance
retrieved
matrix
shapes
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
CN201110106315A
Other languages
Chinese (zh)
Other versions
CN102200999A (en
Inventor
白翔
周瑜
刘文予
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.)
Huazhong University of Science and Technology
Original Assignee
Huazhong University of Science and Technology
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 Huazhong University of Science and Technology filed Critical Huazhong University of Science and Technology
Priority to CN201110106315A priority Critical patent/CN102200999B/en
Publication of CN102200999A publication Critical patent/CN102200999A/en
Application granted granted Critical
Publication of CN102200999B publication Critical patent/CN102200999B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Abstract

The invention discloses a method for retrieving a similarity shape. The method comprises the following steps: (1) extracting shape outlines of an input retrieving image and an image to be retrieved in a database; (2) describing all shapes (including input retrieving shapes and shapes to be retrieved in the database) by using internal distance shape context descriptors; (3) matching all shapes (including the input retrieving shapes and shapes to be retrieved in the database) pairwise by using a dynamic planning method; (4) calculating by using a context sensitive similarity measurement method to acquire a new similarity measurement sequence; and (5) acquiring a result of shape retrieving. In the method, dissimilarity (distance) among pairwise shapes is not used as direct basis for shape retrieving, but the dissimilarity (distance) among initial pairwise shapes is improved by integrating internal differences of shapes and utilizing structure information in a shape similarity space, so that the accuracy of shape retrieving is effectively promoted.

