CN101446980A - Tridimensional partial shape match and retrieval method based on color rotation picture - Google Patents

Abstract

The invention relates to a tridimensional partial shape match and retrieval method based on a color rotation picture, which comprises the steps as follows: a tridimensional shape surface is sampled to obtain a series of base points, and a partial shape characteristic is calculated at each base point; the partial shape characteristic is clustered to obtain a clustering center, and the clustering center is saved to form a partial shape characteristic dictionary; the clustering center proximal to each partial shape characteristic is searched in the partial shape characteristic dictionary, and the characteristic is replaced by a number mark of the clustering center; the rotation picture shape characteristic is calculated according to each base point with the number mark of the clustering center, to generate the color rotation picture characteristic; and the similarity between two color rotation pictures and the similarity between two tridimensional shapes are calculated. The tridimensional partial shape match and retrieval method solves the problems that a traditional tridimensional model retrieval method is not suitable for the search based on the partial shape similarity and is more sensitive to noises and shelters in tridimensional shape data.

Description

A kind of three-dimensional portion form fit and search method based on colored image rotating

Technical field

The present invention relates to computer vision and information retrieval field, be specifically related to a kind of three-dimensional portion form fit and search method based on colored image rotating.

Background technology

Along with the fast development of 3 D scene rebuilding technology and CAD Modeling Technology, three-dimensional model has all obtained extensive application in many fields, as computer animation, and computer-aided manufacturing, Computer Simulation drug design etc.On this basis, the quantity of three-dimensional model increases fast, and some have appearred on the net can be for the three-dimensional modeling data storehouse of downloading.Under many circumstances, from the three-dimensional modeling data storehouse, retrieve relevant three-dimensional model, often can directly satisfy user's needs or reduce the workload that designs brand-new three-dimensional model greatly.

Existing three-dimensional model search technology is mainly based on the three-dimensional shape features of integral body.The whole geometry feature of for example representing a three-dimensional model with a vector or matrix.On this basis, can calculate two similarity/distances between the 3D shape fast by two vectors of contrast or matrix.Like this, can be to each shape in the three-dimensional modeling data storehouse, sort according to the similarity/distance between they and the 3D shape that is used to inquire about, and will be in the database a small amount of and inquire about the most similar three-dimensional model of shape and return to the user, as a reference.

Although existing method is comparatively effective to searching the approaching three-dimensional model of global shape, yet a large amount of three-dimensional datas is to gather in being full of the environment that mixes in a large number and block.In this class three-dimensional scenic, specific 3D shape is mated and identification remains a challenging problem.In addition, search and the on all four three-dimensional model of global shape, often be difficult to provide valuable fresh information in actual applications.The 3D shape searching system of a practicality should be able to find the 3D shape that has part similarity (promptly partially overlapping) with query object.At last, existing method for searching three-dimension model general inapplicable and flexible 3 D shape can change flexibly as the shape of human body and animal, and existing rigid shape descriptor generally is not suitable for describing the similarity of this class 3D shape.

Other method, the purpose of 3D shape coupling is to find two corresponding relations between the similar 3D shape.For this reason, the problem of a key is the reliability of how to evaluate corresponding point matching.Existing work comprises in a large number based on partial 3 d shape description of geometric properties and their similarity measurement.Yet and the global shape feature class seemingly, and this class shape description blocks with the partial shape disappearance comparatively responsive equally to mixing information.Therefore often can not obtain satisfied effect in actual applications.

Summary of the invention

The purpose of this invention is to provide a kind of based on the three-dimensional portion form fit of colored image rotating and the method for retrieval, design a kind of new shape description " colored image rotating " and at its similarity measurement, thereby solving the conventional three-dimensional model retrieval method is not suitable for based on the partial shape similarity search with to the noise in the three-dimensional shape data with block comparatively sensitive issue.

In order to reach the foregoing invention purpose, the invention provides a kind of three-dimensional portion form fit and search method based on colored image rotating, described method comprises step:

S1 samples to each the 3D shape surface in the three-dimensional data base, obtains a series of basic points, and calculates a kind of local shape feature on each basic point;

S2 carries out cluster to described local shape feature and obtains cluster centre, preserves described cluster centre and constitutes the local shape feature lexicon;

S3 searches in described local shape feature lexicon and the immediate cluster centre of each local shape feature, and the numbering of this feature with cluster centre substituted;

S4, according to the building method of image rotating, the basic point that among the S1 each had a cluster centre numbering calculates the image rotating shape facility again, generates colored image rotating feature;

S5 carries out histogram successively and asks friendship, diffuse images, calculates similarity between two colored image rotatings and the similarity between two 3D shapes.

