CN103903245A - Interactive co-segmentation method for three-dimensional model set - Google Patents

Abstract

The invention discloses an interactive co-segmentation method for a three-dimensional model set. The method comprises the following steps: the automatic co-segmentation step in which each model is divided into at least two components; the consistency judgment step in which consistency judgment and cluster analysis are performed on two components in any two models to judge same components; and the interactive step in which user judgment is performed on the same components and the process returns to the consistency judgment step until the best result is generated. According to the invention, the knowledge of a user can be integrated into co-segmentation, interactive co-segmentation is performed on a large-scale model set, advantages of fast calculation, easy operation and high segmentation accuracy can be realized, and the method has broad application prospects in fields, such as digital entertainment and art, consumer electronics, medical image processing, object identification, etc.

Description

Interactive mode for three-dimensional model collection is divided into segmentation method

[technical field]

The present invention relates to graph and image processing technical field, particularly cut apart the segmentation method that is divided into of one group of semantic identical but three-dimensional model set that shape difference is very large simultaneously.

[background technology]

It is that 3-D geometric model is resolved into one group of Limited Number, has a sub-grid of meaning separately that three-dimensional model is cut apart.But the definition of " meaningful " and people's cognition are closely related, it is a very subjective concept.According to the different definition to " meaningful ", people have designed multiple three-dimensional model dividing method, such as dividing ridge method, K-means method, Mean-shift method, region growing method, hierarchy clustering method based on the matching of secondary pel, based on symmetric figure dividing method, based on side the Spectral Clustering, random walk model method, random cutting method etc. of distance.But, for large-scale complex model, the needed condition of these methods is all difficult to meet, and traditional cutting procedure is not considered the relation between different mould shapes, after divided ownership model, set up again individually the consistance that corresponding relation is difficult to guarantee segmentation result, real to be improved.

[summary of the invention]

Because the defect of prior art dividing method is necessary to provide a kind of segmentation method that is divided into based on active semi-supervised learning quick, convenient operation that calculates.

For achieving the above object, the present invention adopts following scheme:

Interactive mode for three-dimensional model collection is divided into a segmentation method, and wherein, the method comprises:

Automatically be total to segmentation step: each model is divided into at least two parts;

Consistance determining step: two parts in any two models are carried out to consistance judgement and do cluster analysis and judge same base part;

Interactive step: will carry out user's judgement and get back to consistance determining step until generate optimum with base part.

The described interactive mode for three-dimensional model collection is divided into segmentation method, and wherein, this is divided at most five parts by each model in segmentation step automatically altogether.

The described interactive mode for three-dimensional model collection is divided into segmentation method, and wherein, this consistance determining step adopts priori unit to carry out consistance judgement.

The described interactive mode for three-dimensional model collection is divided into segmentation method, wherein, if model has between K mark, has n point in class C, and x is a point in class C, and the silhouette coefficient of x is:

Wherein

the mean distance that x arrives point in class C,

$a\left(x\right)=\frac{1}{n_k-1}\underset{y\∈C_k,y\≠x}{\mathrm{\Σ}}d(x,y),$

B (x) is the minimum average B configuration distance that x arrives point in other classes:

$b\left(x\right)=\underset{h\≠k}{\underset{h=1,...,K}{\min }}\left[\frac{1}{n_h}\underset{y\∈C_h}{\mathrm{\Σ}}d(x,y)\right].$

The described interactive mode for three-dimensional model collection is divided into segmentation method, and wherein, s is (x) the number between 0 to 1.

The described interactive mode for three-dimensional model collection is divided into segmentation method, and wherein, when s (x) recommends user near 0 point, the point that in the class of adjacent two classes of s (x), s (x) is very large is simultaneously recommended user together as definite point.

The described interactive mode for three-dimensional model collection is divided into segmentation method, and wherein, in this interactive step, user is judged as similar, inhomogeneity or is unable to explain clearly cluster parts.

The described interactive mode for three-dimensional model collection is divided into segmentation method, wherein, also comprises a step, and the judged result of user in this interactive step is grouped into this priori storage unit.

Compared to existing technology, during the present invention can be dissolved into user's knowledge and be divided into and cut, interactive mode is divided into cuts extensive model set, there is calculating quick, simple to operate, segmentation precision is high, calculating is quick, has broad application prospects in fields such as digital entertainment and art, consumer electronics, Medical Image Processing, target identifications.

[accompanying drawing explanation]

Fig. 1 is the present invention is divided into segmentation method process flow diagram for the interactive mode of three-dimensional model collection.

[embodiment]

Below in conjunction with diagram, the present invention will be described.

Invention thought of the present invention is divided into this problem of cutting as how minimum input completes around user, exploration can be merged user knowledge and be made full use of the segmentation method that is divided into of model information.