Description

A kind of method of retrieving the similarity shape
Technical field
The present invention relates to shape is retrieved, be specifically related to a kind of method of retrieving the similarity shape.
Background technology
Form fit/retrieval is a very important problem in the computer vision.The existing different shape matching method of many kinds, and on the raising matching precision, obtained certain progress.Yet nearly all these methods all emphasis are absorbed in the similarity measurement between the shape in twos, attempt through more rational shape description of proposition and matching algorithm, thereby better measure two similarity or distances between the shape.These methods are all based on a typical viewpoint, and two shapes are similar more, and the difference between them is just more little, and this difference is measured by distance function usually.Yet this viewpoint has been ignored a fact: for shape, the difference between some similar shape maybe be very big, and the difference between some inhomogeneity shape maybe be less relatively on the contrary.This phenomenon causes by the complicacy of shape, because blocking between similar shape, distortion or nonrigid deformation etc., similar shape has very big difference.In other words, for the very big similar shape of difference,, all can not well measure with regard to two very big shapes and compare, make them very similar original difference in any case improve shape description or matching algorithm.Though this species diversity possibly not be the essential distinction between the shape, be difficult to eliminate.Address this problem, need improve the thinking of at present merely carrying out similarity measurement from the characteristic of two shapes itself.
" interior distance is hereinafter in shape " is a kind of effectively shape description that proposes in recent years.It has stronger stability for the non-rigid variation of similar shape.In system of the present invention, shape description uses interior distance hereinafter in shape.In apart from the contextual particular content of shape, show, be published in 2007 periodical " IEEE Transactions onPattern Analysis and Machine Intelligence " the 26th at HaibinLing and David W.Jacobs and roll up in the article " ShapeClassification Using the Inner-Distance " on the o. 11th detailed introduction is arranged.
Dynamic programming method is a branch of operational research, is a kind of mathematical method that solves multistage decision process.In recent years, dynamic programming method has been obtained widely in form fit, shape retrieval and has been used.In the present invention, shape is mated the use dynamic programming method.The content of in form fit, using about dynamic programming method, showing, be published in 2000 in the article " Shape Retrieval Based on Dynamic Programming " of periodical " IEEE Transactions on Image Processing " on the 9th the 1st phase of volume at Evangelos Milios and Euripides G.M.Petrakis has detailed introduction.
Summary of the invention
The object of the present invention is to provide a kind of method of retrieving the similarity shape, this method can improve the accuracy rate of shape retrieval.
A kind of method of retrieving the similarity shape provided by the invention comprises the steps:
(1) the shape profile of image to be retrieved in extraction input inquiry image and the database, query image is referred to as to inquire about shape, and image to be retrieved is referred to as shape to be retrieved;
(2) on the basis of inquiry shape that step (1) is extracted and shape to be retrieved, calculate the characteristic of each shape profile, just descriptor;
(3) on the basis of the shape facility that step (2) calculates; Shape for the inquiry shape and the shape to be retrieved in the database of input are formed is gathered; Mate between any two shapes in the pair set; Obtain its measure of dissimilarity value between any two, form a measure of dissimilarity matrix according to these measure of dissimilarity values;
The measure of dissimilarity matrix of (4) trying to achieve according to step (3) calculates the inquiry shape with the similarity between any shape to be retrieved in the database;
(5) follow the similarity between all shapes to be retrieved in the database based on the input inquiry shape of trying to achieve in the step (4), deterministic retrieval output result.
The present invention possesses following character: each mode to the similarity of shape is only calculated before having replaced in (1), has utilized the stream shape of being made up of known all shapes more; (2) there are not dominance ground study stream shape or geodesic line, bigger because this calculating expends.A better similarity is transmitted similarity through inquiry shape shape to be retrieved in database and is obtained.To be further described effect of the present invention in the embodiment part.
Description of drawings
Fig. 1 is the test pattern example that the present invention imports;
Fig. 2 is a shape profile sample point and interior apart from synoptic diagram, (A) configuration sampling point synoptic diagram, and (B) shape is interior apart from synoptic diagram, and (C) local shape is interior apart from synoptic diagram;
Fig. 3 is interior angle synoptic diagram;
Fig. 4 is the concentric circles of band angular divisions and the application synoptic diagram on contour representation thereof;
Fig. 5 is apart from synoptic diagram in the concentric circles of band angular divisions reaches;
Fig. 6 is a counting statistics histogram synoptic diagram;
Fig. 7 is configuration sampling point and statistic histogram synoptic diagram thereof;
Fig. 8 is the system flowchart of embodiment of the present invention method;
Fig. 9 is the comparison of recall ratio on the MPEG-7 database;
Figure 10 is the result's that on the MPEG-7 database, retrieves comparison;
Figure 11 is that the result of the local shape retrieval of losing of band compares.
Embodiment