Wherein, in step S1, the local shape of being gathered is characterized as 10000-100000, and the basic point number that each 3D shape is sampled is 100-500.

Wherein, in step S2, described local shape feature lexicon comprises the cluster centre of the local shape feature of a series of call numbers and correspondence, wherein the corresponding a kind of color that is used for its corresponding particular shape characteristics of mark of each call number.

Wherein, utilize the K-means method that described local shape feature is carried out cluster.

Wherein, the number of described cluster centre is 500-2000.

Wherein, also comprise step among the described step S3: the numbering of record 3D shape each basic point characteristic of correspondence cluster centre, the i.e. color index of this basic point number.

Wherein, the computing method of the shape facility of described image rotating comprise the steps:

S4-1 for each basic point on the 3D shape, is the axle rotation with a facet along the normal of this basic point place 3D shape, and the basic point with each marker color of 3D shape in rotary course projects on this facet, forms colored two-dimensional points cloud and distributes;

S4-2, when two dimensional surface was carried out discretize, (j), wherein c was a call number for c, i, and (i j) is the discrete coordinates of two-dimensional space to generate 3-D histogram;

S4-3 adds one according to each corresponding unit in 3-D histogram, colored position of putting on the facet;

S4-4 carries out regularization to colored image rotating, be about to 3-D histogram whole unit numeral and normalizing.

Wherein, after described step S4-4, also comprise step

S4-5 writes down the cluster centre call number that is occurred in the described colored image rotating; With

S4-6, each locus of recording colour image rotating (i, j) numerical value of the call number of Chu Xianing and corresponding each call number.

Wherein, it is as follows to calculate the step of the similarity between two colored image rotatings among the described step S5:

S5-1, with the similarity between the colored image rotating be decomposed into each color diagram layer similarity and: $S_{\mathrm{CSI}}(I_1,I_2)=\underset{c}{\mathrm{\Σ}}S(I_1\left(c\right),I_2\left(c\right)),$ I wherein ₁, I ₂Be the histogram of two colored image rotatings, c is the index value of color dimension, I ₁(c), I ₂(c) be I ₁, I ₂C color diagram layer, be expressed as the histogram of two two dimensions;

S5-2 is with two under highest resolution figure layer I ₁(c), I ₂(c) be designated as

Carry out histogram intersection, obtain the numerical value that under this resolution, mates:

$a^0(c,i,j)=\min \{I_1^0(c,i,j),I_2^0(c,i,j)\}$

$n^0\left(c\right)=\underset{i,j}{\mathrm{\Σ}}{a}^0(c,i,j)$

S5-3, to each pixel i, j, with the coupling numerical value from

Corresponding unit in deduct:

$I_1^0(c,i,j)\←I_1^0(c,i,j)-a(c,i,j)$

$I_2^0(c,i,j)\←I_2^0(c,i,j)-a(c,i,j)$

S5-4 adopts the diffuse images method, reduces the resolution of two figure layers:

Will The figure layer that reduces a resolution is designated as To image

Each pixel (i, j), it will be evenly distributed to image

In pixel A wherein, b ∈ 0,1, and [x] is for being no more than the maximum integer of x;

S5-5 repeats the process of S5-2 to S5-4, and the resolution that reduces the figure layer successively is till forming single pixel;

S5-6 on this basis, defines similarities between two figure layers and is the value every layer of coupling, multiply by the weighted sum of pixel length of side inverse under each figure layer resolution:

$S(I_1\left(c\right),I_2\left(c\right))=\underset{i=0}{\overset{L}{\mathrm{\Σ}}}\frac{1}{2^i}{n}^i\left(c\right)$

Here, L is the number of plies of crossing over from the highest resolution to the lowest resolution.

