CN105354593A - NMF (Non-negative Matrix Factorization)-based three-dimensional model classification method - Google Patents
NMF (Non-negative Matrix Factorization)-based three-dimensional model classification method Download PDFInfo
- Publication number
- CN105354593A CN105354593A CN201510695635.2A CN201510695635A CN105354593A CN 105354593 A CN105354593 A CN 105354593A CN 201510695635 A CN201510695635 A CN 201510695635A CN 105354593 A CN105354593 A CN 105354593A
- Authority
- CN
- China
- Prior art keywords
- dimensional model
- nmf
- matrix
- collection
- classification
- 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.)
- Granted
Links
- 239000011159 matrix material Substances 0.000 title claims abstract description 98
- 238000000034 method Methods 0.000 title claims abstract description 70
- 238000012800 visualization Methods 0.000 claims abstract description 15
- 238000004364 calculation method Methods 0.000 claims abstract description 14
- 230000002452 interceptive effect Effects 0.000 claims abstract 4
- 230000000007 visual effect Effects 0.000 claims description 43
- 230000008569 process Effects 0.000 claims description 19
- 238000013461 design Methods 0.000 claims description 13
- 230000001105 regulatory effect Effects 0.000 claims description 10
- 238000013139 quantization Methods 0.000 claims description 8
- 238000006243 chemical reaction Methods 0.000 claims description 4
- 238000009792 diffusion process Methods 0.000 claims description 4
- 230000003993 interaction Effects 0.000 abstract description 2
- 238000004422 calculation algorithm Methods 0.000 description 9
- 238000010586 diagram Methods 0.000 description 8
- 238000004458 analytical method Methods 0.000 description 6
- 230000006870 function Effects 0.000 description 5
- 230000004048 modification Effects 0.000 description 5
- 238000012986 modification Methods 0.000 description 5
- 230000008901 benefit Effects 0.000 description 4
- 238000013480 data collection Methods 0.000 description 3
- 230000006872 improvement Effects 0.000 description 3
- 238000007794 visualization technique Methods 0.000 description 3
- 230000008859 change Effects 0.000 description 2
- 238000000354 decomposition reaction Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 239000000203 mixture Substances 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 230000009467 reduction Effects 0.000 description 2
- 238000004088 simulation Methods 0.000 description 2
- 201000004569 Blindness Diseases 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000007418 data mining Methods 0.000 description 1
- 238000013499 data model Methods 0.000 description 1
- 238000013079 data visualisation Methods 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 230000004992 fission Effects 0.000 description 1
- 238000011478 gradient descent method Methods 0.000 description 1
- 239000004615 ingredient Substances 0.000 description 1
- 238000012804 iterative process Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 230000008520 organization Effects 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 238000004064 recycling Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
- 238000013106 supervised machine learning method Methods 0.000 description 1
- 238000013107 unsupervised machine learning method Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
- G06F18/24133—Distances to prototypes
- G06F18/24137—Distances to cluster centroïds
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T17/00—Three dimensional [3D] modelling, e.g. data description of 3D objects
Abstract
The present invention discloses an NMF (Non-negative Matrix Factorization)-based three-dimensional model classification method. The method comprises: step 1, inputting a three-dimensional model set, and performing characteristic calculation on the input three-dimensional model set to obtain an NMF initial input matrix and an NMF initial classification result of the three-dimensional model set, wherein three-dimensional models in the three-dimensional model set are triangular mesh models that contain three-dimensional coordinates of mesh points and trigonometric relationships of the mesh points; step 2, performing visualization presentation on the three-dimensional model set, wherein visualization refers to providing a visualization interface window for a user to perform a follow-up interaction operation; and step 3, designing an interactive NMF method oriented to the three-dimensional model set, and according to an energy equation, implementing three-dimensional model set dynamic classification based on user drive, wherein the interactive NMF method is separately operation splitting and combination methods.
Description
Technical field
The present invention relates to Computer Image Processing and computer graphics techniques field, particularly based on the three-dimensional model sorting technique of NMF (Non-negative Matrix Factorization, NonnegtiveMatrixFactorization).