The present invention relates to a kind of search method and system of similar shape.A shape database is arranged in the system, comprise several shapes in this database, these shapes are to wait for the shape (being designated hereinafter simply as " shape to be retrieved ") of the output that is retrieved.In addition, as the input of system, an inquiry shape is arranged, the shape to be retrieved in the database can be respectively be carried out similarity relatively with this input inquiry shape, thereby confirms that which shape to be retrieved is enough similar with the input inquiry shape and export as result for retrieval.
Retrieval work receives the input inquiry shape from system and begins, and returns the SOME RESULTS shape that draws through retrieval with system and finishes.
A given shape database, an inquiry shape and a shape distance function; It needs not to be a tolerance; The present invention learns a distance function, and this distance function obtains through the shortest path on the stream shape of the similarity formation between any two of the shape to be retrieved in inquiry shape and the database.The present invention does not need this stream shape of dominance ground study.Experiment showed, that the distance function of new study can integrate the knowledge of inner shape difference.This learning process is non-supervision.
In the method that the present invention proposes, on shape similarity basis, further calculate new similarity based on existing shape description and the acquisition of form fit algorithm computation.Say intuitively, for a given shape S 1Even, shape S 3With shape S 1Dissmilarity, but if with shape S 3Shapes similar S 2Similar in appearance to S 1, shape S so 1With shape S 3Similarity also can be very high.Yet, even shape S 1With shape S 3Very similar, if but with shape S 3Shapes similar S 2With shape S 1And dissmilarity, shape S so 1With shape S 3Similarity will be very low.Therefore, the new similarity that obtains is exactly and its shapes similar for the content of shaped here to content erotic.
Even shape S 1With shape S 3Differ greatly, but a shape S is arranged 2All very little with their difference, still think shape S 1With shape S 3Similar.This situation is possible for most shapes distance, because they do not satisfy the relation of triangle inequality, i.e. and d (S 1, S 3)≤d (S 1, S 2)+d (S 2, S 3) be not to set up always.If to some shape S 1, S 2, S 3,, satisfy d (S 1, S 3)＞d (S 1, S 2)+d (S 2, S 3) this situation, the method for this chapter proposition can be acquired one newly apart from d ' (S so 1, S 3) satisfy d ' (S 1, S 3)≤d (S 1, S 2)+d (S 2, S 3).Equally, if in metric space, exist a paths to make d (S 1, S 3)＞d (S 1, S 2,1)+...+d (S 2, u, S 3), wherein u is the quantity of shape on this path, so this method obtain one new for d ' (S 1, S 3)≤d (S 1, S 2,1)+...+d (S 2, u, S 3).Because this paths is represented one from shape S 1To shape S 3The process of minimal distortion distortion, can ignore the shape difference that is not very crucial, and concentrate on the crucial shape difference, so just obtained one new for d '.The present invention has promoted the accuracy of shape retrieval through this new distance metric.
Below in conjunction with accompanying drawing and instance the inventive method is done further detailed explanation.
As shown in Figure 8, the inventive method comprises the steps:
(1) profile of image to be retrieved in extraction input inquiry image and the database.The query image of input and the image to be retrieved in the database all are bianry images.If the query image of input is not a bianry image, convert thereof into bianry image earlier.The method of conversion comprises artificial manual mark or image partition method (for example, figure normalization cutting scheduling algorithm) etc.As shown in Figure 1, the shape shown in the figure is an input picture of the present invention.From the image to be retrieved of query image and the database of input, obtain the shape profile through the profile extraction algorithm.
The shape profile that extracts from the query image of input is called the inquiry shape, and the shape profile that image to be retrieved extracted from database is called shape to be retrieved.Because the shape profile has comprised the key message of bianry image; Profile extracts and is very easy to realize simultaneously; So the query image of input also can be described as the input inquiry shape, the binary image data storehouse also can be described as shape database, and the image to be retrieved in the database also can be described as shape to be retrieved.The back is adopted the saying of " inquiry shape ", " shape database " and " shape to be retrieved " respectively in the literary composition.
For the image to be retrieved in the database, obtain the work of treatment of corresponding shape profile from bianry image, not be used in retrieval phase and carry out, can accomplish in advance.For the query image of input, its work of extracting the shape profile can only be accomplished in retrieval phase.
(2) on the basis of inquiry shape that step (1) is extracted and shape to be retrieved, calculate the characteristic of each shape profile, just descriptor.The present invention use in the distance in shape hereinafter as the descriptor of shape.Specific practice is following:
(2.1) each shape profile is carried out uniform sampling.N point of each profile up-sampling, the value of N are set in different application environments as required, like N=100.See shown in Fig. 2 (A) that the circular solids point on the outline of scissors is the sampled point on the profile in the figure.
(2.2) for each the sampled point p on each shape profile i, i representes the sequence number of sampled point, i=1 ..., N, the interior distance of use hereinafter descriptor is in shape described.For certain shape, all N sampled point on its shape profile interior apart from shape description formed the characteristic of this shape jointly.
The interior distance of calculating each sampled point is the hereinafter descriptor in shape, and specific practice is following:
(2.2.1) calculate interior distance.Interior distance between the definition sampled point is in the inner length that connects the shortest path of two sampled points of shape.Use shortest path first to calculate certain sampled point p iInterior distance on the shape border between other sampled point, computing method are following: 1. at first construct a graph structure, in this graph structure, each sampled point p iAs the summit of figure, establishing wherein any two sampled points is p 1And p 2, judge whether the limit that connects these two summits drops on the inside of shape, if drop on the inside of shape, so just between these two summits, keep the connection on a limit, the limit weight equals the Euclidean distance between them || p 1-p 2||.If the inside of shape is not dropped on this limit fully, then there is not a limit directly to link to each other between these two summits.2. on whole weight map, use shortest path first (like enlightening Coase thorough (Dijkstra) algorithm), obtain the shortest connection the between any two points, this bee-line is interior distance.It in Fig. 2 (C) the shortest path line segment on the shape local detail.
(2.2.2) calculate interior angle.Define a sampled point p 1With respect to another sampled point p 2Interior angle be configuration sampling point p 1Tangential direction and from a p 1The p that sets out 1, p 2Between the shortest path direction between angle.In Fig. 3, p 1And p 2Be two sampled points on the profile, the angle θ shown in the figure is p 1And p 2Between interior angle.
(2.2.3) calculate interior distance hereinafter descriptor in shape.The concentric circles that uses the division of band direction among the present invention is as the instrument that calculates descriptor, and so-called band direction is divided and is meant with sampled point p iTangential direction be the zero angle direction, with 360 degree spaces evenly be divided into G interval, as shown in Figure 4, the concentrically ringed center of circle is in certain configuration sampling point p i, each radius of a circle of concentric circles and concentrically ringed orientation angle are divided according to the concrete occasion of using and are decided, and suppose that the zero degree direction is sampled point p iTangential direction, suppose angular divisions be G interval, use in this example be with 360 degree be divided into G=12 interval, each zone is 30 to spend.Hypothesis has R concentric circles again, R=5 in this example, and then R concentric circles and G angular interval combine and have constituted (R+1) * G interval.The interior distance and the interior angle that combine (2.2.1) and (2.2.2) acquisition then, the point of confirming on the profile other is with a p iFor the band direction in the center of circle apart from the distribution of dividing in concentrically ringed (R+1) * G interval, obtain a statistic histogram, this histogram is seen shown in Fig. 7 right side.This statistic histogram is interior distance hereinafter descriptor in shape.Its mathematical definition is:
h i(k)＝#{p j：1≤j≤N，j≠i，p j-p i∈bin(k)}
In following formula, h iThe interior distance hereinafter statistic histogram in shape of i sampled point on the expression profile, h i(k) expression drops on the number of putting in k the interval of this statistic histogram, and the span of k is from 1 integer to (R+1) * G.p iRepresent i sampled point, the object described of this statistic histogram just, p jThen represent in N the sampled point on the shape profile, except p iIn addition other N-1 the point in any one.p iBe the above-mentioned concentrically ringed center of circle, p j-p iRepresent a vector, this vector points to p jBin representes the interval division that the above-mentioned concentric circles with being with direction apart from division obtains, and k of bin (k) expression is interval.The implication of expression formula is if vectorial p j-p iDrop among the interval bin (k), then think a p jDrop among the interval k of statistic histogram, the quantity h of the point among the interval k of corresponding statistic histogram i(k) add 1.Symbol # representes the implication that adds up.
Explain further that in conjunction with Fig. 4, Fig. 5 in Fig. 4, the position of establishing place, the center of circle is p i, the position of arrow points is p j, through an intermediate point p wLength from the broken line in the center of circle among Fig. 5 is interior distance; In calculating during apart from hereinafter descriptor in shape; At first the broken line among Fig. 5 is done partial rotation and obtain the straight-line segment shown in Fig. 6, see then which distance regions this straight-line segment terminal point drops between, judge this straight-line segment and some p iThe number of degrees of angle (being among Fig. 6) of tangential direction near angle θ that home position indicated drop in which angular interval, in conjunction with above-mentioned two parts information, finally confirm a some p jDrop in which interval, accumulated value that then should the interval before the basis on add 1.
The final statistic histogram that obtains is as shown in Figure 7, and the dark more place of color is represented to count few more, and the shallow more place of color is represented to count many more.
In the statistic histogram of Fig. 7, X direction is represented the angular interval division, and y direction is represented apart from interval division.
(2.3) calculate the characteristic of each shape profile
Repeat the process of (2.2), calculate the interior distance hereinafter descriptor in shape that obtains each sampled point on certain shape profile, thereby obtain description this shape.This description to whole shape, be through one group of statistic histogram of describing like (2.2.3) composition, each statistic histogram all is the description of some points on this shape profile.Such one group of descriptor is exactly foregoing shape facility.