Wherein, it is as follows to calculate the step of the similarity between two 3D shapes among the described step S5:

S5-7: define two colored image rotating f _i, g _jBetween basis tolerance be:

d(f _i，g _j)＝A-S _CSI(f _i，g _j)

Wherein A is arbitrarily greater than 1 constant;

S5-8 according to above-mentioned basis tolerance, utilizes the method for dozer distance, and the distance that defines between two 3D shapes is to satisfy under the column constraint down:

$e_{\mathrm{ij}}\≥0,\underset{i}{\mathrm{\Σ}}{e}_{\mathrm{ij}}\≤v_j,\underset{j}{\mathrm{\Σ}}{e}_{\mathrm{ij}}\≤u_i,\underset{\mathrm{ij}}{\mathrm{\Σ}}{e}_{\mathrm{ij}}=\min \{\underset{i}{\mathrm{\Σ}}{u}_i,\underset{j}{\mathrm{\Σ}}{v}_j\}.$

Minimum transport distance:

$D_{\mathrm{EMD}}(F,G)=\underset{\left\{e_{\mathrm{ij}}\right\}}{\min }\frac{\underset{i,j}{\mathrm{\Σ}}{e}_{\mathrm{ij}}d(f_i,g_j)}{\underset{i,j}{\mathrm{\Σ}}{e}_{\mathrm{ij}}}$

Wherein, e _IjExpression is from f _iBe transferred to g _jThe amount of feature, d (f _i, g _j) for describing the feature of a unit from f _iBe transferred to g _jThe basis tolerance of cost, f _iAnd g _jBe respectively two colored image rotating features, u _iAnd v _jIt is its corresponding weight.

Three-dimensional portion form fit and search method based on colored image rotating provided by the present invention, having solved the conventional three-dimensional model retrieval method is not suitable for based on the partial shape similarity search with to the noise in the three-dimensional shape data with block comparatively sensitive issue.The colored image rotating that generates among the present invention not only can reflect the geometric distributions of the three-dimensional point near the local shape of basic point, and has reflected the shape semantic information of these points self.The calculation procedure of the similarity measurement of Ti Chuing has been simulated the real process of level form fit preferably on this basis, in true form retrieval and coupling result is preferably arranged.

Description of drawings

Fig. 1 is three-dimensional portion form fit and the search method process flow diagram based on colored image rotating of the present invention;

Fig. 2 is the computation process synoptic diagram of colored image rotating local shape feature of the present invention.

Embodiment

Following examples are used to illustrate the present invention, but are not used for limiting the scope of the invention.

Fig. 1 is three-dimensional portion form fit and the search method process flow diagram based on colored image rotating of the present invention.As shown in Figure 1, the invention provides a kind of three-dimensional portion form fit and search method based on colored image rotating, described method comprises step: S1, sampled in each 3D shape surface in the three-dimensional data base, obtain a series of basic points, and on each basic point, calculate a kind of local shape feature; S2 carries out cluster to described local shape feature and obtains cluster centre, preserves described cluster centre and constitutes the local shape feature lexicon; S3 searches in described local shape feature lexicon and the immediate cluster centre of each local shape feature, and the numbering of this feature with cluster centre substituted; S4, according to the building method of image rotating, the basic point that among the S1 each had a cluster centre numbering calculates the image rotating shape facility again, generates colored image rotating feature; S5 carries out histogram successively and asks friendship, diffuse images, calculates similarity between two colored image rotatings and the similarity between two 3D shapes.

Wherein, in step S1, the local shape of being gathered is characterized as 10000-100000, and the basic point number that each 3D shape is sampled is 100-500.Wherein, in step S2, described local shape feature lexicon comprises the cluster centre of the local shape feature of a series of call numbers and correspondence, wherein the corresponding a kind of color that is used for its corresponding particular shape characteristics of mark of each call number.Wherein, utilize the K-means method that described local shape feature is carried out cluster.Wherein, the number of described cluster centre is 500-2000.Wherein, also comprise step among the described step S3: the numbering of record 3D shape each basic point characteristic of correspondence cluster centre, the i.e. color index of this basic point number.

Wherein, the computing method of the shape facility of described image rotating comprise the steps: S4-1, for each basic point on the 3D shape, is the axle rotation with a facet along the normal of this basic point place 3D shape, basic point with each marker color of 3D shape in rotary course projects on this facet, forms colored two-dimensional points cloud and distributes; S4-2, when two dimensional surface was carried out discretize, (j), wherein c was a call number for c, i, and (i j) is the discrete coordinates of two-dimensional space to generate 3-D histogram; S4-3 adds one according to each corresponding unit in 3-D histogram, colored position of putting on the facet; S4-4 carries out regularization to colored image rotating, be about to 3-D histogram whole unit numeral and normalizing; S4-5 writes down the cluster centre call number that is occurred in the described colored image rotating; And S4-6, each locus of recording colour image rotating (i, j) numerical value of the call number of Chu Xianing and corresponding each call number.