Background technology
In recent years, along with the raising day by day of powerful modeling software and three-dimensional acquisition equipment availability, digital geometry model resource is increasing fast, the scale of resources bank becomes content that is huge all the more, that simultaneously comprise and also becomes abundant (as Trimble/Google3Dwarehouse all the more, all contain tens0000 digital geometry models even up to a million in the databases such as Turbosquid), these abundant digital geometry resources of rationally recycling contain huge using value.
An important prerequisite for the effective huge profit of resource should be the tissue typing rationalized Models Sets, so that user better understands and uses the model comprised in storehouse.Traditional many employings of Models Sets tissue classification procedure have supervision, semi-supervised or unsupervised machine learning method has come, but these methods are mainly with centered by computing machine, only bottom operation is carried out to storehouse inner model---namely according to low-level feature, taxonomic organization is carried out to Models Sets, user seldom participates in assorting process, and classification results is seldom got information about, this just makes user be difficult to histological structure's situation of profound understanding Models Sets, therefore causes effective huge profit of resource with being obstructed.Especially in Models Sets sharp increase, comprise kind overall background various all the more under, adopt above-mentioned classic method will more and more be difficult to competent related work.
The Models Sets tissue typing mode having more prospect should be make user and computer cooperative design also play respective advantage respectively, under the driving of user view, dynamic cataloging tissue is carried out to Models Sets, and classification results can be presented to user clearly, so that user is when intuitively understanding Models Sets, by obtaining its resource requirement based on the heuristic search browsed.But these class methods are still faced with two large problems: take which kind of method to represent the representation vector of three-dimensional model and how to come tri-vector is carried out to visual display and classified alternately.
Summary of the invention
Goal of the invention: technical matters to be solved by this invention is for the deficiencies in the prior art, provides a kind of three-dimensional model sorting technique based on NMF.
Technical scheme: in order to solve the problems of the technologies described above, the invention discloses a kind of three-dimensional model sorting technique based on NMF, the method carries out semisupervised classification to three-dimensional model collection, comprises the following steps:
Step 1, input three-dimensional model collection, carry out to the three-dimensional model collection of input the NMF preliminary classification result that feature calculation obtains NMF initial input matrix and three-dimensional model collection, the three-dimensional model that wherein three-dimensional model is concentrated is the triangle grid model comprising the three-dimensional coordinate of net point and the triangle relation of net point;
Step 2, carries out visual presenting to three-dimensional model collection, so that user browse data collection and subsequent modification, and the wherein visual operation being to provide visualization interface window and supplying user follow-up; The present invention is based on the t-SNE visualization technique of existing classification annotation information to advanced person to improve, and browse Models Sets and follow-up operation for ease of user, design the visual analyzing prototype system that achieves for three-dimensional model collection, this is also that first passage provides visual and adds operation and is incorporated into three-dimensional model classification;
Step 3, design surface can NMF method alternately to three-dimensional model collection, according to energy equation, realizes the three-dimensional model collection dynamic cataloging driven based on user, wherein can NMF method be respectively and operates Abruption and mergence method alternately.
The geological information that computing machine can comprise from the angle of signature analysis to understand model is utilized in step 1 of the present invention, and adopt simple effective method to distinguish similarity and class inherited in Models Sets class, and then complete stable the presorting of Models Sets, concrete steps are:
Step 1-1, builds word bag model BOW (the Bagofwordsmodel word bag model) feature of three-dimensional model collection, obtains NMF initial input matrix;
Step 1-2, utilizes NMF to complete calculating of presorting to three-dimensional model collection, obtains the preliminary classification result of three-dimensional model collection.