Repeat this process, calculate the characteristic that obtains all shapes to be retrieved in input inquiry shape and the database.
For the shape to be retrieved in the database, the calculating of its shape contour feature not be used in retrieval phase to be carried out, and can accomplish in advance.For the inquiry shape of input, the calculating of its shape contour feature can only be accomplished in retrieval phase.
(3) form fit
On the basis of the shape facility that step (2) calculates; Shape for the inquiry shape and the shape to be retrieved in the database of input are formed is gathered; Mate between any two shapes in the pair set, obtain its measure of dissimilarity value (distance) between any two.
Coupling between (3.1) two shapes
The present invention uses dynamic programming method, finds the solution the best correspondence of any two shape profile up-sampling points and the measure of dissimilarity value (distance) between this two shapes, accomplishes form fit.Specific practice is following:
Suppose that to be matched one of them is shaped as A, another one is shaped as B, forms two arrangement sets by the configuration sampling point of A and B, is designated as { p respectively i, i=1 ..., N}, { q j, j=1 ..., N}.
(3.1.1) the difference sequence of calculation { p i, i=1 ..., any some p among the N} iWith sequence { q j, j=1 ..., any some q among the N} jBetween characteristic distance c (i, j), this distance definition point p that (2.2) calculate that serves as reasons iStatistic histogram with the some q jStatistic histogram between distance, its mathematical definition is following:
c ( i , j ) = 1 2 Σ k = 1 ( R + 1 ) × G [ h A , i ( k ) - h B , j ( k ) ] 2 h A , i ( k ) + h B , j ( k )
H wherein A, i, h B, jBe the statistic histogram of definition in (2.2.3), subscript A and B represent h A, i(k), h B, j(k) be the statistic histogram of shape A and shape B respectively, c (i, j) expression point p iWith a q iCharacteristic distance.
Through calculating the characteristic distance of being had a few on the have a few and shape B on the shape A, the characteristic distance matrix of acquisition N * N c (i, j), i=1 ..., N, j=1 ..., N}:
C = c ( 1,1 ) c ( 1,2 ) . . . c ( 1 , N - 1 ) c ( 1 , N ) c ( 2,1 ) c ( 2,2 ) . . . c ( 2 , N - 1 ) c ( 2 , N ) . . . . . . . . . . . . . . . c ( N - 1,1 ) c ( N - 1,2 ) . . . c ( N - 1 , N - 1 ) c ( N - 1 , N ) c ( N , 1 ) c ( N , 2 ) . . . c ( N , N - 1 ) c ( N , N )
(3.1.2) two sequence { p of hypothesis i, i=1 ..., N} and { q j, j=1 ..., N} is at a p 1And q 1Place's alignment.Use dynamic programming method, try to achieve dissimilar degree (distance) D1 of two sequences, its solution procedure is following:
1. generate the empty matrix M of a N * N, be called the dynamic programming matrix.
2. initialization marginal condition:
M(0，0)＝0；
M(i，0)＝γ×i，i＝1，...，N；
M(0，j)＝γ×j，j＝1，...，N；
These numerical value are not included in the dynamic programming matrix M, because the scope of the line index of the element among the M and column index is from 1 to N integer.The scalar penalty factor of γ for setting.
3. according to from left to right, from top to bottom order, calculate successively each element in the dynamic programming matrix M value S (i, j), i=1 ..., N, j=1 ..., N.Computing formula is following:
M ( i , j ) = min M ( i - 1 , j - 1 ) + c ( i , j ) M ( i - 1 , j ) + γ M ( i , j - 1 ) + γ
Wherein (i j) is a p to c iWith a q jBetween characteristic distance.
4. the value of each element is all calculated and is finished in the dynamic programming matrix, and then dynamic programming method finishes.The dissimilar degree D of two sequences 1=M (N, N).
(3.1.4) process of repetition (3.1.3) is supposed two sequence { p respectively iAnd { q iAt p 2With q 1Place's alignment, p 3With q 1Place's alignment ..., p NWith q 1Place's alignment uses dynamic programming algorithm to try to achieve dissimilar degree (distance) D respectively 2, D 3..., D N
(3.1.5) choose minimum that in try to achieve in the step (3.1.3) to (3.1.4) N the distance, as shape A, the measure of dissimilarity value (distance) between the B:
D(A，B)＝min(D i)，i＝1，...，N
(3.2) coupling between the shape in twos in the shape set
For the shape set of forming by the inquiry shape and the shape to be retrieved in the database of input; Utilize the described method of step (3.1); Any two shapes in the pair set utilize dynamic programming algorithm to mate, and obtain its measure of dissimilarity value (distance).
Contain n shape to be retrieved in the tentation data storehouse, n is a positive integer, and the n of inquiry shape in a database shape of input has been formed n+1 shape { x altogether i, i=1 ..., n+1}.According to step (3.1), can obtain this n+1 shape measure of dissimilarity value (distance) between any two respectively, form a measure of dissimilarity matrix (distance matrix) { D I, j, i, j=1 ..., n, n+1}:
Wherein, D I, jRepresent the measure of dissimilarity value (distance) between i shape and j the shape.In addition, the inquiry shape with input is placed in this n+1 shape first, i.e. x 1
For the n in the database shape to be retrieved, its coupling and measure of dissimilarity between any two calculates, and not be used in retrieval phase and carries out, and can accomplish in advance.If but comprised the inquiry shape of input in two shapes of mating, then its coupling and measure of dissimilarity work could only be accomplished in retrieval phase.That is to say that for matrix D, the element value of its first row and first row must calculate in retrieval phase, but the equal calculated in advance of other element values is tried to achieve.
Measure of dissimilarity (distance) matrix of (4) trying to achieve according to step (3) calculates the inquiry shape with the similarity between any shape to be retrieved in the database.
This step just obtains result for retrieval through the shape method for measuring similarity that uses content erotic, and its specific practice is following:
(4.1) defining relation matrix w I, j, its computing formula is following:
w i , j = exp ( - D i , j 2 λ i , j 2 )