Wherein, the step of calculating the similarity between two colored image rotatings among the described step S5 is as follows: S5-1, with the similarity between the colored image rotating be decomposed into each color diagram layer similarity and: $S_{\mathrm{CSI}}(I_1,I_2)=\underset{c}{\mathrm{\Σ}}S(I_1\left(c\right),I_2\left(c\right)),$ I wherein ₁, I ₂Be the histogram of two colored image rotatings, c is the index value of color dimension, I ₁(c), I ₂(c) be I ₁, I ₂C color diagram layer, be expressed as the histogram of two two dimensions; S5-2 is with two under highest resolution figure layer I ₁(c), I ₂(c) be designated as

Carry out histogram intersection, obtain the numerical value that under this resolution, mates:

$a^0(c,i,j)=\min \{I_1^0(c,i,j),I_2^0(c,i,j)\}$

$n^0\left(c\right)=\underset{i,j}{\mathrm{\Σ}}{a}^0(c,i,j)$

S5-3, to each pixel i, j, with the coupling numerical value from

Corresponding unit in deduct:

$I_1^0(c,i,j)\←I_1^0(c,i,j)-a(c,i,j)$

$I_2^0(c,i,j)\←I_2^0(c,i,j)-a(c,i,j)$

S5-4 adopts the diffuse images method, reduces the resolution of two figure layers: will

The figure layer that reduces a resolution is designated as , to image Each pixel (i, j), it will be evenly distributed to image

In pixel

A wherein, b ∈ 0,1, and [x] is for being no more than the maximum integer of x;

S5-5 repeats the process of S5-2 to S5-4, and the resolution that reduces the figure layer successively is till forming single pixel; S5-6 on this basis, defines similarities between two figure layers and is the value every layer of coupling, multiply by the weighted sum of pixel length of side inverse under each figure layer resolution:

$S(I_1\left(c\right),I_2\left(c\right))=\underset{i=0}{\overset{L}{\mathrm{\Σ}}}\frac{1}{2^i}{n}^i\left(c\right)$

Here, L is the number of plies of crossing over from the highest resolution to the lowest resolution.

Wherein, it is as follows to calculate the step of the similarity between two 3D shapes among the described step S5: S5-7: define two colored image rotating f _i, g _jBetween basis tolerance be:

d(f _i，g _j)＝A-S _CSI(f _i，g _j)

Wherein A is arbitrarily greater than 1 constant;

S5-8 according to above-mentioned basis tolerance, utilizes the method for dozer distance, and the distance that defines between two 3D shapes is to satisfy under the column constraint down:

$e_{\mathrm{ij}}\≥0,\underset{i}{\mathrm{\Σ}}{e}_{\mathrm{ij}}\≤v_j,\underset{j}{\mathrm{\Σ}}{e}_{\mathrm{ij}}\≤u_i,\underset{\mathrm{ij}}{\mathrm{\Σ}}{e}_{\mathrm{ij}}=\min \{\underset{i}{\mathrm{\Σ}}{u}_i,\underset{j}{\mathrm{\Σ}}{v}_j\}.$

Minimum transport distance:

$D_{\mathrm{EMD}}(F,G)=\underset{\left\{e_{\mathrm{ij}}\right\}}{\min }\frac{\underset{i,j}{\mathrm{\Σ}}{e}_{\mathrm{ij}}d(f_i,g_j)}{\underset{i,j}{\mathrm{\Σ}}{e}_{\mathrm{ij}}}$

Wherein, e _IjExpression is from f _iBe transferred to g _jThe amount of feature, d (f _i, g _j) for describing the feature of a unit from f _iBe transferred to g _jThe basis tolerance of cost, f _iAnd g _jBe respectively two colored image rotating features, u _iAnd v _jIt is its corresponding weight.

Introduce the technical scheme among the present invention below in detail.

1. the computation process of colored image rotating

Colored image rotating generates by two steps.The target of phase one is to give to each point on the 3D shape and represent its index value of shape attribute on every side.At first from the three-dimensional modeling data storehouse, gather a large amount of three-dimensional local shape features.This is by to each model in the database, and at a large amount of basic point of its surperficial up-sampling, and utilization local shape descriptor extraction feature obtains on each basic point.For each model, use Monte-Carlo method uniform sampling on 3D shape.Then, we use K-Means (or other) clustering method, and the shape facility of these magnanimity is carried out cluster.Class center for obtaining after the cluster can be considered as independently, recurrent feature mode.Because all class centers have constituted " dictionary " of local feature, so also usually be known as a word in " dictionary " of each cluster centre, its call number is arranged in dictionary.This process off-line of building dictionary is finished, and " dictionary " that will generate afterwards stores, so that use in the future.