In general Models Sets all comprises hundreds and thousands of models conventionally, and each model be divided into cut after again by

$S\left(x\right)=\frac{b\left(x\right)-a\left(x\right)}{\max [b\left(x\right),a\left(x\right)]},$

At least be decomposed into 2-5 parts, the correctness of the each segmentation result of manual examination (check) is a very loaded down with trivial details task.Therefore, be uncertain point by being positioned at class edge in detection cluster result, they are recommended to user, allow user determine whether similar or inhomogeneity, thereby improve the efficiency of Interactive Segmentation.Hypothesized model has between K mark, has n point in class C, and x is a point in class C, and the silhouette coefficient of x can be defined as:

Wherein a (x) is the mean distance that x arrives point in class C,

$a\left(x\right)=\frac{1}{n_k-1}\underset{y\∈C_k,y\≠x}{\mathrm{\Σ}}d(x,y),$

B (x) is the minimum average B configuration distance that x arrives point in other classes

$b\left(x\right)=\underset{h\≠k}{\underset{h=1,...,K}{\min }}\left[\frac{1}{n_h}\underset{y\∈C_h}{\mathrm{\Σ}}d(x,y)\right].$

S is (x) the number between 0 to 1.When s (x) is near 0 time, this point is just on the edge of class, and we find out such point and recommend user, and the point that in the class of adjacent two classes of s (x), s (x) is very large is simultaneously as determining that point recommends user together.

Be exactly more than in Fig. 1 based on being divided into of priori cut and be total to segmentation result consistency analysis the two interdepend, the former is by user to being dissolved into be alternately divided into and cutting of model, the latter chooses suitable model allows user carry out alternately.These two is to be divided into and to cut for alone family, is that multi-user collaborative interactive mode is divided into the basis of cutting.

Carry out consistance judgement so this priori unit is equivalent to an experts database, can constantly judge to increase according to user again simultaneously.Particularly, after multiuser state repeatedly uses, accumulate more prioris and more can make to judge common splitting speed quickening, raise the efficiency.

As shown in Figure 1, after cutting based on being divided into of priori unit, between user's designated model sub-block, whether necessarily belong to similar or inhomogeneity, then system cluster again.In most preferred embodiment, in user interface, adopt blue line to represent necessarily to belong to similar, the red inhomogeneity that represents, yellow line represents uncertain.But they are neighbours in attribute space.When user has inputted blue line and red line, the position of the each point of system iterative is until convergence.

Through Fig. 1 circulation so repeatedly, until there is optimum.This is divided into segmentation method is the segmentation method that is divided into that can learn user knowledge, thereby improves segmentation result according to user's knowledge.In order to reduce user's workload, adopt interactive mode, a large amount of not model informations of interaction process can be combined with a small amount of user's input information, thus the accuracy rate of raising method.

Above the present invention is described in detail, has applied specific case herein principle of the present invention and embodiment are set forth, the explanation of above embodiment is just for helping to understand core concept of the present invention; , for one of ordinary skill in the art, according to thought of the present invention, all will change in specific embodiments and applications, in sum, this description should not be construed as limitation of the present invention meanwhile.

Claims (8)

1. be divided into a segmentation method for the interactive mode of three-dimensional model collection, it is characterized in that, the method comprises:

Automatically be total to segmentation step: each model is divided into at least two parts;

Consistance determining step: two parts in any two models are carried out to consistance judgement and do cluster analysis and judge same base part;

Interactive step: will carry out user's judgement and get back to consistance determining step until generate optimum with base part.

2. the interactive mode for three-dimensional model collection according to claim 1 is divided into segmentation method, it is characterized in that, this is divided at most five parts by each model in segmentation step automatically altogether.

3. the interactive mode for three-dimensional model collection according to claim 1 is divided into segmentation method, it is characterized in that, this consistance determining step adopts priori unit to carry out consistance judgement.

4. be divided into segmentation method according to the interactive mode for three-dimensional model collection described in claim 1 or 3, it is characterized in that, if model has between K mark, have n point in class C, x is a point in class C, and the silhouette coefficient of x is: $S\left(x\right)=\frac{b\left(x\right)-a\left(x\right)}{\max [b\left(x\right),a\left(x\right)]},$

Wherein a (x) is the mean distance that x arrives point in class C,

$a\left(x\right)=\frac{1}{n_k-1}\underset{y\∈C_k,y\≠x}{\mathrm{\Σ}}d(x,y),$

B (x) is the minimum average B configuration distance that x arrives point in other classes:

$b\left(x\right)=\underset{h\≠k}{\underset{h=1,...,K}{\min }}\left[\frac{1}{n_h}\underset{y\∈C_h}{\mathrm{\Σ}}d(x,y)\right].$

5. the interactive mode for three-dimensional model collection according to claim 4 is divided into segmentation method, it is characterized in that, s is (x) the number between 0 to 1.

6. the interactive mode for three-dimensional model collection according to claim 5 is divided into segmentation method, it is characterized in that, when s (x) recommends user near 0 point, the point that in the class of adjacent two classes of s (x), s (x) is very large is simultaneously recommended user together as definite point.

7. the interactive mode for three-dimensional model collection according to claim 1 is divided into segmentation method, it is characterized in that, in this interactive step, user is judged as similar, inhomogeneity or is unable to explain clearly cluster parts.

8. the interactive mode for three-dimensional model collection according to claim 7 is divided into segmentation method, it is characterized in that, also comprises a step, and the judged result of user in this interactive step is grouped into this priori storage unit.

Images