Wherein, D I, jBe the dissimilar metric matrix (distance matrix) that obtains in the step (3.2), λ I, jBe normalized parameter, the computing method of this normalized parameter are following:
λ i，j＝α·mean({knnd(x i)，knnd(x j)})
In following formula, mean ({ knnd (x i), knnd (x j)) expression shape x i, x jThe mean distance of preceding b nearest neighbor distance.So-called nearest neighbor distance is meant and middle b the minimum distance of the measure of dissimilarity value (distance) of this shape.Parameter b and α be rule of thumb decision in actual application environment, b=10 in this example, α=0.27.
(4.2) definition probability transfer matrix P, this probability transfer matrix passes through w I, jFollow direction normalization and obtain, computing formula is following:
P i , j = w i , j Σ h = 1 n + 1 w i , h
W wherein I, jBe above-mentioned relational matrix.
(4.3) define crucial function f, and find the solution crucial function through the process of following recurrence:
f t + 1 ( x i ) = Σ j = 1 n + 1 P i , j f t ( x j ) , i＝2，...，(n+1)
And have
f t+1(x 1)＝1
P wherein I, jBe above-mentioned probability transfer matrix; This recursive procedure need be passed through iteration T time, and parameter T is predetermined in actual application environment, T=5000 in this example.
(4.4) the crucial function f of using the recursive iteration process of step (4.3) to obtain, result's note is made f T(x i), i=1 ..., n, n+1, solving result are the matrixes of capable 1 row of n+1.Remove f T(x 1), in the matrix of capable 1 row of remaining n, i the shape in the capable value representation database of i and the similarity of input inquiry shape.
(5) follow the similarity f between the individual shape to be retrieved of n in the database based on the input inquiry shape of trying to achieve in (4) T(x i), i=2 ..., n, n+1, deterministic retrieval output result.
To f T(x i), i=2 ..., the n among the n+1 is capable, and value sorts from big to small, the order later according to ordering, the retrieval output of shape to be retrieved order in the specified data storehouse.For example, i value is maximum, the result for retrieval of output then, and i-1 shape comes first of result for retrieval, obtains second result for retrieval, the 3rd result for retrieval by that analogy, or the like.
What retrieval was exported is not the shape profile, but the original bianry image in the database.Like Figure 10, shown in Figure 11.
Set a parameter l, the quantity of the image result that the expression retrieval is returned.That is to say, retrieval is returned be in the database with preceding l maximum image of input inquiry image similarity.The value of parameter l is set based on the demand of practical application by the user flexibly, l=40 in this example.
The present invention of experimental result proof can improve the retrieval performance of existing shape matching method.
Provided the comparison that is published in the main method in international important periodical and the meeting in the present invention on the MPEG-7CE-Shape-1 database and over nearly 10 years in the table 1; The present invention has obtained 91.61% retrieval precision, is the highest among the result who on this database, obtains at present.
Provided the result of on Kimia99 data set the present invention and in recent years important method in the table 2 and compare, result's demonstration, the result of method of the present invention on each rank all is higher than additive method.
In embodiment (4.1), exist two parameter b and α, table 3 has provided the retrieval precision that system can reach under different parameters, can see, and under the situation of fine setting parameter, Supreme Procuratorate of the present invention rope precision can reach 92.57%.
What provide among Fig. 9 is apart from the comparison of the recall ratio (curve of " * " labelled notation) of the recall ratio of hereinafter method (curve of circles mark) and use the inventive method in shape on the MPEG-7 database, directly using.Can see that the method effect that the present invention provides obviously is superior to directly using interior distance hereinafter method in shape.
That Figure 10 provides is the result who on the MPEG-7 database, retrieves; First row are shapes to be retrieved; The right 10 row are the distance result for retrieval of hereinafter method (odd-numbered line) and result for retrieval of the inventive method (even number line) in shape in directly using, and have provided preceding 10 results the most close among the figure.Can see that the method effect that the present invention provides obviously is superior to directly using interior distance hereinafter method in shape.
What Figure 11 provided is the local shape retrieval experiment of losing of band.First classifies the image to be retrieved of having lost a part as; The right 10 row be in directly using distance in shape hereinafter method (odd-numbered line) retrieve and use the inventive method (even number line) to retrieve the result for retrieval that obtains, provided preceding 10 result for retrieval the most similar among the figure.
The recall ratio of table 1 distinct methods on the MPEG-7 data set (bull ' s eyes)
Figure BDA0000057703650000131
The result for retrieval of table 2Kimia 99 data sets
Figure BDA0000057703650000132
Bull ' the s eye recall ratio of table 3 under different parameters
b＝3 b＝5 b＝7 b＝9
α＝0.1 83.9％ 88.11％ 89.26％ 89.84％
α＝0.15 84.33％ 88.67％ 89.66％ 90.31％
α＝0.2 85.77％ 90.29％ 91.34％ 91.84％
α＝0.25 88.71％ 92.17％ 92.57％ 92.56％
α＝0.3 89.69％ 91.16％ 91.41％ 91.16％
α＝0.35 89.03％ 90.39％ 90.30％ 90.20％
α＝0.4 88.74％ 89.99％ 89.97％ 89.84％
The present invention not only is confined to above-mentioned embodiment; Persons skilled in the art are according to content disclosed by the invention; Can adopt other multiple embodiment embodiment of the present invention, therefore, every employing project organization of the present invention and thinking; Do some simple designs that change or change, all fall into the scope of the present invention's protection.