For online characteristic extraction procedure,,, obtain a series of basic points at the 3D shape up-sampling to use above-mentioned Monte-Carlo algorithm as the 3D shape of query object.In each basic point position, utilization local shape descriptor extracts a feature, and searches in dictionary and " word " that this characteristic distance is nearest.To be somebody's turn to do " word " call number in " dictionary " and give this basic point, indicate its local shape on every side.

In this article, " word " is called as " color " at the index of dictionary.Therefore the process of front can be regarded as each point on the 3D shape, according to the local shape around this point, gives a color to it.Therefore, shape description that is proposed is called " colored image rotating ".Yet, all give a color for all the dense sampled points on the 3D shape, have bigger computational complexity.The shortcut of handling this problem is such: only in shape small sample point (for example 500) is dyeed.And then, give its color of closest approach in these 500 points to each dense sampled point in shape.

Above-mentioned is the process of calculating the phase one of colored image rotating.In this stage, there are a lot of degree of freedom to select for flexible.At first, the kind of local shape descriptor can be selected arbitrarily.For example, can use original image rotating method, in shape point is dyeed.Secondly, the size of the support scope of local shape descriptor can be specified flexibly according to particular problem.Must, for the higher three-dimensional data of noise, the reply descriptor is set bigger support scope.On the contrary, for there being more 3D shape of blocking, the support scope should not established too much, to avoid obscuring of shape information.

Introduce second step of the colored image rotating of structure now.Before introducing detail, at first introduce original image rotating shape description.As shown in Figure 2, image rotating is expressed as the vector of a two dimension and traditional image similarity.As its name suggests, image rotating is to be the facet of axle by rotation with the basic point normal direction, scanning three-dimensional dot generation in shape.We introduce more ins and outs here.As shown in the figure, suppose that B is certain point in shape, here as the basic point of image rotating.P is another point in shape, in the support scope of this image rotating.Γ is positioned at a section of the shape at B place.Thus, motion vector

, can be broken down into two parts, be positioned at the part in section

With perpendicular to the part in section

Note Length be w,

Length be h.Like this, (w, (i j) casts a ticket to pixel h) to correspondence on the image rotating.The support scope of image rotating (W H) has guaranteed to have only satisfied in shape following constraint | h|≤H, | the point of w|≤W just can project on the Plane of rotation, otherwise be left in the basket.In other words, the description scope of image rotating that supported scope definition to the local shape around the basic point.

In the original that proposes image rotating, only with the summit of grid, the sampled point as shape generates image rotating.These are being irregular distribution in shape, and may be very sparse, so effect tends to be a greater impact.In the method, the utilization monte carlo method has been gathered evenly, and dense shape sampled point has been avoided this problem.

For colored image rotating, the crucial difference of it and original image rotating is: the sampled point of shape has been endowed color (call number), therefore, and need be to Plane of rotation with the spot projection of colour.In addition, different with the image rotating method of two dimension, colored image rotating is represented by 3-D histogram.Its each unit by three round valuess (c, i, j) index, c represents a specific color here.At last, also need to carry out regularization to colored image rotating, be about to it whole unit numeral and normalizing.

2, the space of colored image rotating storage

Have only three Color Channels different with natural image, it is individual to several thousand Color Channels that colored image rotating has hundreds of.See that from the teeth outwards people may think that colored image rotating need be than rotation image much bigger storage space, yet in fact do not need.Because colored image rotating is a kind of sparse expression, can utilize this point to reduce its storage space greatly.In fact, although " characteristics dictionary " has hundreds and thousands of " word ", " word " that occurs in a colored image rotating (being color) may be also few.Especially, for fixing _{(i, j)}, may have only few color to appear on this pixel location.Therefore, at first at the place that begins of colored image rotating file, be recorded in the color index that occurred among it number.Then, for each _{(i, j)}The position only need be recorded in the index value of a few color that this position occurs and the numerical value of corresponding each color.In our test, although 1500 words are arranged in the dictionary, the storage space of colored image rotating only is about 3 times of common image rotating.

Index structure makes the similarity of colored image rotating calculate more quick.Before two colored image rotatings of comparison, at first the color index that occurs among them is asked friendship, a need compare these common color diagram layers and get final product then.