Step 1-1 comprises the steps:
Step 1-1-1, three-dimensional model collection HKS calculates: utilize multi-scale diffusion thermonuclear HKS (HeatKernelSignature) method to carry out feature calculation to each net point of input three-dimensional model collection, obtain HKS descriptor, to represent the local feature information of three-dimensional model collection;
Step 1-1-2, vector quantization: obtain corresponding word list (geometricwords) by the HKS descriptor computation of k-means (hard clustering algorithm) cluster three-dimensional model collection;
Step 1-1-3, the probability distribution that the corresponding word list of each three-dimensional model concentrated by three-dimensional model is built by statistics, obtain the word bag model BOW feature of each three-dimensional model, utilize the BOW feature calculated to obtain the input matrix V of NMF, i.e. three-dimensional model collection eigenmatrix simultaneously.
One is developed towards the Visualized Analysis System of three-dimensional model collection in step 2 of the present invention, can for user provide in visual space clearly visual presentation modes to show the hoc scenario of Models Sets, and then user and machine can be made better to cooperate, utilize respective distinctive ability, go to obtain more efficiently classification results, specifically comprise the following steps:
Step 2-1, completes the conversion of word bag model BOW feature from higher dimensional space to two dimensional surface of each three-dimensional model by t-SNE method (t-DistributedStochasticNeighborEmbeddingt distribution random neighbor embedded mobile GIS);
Step 2-2, add the visualization interface of three-dimensional model collection, t-SNE is transformed projection result to be shown thus the visual display completing input three-dimensional model collection, utilize NMF preliminary classification result to carry out correspondence markings to the subpoint of each model simultaneously, complete visual display, for follow-up operation, visualization interface window comprises assistant browsing window and viewable area, and wherein the three-dimensional model of assistant browsing window area and the display point of viewable area carry out one_to_one corresponding display.
Step 3 comprises the following steps:
Step 3-1, design surface can NMF energy equation alternately to three-dimensional model collection, and the minimum value asking for this expression formula obtains optimum solution, can NMF energy equation be alternately:
Wherein V is the input matrix of the NMF that step 1-1-3 obtains, i.e. three-dimensional model collection eigenmatrix; W, H are respectively the center matrix got required by NMF and refer to matrix, and L is the R-matrix relative to W; M
wfor the diagonal matrix of regulating parameter, matrix M
wvalue on diagonal line is between 0 ~ 1, and α is regulating parameter, and regulate the weight in equation result convergence direction, its scope is between 0 ~ 1;
Wherein the present invention carrys out the classification direction convergence of driving N MF optimum solution towards user's request mainly through amendment L R-matrix, and for this reason by amendment L R-matrix, the present invention devises two kinds of operations: classify and merge.
Step 3-2, divides two generic operations by three-dimensional model collection categories combination and classification and dynamically changes three-dimensional model collection:
Three-dimensional model collection categories combination: after the representative three-dimensional model of user's specified three-dimensional Models Sets classification, its column information in Current central matrix W is located according to this classification represented belonging to three-dimensional model, in union operation process, new column information is obtained according to specified three-dimensional model, the column information do not merged in former center matrix W is directly stored in matrix L, the newer cluster centre need of trying to achieve being merged three-dimensional model joins in matrix L;
Three-dimensional model collection classification divides: user creates brand-new classification on existing classification results basis, namely user is by browsing the classification determining required division, and select correspondingly to represent the new classified information of obtaining three-dimensional model, when dividing, original basis of classification increases new cluster centre, instruct NMF to calculate, R-matrix L increases the column information representing new cluster centre on original center matrix W basis.
The wherein selected model of Abruption and mergence operation of the present invention, based on the information that visualization interface provides, user selects oneself will select the Models Sets of Abruption and mergence very easily, by iteration for several times, finally obtains optimum.
Beneficial effect: category of model method advantage of the present invention is: one of research method enliven data mining and machine learning field the most NMF first time is incorporated into three-dimensional model classification field.In addition, utilize NMF classification acquired results information, the visual tactful t-SNE of the present invention to advanced person improves, and makes the visual display of its classification more clear and intuitive, follow-up user operation of being more convenient for.Finally, present result in conjunction with the visual of three-dimensional model collection, propose a kind of semi-supervised NMF method, make it can instruct the amendment of existing Non-negative Matrix Factorization result, and then realize directly perceived, the dynamically change in real time to Models Sets classification results.