Claims (3)

1. a method of retrieving the similarity shape comprises the steps:
(1) the shape profile of image to be retrieved in extraction input inquiry image and the database, query image is referred to as to inquire about shape, and image to be retrieved is referred to as shape to be retrieved;
(2) on the basis of inquiry shape that step (1) is extracted and shape to be retrieved, calculate the characteristic of each shape profile, just descriptor;
(3) on the basis of the shape facility that step (2) calculates; Shape for the inquiry shape and the shape to be retrieved in the database of input are formed is gathered; Mate between any two shapes in the pair set; Obtain its measure of dissimilarity value between any two, form the measure of dissimilarity matrix according to the measure of dissimilarity value of obtaining;
The measure of dissimilarity matrix of (4) trying to achieve according to step (3) calculates the inquiry shape with the similarity between any shape to be retrieved in the database;
(5) follow the similarity between the individual shape to be retrieved of n in the database, deterministic retrieval output result based on the input inquiry shape of trying to achieve in (4);
Step (2) comprises following process:
(2.1) each shape profile is carried out uniform sampling;
(2.2) for each the sampled point p on each shape profile i, i representes the sequence number of sampled point, i=1 ..., N, the interior distance of use hereinafter descriptor is in shape described, and N representes the quantity of the sampled point on the shape profile;
(2.3) repeat the process of (2.2), calculate the interior distance hereinafter descriptor in shape that obtains each sampled point on certain shape profile, thereby obtain description this shape;
Step (3) specifically comprises following process:
Coupling between (3.1) two shapes:
Use dynamic programming method, find the solution the best correspondence of any two shape profile up-sampling points and the measure of dissimilarity value between this two shapes, accomplish form fit;
(3.2) coupling between the shape in twos in the shape set obtains dissimilar metric matrix { D I, j, i, j=1 ..., n, n+1}:
Shape set for being made up of the inquiry shape and the shape to be retrieved in the database of input utilizes the described method of step (3.1), and any two shapes in the pair set utilize dynamic programming algorithm to mate, and obtain its measure of dissimilarity value;
Contain n shape to be retrieved in the tentation data storehouse, n is a positive integer, and the n of inquiry shape in a database shape of input has been formed n+1 shape { x altogether i, i=1 ..., n+1}; According to step (3.1), obtain this n+1 shape measure of dissimilarity value between any two respectively, form a dissimilar metric matrix { D I, j, i, j=1 ..., n, n+1}:
D i , j = D ( 1,1 ) D ( 1,2 ) . . . D ( 1 , n ) D ( 1 , n + 1 ) D ( 2,1 ) D ( 2,2 ) . . . D ( 2 , n ) D ( 2 , n + 1 ) . . . . . . . . . . . . . . . D ( n , 1 ) D ( n , 2 ) . . . D ( n , n ) D ( n , n + 1 ) D ( n + 1,1 ) D ( n + 1,2 ) . . . D ( n + 1 , n ) D ( n + 1 , n + 1 )
Wherein, D I, jRepresent the measure of dissimilarity value between i shape and j the shape;
Step (4) specifically comprises following process:
(4.1) defining relation matrix w I, j, its computing formula is following:
w i , j = exp ( - D i , j 2 λ i , j 2 )
Wherein, D I, jBe the dissimilar metric matrix that obtains in the step (3.2), λ I, jBe normalized parameter, the computing method of this normalized parameter are following:
λ i,j＝α·mean({knnd(x i),knnd(x j)})
In following formula, mean ({ knnd (x i), knnd (x j)) expression shape x i, x jThe mean distance of preceding b nearest neighbor distance, so-called nearest neighbor distance is meant b minimum distance in the measure of dissimilarity value with this shape;
(4.2) definition probability transfer matrix P, this probability transfer matrix passes through w I, jFollow direction normalization and obtain, computing formula is following:
P i , j = w i , j Σ h = 1 n + 1 w i , h
(4.3) define crucial function f, and find the solution crucial function through the process of following recurrence:
f t + 1 ( x i ) = Σ j = 1 n + 1 P i , j f t ( x j ) , i = 2 , . . . , ( n + 1 )
And have
f t+1(x 1)＝1
P wherein I, jBe above-mentioned transition probability matrix; This recursive procedure need be passed through iteration T time, and parameter T is predefined iterations;
(4.4) use the recursive iteration process of step (4.3) to obtain crucial function f, result's note is made f T(x i), i=1 ..., n, n+1, solving result are the matrixes of capable 1 row of n+1; Remove f T(x 1), in the matrix of capable 1 row of remaining n, i the shape in the capable value representation database of i and the similarity of input inquiry shape.
2. the method for retrieval similarity shape according to claim 1 is characterized in that, step (2.2) comprises following process:
(2.2.1) calculate interior distance: the interior distance between the definition sampled point is in the inner length that connects the shortest path of two sampled points of shape, uses certain sampled point p of shortest path first calculating iInterior distance on the shape border between other sampled point;
(2.2.2) calculate interior angle: define a sampled point p 1With respect to another sampled point p 2Interior angle be configuration sampling point p 1Tangential direction and from a p 1The p that sets out 1, p 2Between the shortest path direction between angle;
(2.2.3) calculate interior distance hereinafter descriptor in shape: use the band direction apart from the instrument of the concentric circles of dividing as the calculating descriptor, the concentrically ringed center of circle is in certain configuration sampling point p i, suppose angular divisions be G interval, hypothesis has R concentric circles again, then R concentric circles and G angular interval combine and have constituted (R+1) * G interval; The interior distance and the interior angle that combine (2.2.1) and (2.2.2) acquisition then, the point of confirming on the profile other is with a p iFor the band direction in the center of circle apart from the distribution of dividing in concentrically ringed (R+1) * G interval, obtains a statistic histogram, this statistic histogram is interior apart from hereinafter descriptor in shape; So-called band direction is meant with sampled point p apart from division iTangential direction be the zero angle direction, with 360 degree spaces evenly be divided into G interval.
3. the method for retrieval similarity shape according to claim 1 and 2 is characterized in that, step (3.1) specifically comprises following process:
(3.1.1) the difference sequence of calculation { p i, i=1 ... N, in any some p iWith sequence { q j, j=1 ... any some q among the N jBetween characteristic distance c (i, j), this distance definition point p that (2.2) calculate that serves as reasons iStatistic histogram with the some q jStatistic histogram between distance, its mathematical definition is following:
c ( i , j ) = 1 2 Σ k = 1 ( R + 1 ) × G [ h A , i ( k ) - h B , j ( k ) ] 2 h A , i ( k ) + h B , j ( k )
H wherein A, i, h B, jBe the statistic histogram of definition in the step (2.2.3), subscript A and B represent h A, i(k), h B, j(k) be the statistic histogram of shape A and shape B respectively, c (i, j) expression point p iWith a q iCharacteristic distance;
Through calculating the characteristic distance of being had a few on the have a few and shape B on the shape A, the characteristic distance matrix of acquisition N * N c (i, j), i=1 ..., N, j=1 ..., N}:
C = c ( 1,1 ) c ( 1,2 ) . . . c ( 1 , N - 1 ) c ( 1 , N ) c ( 2,1 ) c ( 2,2 ) . . . c ( 2 , N - 1 ) c ( 2 , N ) . . . . . . . . . . . . . . . c ( N - 1,1 ) c ( N - 1,2 ) . . . c ( N - 1 , N - 1 ) c ( N - 1 , N ) c ( N , 1 ) c ( N , 2 ) . . . c ( N , N - 1 ) c ( N , N )
(3.1.2) two sequence { p of hypothesis i, i=1 ..., N} and { q j, j, j=1 ..., N} is at a p 1And q 1Dynamic programming method is used in place's alignment, tries to achieve the dissimilar degree of two sequences, as D 1, its solution procedure is following:
1. generate the empty matrix M of a N * N, be called the dynamic programming matrix;
2. initialization marginal condition:
M(0,0)＝0；
M(i,0)＝γ×i,i＝1,...,N；
M(0,j)＝γ×j,j＝1,...,N；
The scalar penalty factor of γ for setting;
3. according to from left to right, from top to bottom order, calculate successively each element in the dynamic programming matrix M value S (i, j), i=1 ..., N, j=1 ..., N, computing formula is following:
M ( i , j ) = min M ( i - 1 , j - 1 ) + c ( i , j ) M ( i - 1 , j ) + γ M ( i , j - 1 ) + γ
Wherein (i j) is a p to c iWith a q jBetween characteristic distance;
4. the value of each element is all calculated and is finished in the dynamic programming matrix, and then dynamic programming method finishes, the dissimilar degree D of two sequences 1=M (N, N);
(3.1.3) process of repetition (3.1.2) is supposed two sequence { p respectively iAnd { q iAt p 2With q 1Place's alignment, p 3With q 1Place's alignment ..., p NWith q 1Place's alignment uses dynamic programming algorithm to try to achieve dissimilar degree D respectively 2, D 3..., D N
(3.1.4) choose minimum that in try to achieve in the step (3.1.2) to (3.1.3) N the distance, as shape A, the measure of dissimilarity value between the B:
D(A,B)＝min(D i),i＝1,...,N。
CN201110106315A 2011-04-27 2011-04-27 Method for retrieving similarity shape Active CN102200999B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201110106315A CN102200999B (en) 2011-04-27 2011-04-27 Method for retrieving similarity shape