3, the similarity measurement of colored image rotating

Introduce below and how to calculate two similaritys between the colored image rotating.At first, regard the colored image rotating of three-dimensional the image of a two dimension as, at each pixel location _{(i, j)}All placed the point of a lot of different colours.Like this, relatively the task of two colored image rotatings has just converted the point that mates two groups of colours that distribute on two planes of delineation to.Since every kind of color is represented a kind of local shape of particular type, have only the point of same color to mate.Therefore, the similarity between the colored image rotating can be decomposed into similarity between each color layers and:

$S_{\mathrm{CSI}}(I_1,I_2)=\underset{c}{\mathrm{\Σ}}S(I_1\left(c\right),I_2\left(c\right))$

I wherein ₁, I ₂Be the colored image rotating histogram of two three-dimensionals, c is the index value of color dimension, I ₁(c), I ₂(c) be I ₁, I ₂C color diagram layer, be expressed as the histogram of two two dimensions.

Problem has just converted to and has calculated two similarities between the two-dimensional histogram now.For example, can calculate their similarity with the normalized degree of correlation.But this is very unsuitable to colored image rotating.Because in colored image rotating, the point on each 3D shape is color on the mark all, and (i, j) Pi Pei probability is often very little and the point of same color is just at same location of pixels.Therefore, should allow to be positioned at the some coupling of the same color of different pixels position on the plane.

Therefore, in order to describe the similarity between two image layer of colored image rotating, should adopt the similarity measurement of unit cross-matched.We provide one very efficiently, and the histogram of multiresolution is asked the friendship method, calculate two similaritys between the figure layer.With two initial figure layer I ₁(c), I ₂(c) be designated as Because they are in the highest resolution.These two X-Y scheme layers are intersected, calculate numerical value in this layer coupling:

$a^0(c,i,j)=\min \{I_1^0(c,i,j),I_2^0(c,i,j)\}$

$n^0\left(c\right)=\underset{i,j}{\mathrm{\Σ}}{a}^0(c,i,j)$

To each pixel _{(i, j)}, with the coupling numerical value from Corresponding unit in deduct:

$I_1^0(c,i,j)\←I_1^0(c,i,j)-a(c,i,j)$

$I_2^0(c,i,j)\←I_2^0(c,i,j)-a(c,i,j)$

Next,

Size all narrow down to originally 1/2nd, generate two width of cloth images of low resolution.Unit different in the histogram calculate the different problem (being space bias) of contribution to similarity, have adopted the method for diffuse images.Two figure layers that are located at next low resolution are To image

Each pixel (i j), is diffused into image with it In pixel

A wherein, b ∈ 0,1, and [x] is for being no more than the maximum integer of x.Right

Carry out similar operation.

Be not difficult to verify under the situation of this diffusion, two pixels that any side is adjacent, their values of 50% all can merge under next resolution.For the diagonal angle neighboring pixels, their values of 25% or 50% can be also laminated at next.Therefore, directly merge the space bias that histogram unit brings, just greatly reduce.

At last, two figure layers of definition I ₁(c), I ₂(c) similarity is the value every layer of coupling, multiply by the weighted sum of pixel length of side inverse:

$S(I_1\left(c\right),I_2\left(c\right))=\underset{i=0}{\overset{L}{\mathrm{\Σ}}}\frac{1}{2^i}{n}^i\left(c\right)$

Here, L is the number of plies of lowest resolution correspondence, and this moment, original figure layer was shrunk to a pixel.By top structure, be easy to prove that similarity measurement S has the computation complexity of the O of a linearity (n), here the number of pixel in the n representative image.Like this, described similarity measurement, the histogram similarity measurement with general classics on the magnitude of complexity computing time is consistent.

4, the similarity between the calculating 3D shape

Above by the agency of the building method of colored image rotating, and computing method of similarity measurement between this feature.Yet,, need to calculate the similarity between two 3D shapes for three-dimensional model search.Below will utilize the method for dozer distance, the similarity between two colored image rotatings will be generalized on two 3D shapes.

Briefly, the set F:{ (f of given two local features ₁, u ₁), (f ₂, u ₂) ... (f _m, u _m) and G:{ (g ₁, v ₁), (g ₂, v ₂) ... (g _n, v _n), the dozer distance calculation goes out the minimum cost on the feature that feature with a set is transported to another set.Here f _iAnd g _jBe respectively the local feature of two set, u _iAnd v _jIt is corresponding weight.Satisfying under the following constraint in dozer distance on the mathematics:

$e_{\mathrm{ij}}\≥0,\underset{i}{\mathrm{\Σ}}{e}_{\mathrm{ij}}\≤v_j,\underset{j}{\mathrm{\Σ}}{e}_{\mathrm{ij}}\≤u_i,\underset{\mathrm{ij}}{\mathrm{\Σ}}{e}_{\mathrm{ij}}=\min \{\underset{i}{\mathrm{\Σ}}{u}_i,\underset{j}{\mathrm{\Σ}}{v}_j\}$

Found the solution following optimization problem:

$D_{\mathrm{EMD}}(F,G)=\underset{\left\{e_{\mathrm{ij}}\right\}}{\min }\frac{\underset{i,j}{\mathrm{\Σ}}{e}_{\mathrm{ij}}d(f_i,g_j)}{\underset{i,j}{\mathrm{\Σ}}{e}_{\mathrm{ij}}}$

Wherein, g _IjExpression is from f _iBe transferred to g _jThe amount of feature, d (f _i, g _j) for describing the feature of a unit from f _iBe transferred to g _jThe basis tolerance of cost.Therefore, above-mentioned several constraint has reflected the natural restriction relation of feature " mass conservation " in the transport process.

Because basic point uniform sampling on 3D shape of colored image rotating obtains, therefore, the weight of all features is identical.Definition basis tolerance d is that a constant deducts two similarities between the colored image rotating:

d(f _i，g _j)＝A-S _CSI(f _i，g _j)

As previously described, because colored image rotating descriptor is normalized, be easy to prove S _CSI(f _i, g _j)≤1.Therefore as long as constant is made as A〉1 just just be enough to guarantee basis tolerance perseverance.Further, be easy to prove that the value of constant A can freely select under these conditions, because finally calculate, the distance D between two three-dimensional models _EMD(F G) differs a constant.

Above embodiment only is used to illustrate the present invention; and be not limitation of the present invention; the those of ordinary skill in relevant technologies field; under the situation that does not break away from the spirit and scope of the present invention; can also make various variations; therefore all technical schemes that are equal to also belong to category of the present invention, and scope of patent protection of the present invention should be limited by its claim.

Claims (10)

1, a kind of three-dimensional portion form fit and search method is characterized in that, said method comprising the steps of:

S1 samples to each the 3D shape surface in the three-dimensional data base, obtains a series of basic points, and calculates any one local shape feature on each basic point;

S2 carries out cluster to described local shape feature and obtains cluster centre, numbering is set for each cluster centre, preserves described cluster centre and numbering thereof, constitutes the local shape feature lexicon;

S3 searches in described local shape feature lexicon and the immediate cluster centre of each local shape feature, and the numbering of this feature with cluster centre substituted;

S4, according to the building method of image rotating, the basic point that among the S1 each had a cluster centre numbering calculates the image rotating shape facility again, generates colored image rotating feature;

S5 carries out histogram successively and asks friendship, diffuse images, calculates similarity between two colored image rotatings and the similarity between two 3D shapes.

2, three-dimensional portion form fit as claimed in claim 1 and search method is characterized in that, in step S1, the local shape of being gathered is characterized as 10000-100000, and the basic point number that each 3D shape is sampled is 100-500.

3, three-dimensional portion form fit as claimed in claim 1 and search method, it is characterized in that, in step S2, described local shape feature lexicon comprises the cluster centre of the local shape feature of call number and correspondence, wherein the corresponding a kind of color that is used for its corresponding particular shape characteristics of mark of each call number.

4, three-dimensional portion form fit as claimed in claim 3 and search method is characterized in that, utilize the K-means method that described local shape feature is carried out cluster.

As each described three-dimensional portion form fit and search method among the claim 1-4, it is characterized in that 5, the number of described cluster centre is 500-2000.

6, three-dimensional portion form fit as claimed in claim 1 and search method is characterized in that, also comprise step among the described step S3: the numbering of record 3D shape each basic point characteristic of correspondence cluster centre, the i.e. color index of this basic point number.

7, three-dimensional portion form fit as claimed in claim 1 and search method is characterized in that, the computing method of the shape facility of described image rotating comprise the steps:

S4-1 for each basic point on the 3D shape, is the axle rotation with a facet along the normal of this basic point place 3D shape, and the basic point with each marker color of 3D shape in rotary course projects on this facet, forms colored two dimensional surface;

S4-2, when two dimensional surface was carried out discretize, (j), wherein c was a call number for c, i, and (i j) is the discrete coordinates of two-dimensional space to generate 3-D histogram;

S4-3 adds one according to each corresponding unit in 3-D histogram, colored position of putting on the facet;

S4-4 carries out regularization to colored image rotating, be about to 3-D histogram whole unit numeral and normalizing.