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201110106315A CN102200999B (en) 2011-04-27 2011-04-27 Method for retrieving similarity shape

Publications (2)

Publication Number Publication Date
CN102200999A CN102200999A (en) 2011-09-28
CN102200999B true CN102200999B (en) 2012-10-10

Family

ID=44661674

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201110106315A Active CN102200999B (en) 2011-04-27 2011-04-27 Method for retrieving similarity shape

Country Status (1)

Country Link
CN (1) CN102200999B (en)

Families Citing this family (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102306202B (en) * 2011-09-30 2013-09-04 中国传媒大学 High-dimension vector rapid searching algorithm based on block distance
CN102521287A (en) * 2011-11-28 2012-06-27 上海交通大学 Expectation model-based image re-ranking method
US9239395B2 (en) 2012-03-31 2016-01-19 University Of Science And Technology Of China Method and system of determining earthquake parameters, earthquake search engine
CN103365916B (en) * 2012-03-31 2016-12-14 中国科学技术大学 Earthquake parameter estimates acquisition methods and system, seismic events search engine
EP2915059B1 (en) * 2012-10-30 2021-05-26 Ent. Services Development Corporation LP Analyzing data with computer vision
CN103093461B (en) * 2013-01-14 2015-11-18 大连理工大学 A kind of shape matching method based on metric
US9305233B2 (en) 2013-09-12 2016-04-05 The Boeing Company Isotropic feature matching
CN104299226B (en) * 2014-09-22 2017-07-28 大冶市华兴玻璃有限公司 A kind of method of circular Bottle ＆ Can profile similarity retrieval
CN104680132B (en) * 2015-01-30 2017-11-21 哈尔滨工程大学 A kind of sonar target recognition methods based on Shape context method
CN106682693A (en) * 2016-12-23 2017-05-17 浙江大学 Identification method for plastic bottle body overlapping image
CN109426791B (en) * 2017-09-01 2022-09-16 深圳市金溢科技股份有限公司 Multi-site and multi-vehicle matching method, server and system
CN108563767B (en) 2018-04-19 2020-11-27 深圳市商汤科技有限公司 Image retrieval method and device
CN109685837B (en) * 2018-11-07 2024-03-22 中国矿业大学 Heterologous remote sensing image registration method based on feature structure similarity
CN113379777A (en) * 2021-05-26 2021-09-10 西安电子科技大学 Shape description and retrieval method based on minimum circumscribed rectangle vertical internal distance proportion

Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN1379364A (en) * 2001-03-29 2002-11-13 日本电气株式会社 Graph comparing device and graph comparing method

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2001011563A1 (en) * 1999-08-05 2001-02-15 Koninklijke Philips Electronics N.V. Detection of auxiliary data in an information signal

Patent Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN1379364A (en) * 2001-03-29 2002-11-13 日本电气株式会社 Graph comparing device and graph comparing method

Also Published As

Publication number Publication date
CN102200999A (en) 2011-09-28

Similar Documents

Publication Publication Date Title
CN102200999B (en) Method for retrieving similarity shape
Cucuringu Sync-rank: Robust ranking, constrained ranking and rank aggregation via eigenvector and SDP synchronization
Chang et al. Sparse Regularization in Fuzzy $ c $-Means for High-Dimensional Data Clustering
Rayana et al. Sequential ensemble learning for outlier detection: A bias-variance perspective
CN101567051B (en) Image matching method based on characteristic points
Fogel et al. Spectral ranking using seriation
CN107291847A (en) A kind of large-scale data Distributed Cluster processing method based on MapReduce
CN110188225B (en) Image retrieval method based on sequencing learning and multivariate loss
CN109325510B (en) Image feature point matching method based on grid statistics
Yang et al. A feature-metric-based affinity propagation technique for feature selection in hyperspectral image classification
Ma et al. Coronal mass ejection data clustering and visualization of decision trees
CN102495886A (en) Method for searching fingerprint database based on quantum algorithm
CN104978742A (en) Image registration method and image registration device based on cascade structure
CN111369601A (en) Remote sensing image registration method based on twin network
CN111507297A (en) Radar signal identification method and system based on measurement information matrix
CN110647647B (en) Closed graph similarity searching method based on time sequence complexity difference
Martínez et al. Adaptive sampling strategies for quickselects
Tsakiris Low-rank matrix completion theory via Plücker coordinates
Zhou et al. Hyperspectral image change detection by self-supervised tensor network
Amir et al. Period recovery of strings over the Hamming and edit distances
Schlegel et al. Visual localization and loop closing using decision trees and binary features
Jiang Sparse estimation based on square root nonconvex optimization in high-dimensional data
CN105512480B (en) The data-optimized processing method of wearable device based on editing distance
Zheng et al. Boundary adjusted network based on cosine similarity for temporal action proposal generation
Choi et al. Learning latent tree graphical models

Legal Events

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