8, three-dimensional portion form fit as claimed in claim 7 and search method is characterized in that, after described step S4-4, also comprise step

S4-5: write down the cluster centre call number that is occurred in the described colored image rotating;

S4-6: each locus of recording colour image rotating (i, j) numerical value of the call number of Chu Xianing and corresponding each call number.

9, three-dimensional portion form fit as claimed in claim 1 and search method is characterized in that, the step of calculating among the described step S5 similarity between two colored image rotatings is as follows:

S5-1, with the similarity between the colored image rotating be decomposed into each color diagram layer similarity and: $S_{\mathrm{CSI}}(I_1,I_2)=\underset{c}{\mathrm{\Σ}}S(I_1\left(c\right),I_2\left(c\right)),$ I wherein ₁, I ₂Be the histogram of two colored image rotatings, c is the index value of color dimension, I ₁(c), I ₂(c) be I ₁, I ₂C color diagram layer, be expressed as the histogram of two two dimensions;

S5-2 is with two under highest resolution figure layer I ₁(c), I ₂(c) be designated as

,

, carry out histogram intersection, obtain the numerical value that under this resolution, mates:

$a^0(c,i,j)=\min \{I_1^0(c,i,j),I_2^0(c,i,j)\}$

$n^0\left(c\right)=\underset{i,j}{\mathrm{\Σ}}{a}^0(c,i,j)$

S5-3, to each pixel i, j, with the coupling numerical value from

Corresponding unit in deduct:

$I_1^0(c,i,j)\←I_1^0(c,i,j)-a(c,i,j)$

$I_2^0(c,i,j)\←I_2^0(c,i,j)-a(c,i,j)$

S5-4 adopts the diffuse images method, reduces the resolution of two figure layers:

Will

The figure layer that reduces a resolution is designated as To image

Each pixel (i, j), it will be evenly distributed to image In pixel A wherein, b ∈ 0,1, and [x] is for being no more than the maximum integer of x;

S5-5 repeats the process of S5-2 to S5-4, and the resolution that reduces the figure layer successively is till forming single pixel;

S5-6 on this basis, defines similarities between two figure layers and is the value every layer of coupling, multiply by the weighted sum of pixel length of side inverse under each figure layer resolution:

$S(I_1\left(c\right),I_2\left(c\right))=\underset{i=0}{\overset{L}{\mathrm{\Σ}}}\frac{1}{2^i}{n}^i\left(c\right)$

Wherein, L is the number of plies of crossing over from the highest resolution to the lowest resolution.

As claim 1 or 9 described three-dimensional portion form fit and search methods, it is characterized in that 10, the step of calculating the similarity between two 3D shapes among the described step S5 is as follows:

S5-7: define two colored image rotating f _i, g _jBetween basis tolerance be:

d(f _i，g _j)＝A-S _CSI(f _i，g _j)

Wherein A is arbitrarily greater than 1 constant;

S5-8, according to above-mentioned basis tolerance, the distance between two 3D shapes satisfies column constraint down:

$e_{\mathrm{ij}}\≥0,\underset{i}{\mathrm{\Σ}}{e}_{\mathrm{ij}}\≤v_j,\underset{j}{\mathrm{\Σ}}{e}_{\mathrm{ij}}\≤u_i,\underset{\mathrm{ij}}{\mathrm{\Σ}}{e}_{\mathrm{ij}}=\min \{\underset{i}{\mathrm{\Σ}}{u}_i,\underset{j}{\mathrm{\Σ}}{v}_j\}.$

Minimum transport distance be:

$D_{\mathrm{EMD}}(F,G)=\underset{\left\{e_{\mathrm{ij}}\right\}}{\min }\frac{\underset{i,j}{\mathrm{\Σ}}{e}_{\mathrm{ij}}d(f_i,g_j)}{\underset{i,j}{\mathrm{\Σ}}{e}_{\mathrm{ij}}}$

Wherein, e _IjExpression is from f _iBe transferred to g _jThe amount of feature, d (f _i, g _j) for describing the feature of a unit from f _iBe transferred to g _jThe basis tolerance of cost, f _iAnd g _jBe respectively two colored image rotating features, u _iAnd v _jIt is its corresponding weight.

Images