Accompanying drawing explanation
To do the present invention below in conjunction with the drawings and specific embodiments and further illustrate, above-mentioned and/or otherwise advantage of the present invention will become apparent.
Fig. 1 is main flow figure of the present invention.
Fig. 2 is embodiment 1 three-dimensional model BOF calculation flow chart.
Fig. 3 is the result schematic diagram of presorting of embodiment input model collection, is corresponding classification results in each square frame.
To be that t-SNE is visual present schematic diagram to Fig. 4.
Fig. 5 is that embodiment carries out merging schematic diagram.
Fig. 6 is that embodiment carries out division schematic diagram.
Fig. 7 is final classification results schematic diagram, is corresponding classification results in each square frame.
Embodiment
The present invention, to the three-dimensional model collection of input, first calculates its three-dimensional model HKS feature, then builds three-dimensional model BOF by vector quantization, finally calculate the input matrix V of NMF, and completes preliminary NMF to input three-dimensional model collection and presort; Then utilize t-SNE algorithm complete three-dimensional model high dimensional feature to low dimensional feature projection transform and visual presenting facilitate user's subsequent operation, finally devise semi-supervised NMF algorithm and iteration is carried out to result of presorting, finally under visual interface, obtain satisfied three-dimensional model classification results by Abruption and mergence two generic operation.
As shown in Figure 1, the present invention includes following steps:
Step 1, input three-dimensional model collection, carry out to the three-dimensional model collection of input the NMF preliminary classification result that feature calculation obtains NMF initial input matrix and three-dimensional model collection, the three-dimensional model that wherein three-dimensional model is concentrated is the triangle grid model comprising the three-dimensional coordinate of net point and the triangle relation of net point;
Step 2, carries out visual presenting to three-dimensional model collection, so that user browse data collection and subsequent modification, and the wherein visual operation being to provide visualization interface window and supplying user follow-up; The present invention is based on the t-SNE visualization technique of existing classification annotation information to advanced person to improve, and browse Models Sets and follow-up operation for ease of user, design the visual analyzing prototype system that achieves for three-dimensional model collection, this is also that first passage provides visual and adds operation and is incorporated into three-dimensional model classification;
Step 3, design surface can NMF method alternately to three-dimensional model collection, and proposes corresponding accounting equation, realizes the three-dimensional model collection dynamic cataloging driven based on user, wherein NMF method can comprise two kinds of operation Abruption and mergence methods alternately, operate based on visualization interface.
The geological information that computing machine can comprise from the angle of signature analysis to understand model is utilized in step 1 of the present invention, and adopt simple effective method to distinguish similarity and class inherited in Models Sets class, and then complete stable the presorting of Models Sets, concrete steps are:
Step 1-1, builds word bag model BOW (the Bagofwordsmodel word bag model) feature of three-dimensional model collection, obtains NMF initial input matrix;
Step 1-2, utilizes NMF to complete calculating of presorting to three-dimensional model collection, obtains the preliminary classification result of three-dimensional model collection.
Step 1-1 comprises the steps:
Step 1-1-1, three-dimensional model collection HKS calculates: utilize multi-scale diffusion thermonuclear HKS (HeatKernelSignature) method to carry out feature calculation to each net point of input three-dimensional model collection, obtain HKS descriptor, to represent the local feature information of three-dimensional model collection;
Step 1-1-2, vector quantization: obtain corresponding word list geometricwords by the HKS descriptor computation of k-means cluster three-dimensional model collection;
Step 1-1-3, the probability distribution that the corresponding word list of each three-dimensional model concentrated by three-dimensional model is built by statistics, obtain the word bag model BOW feature of each three-dimensional model, utilize the BOW feature calculated to obtain the input matrix V of NMF, i.e. three-dimensional model collection eigenmatrix simultaneously.
One is developed towards the Visualized Analysis System of three-dimensional model collection in step 2 of the present invention, can for user provide in visual space clearly visual presentation modes to show the hoc scenario of Models Sets, and then user and machine can be made better to cooperate, utilize respective distinctive ability, go to obtain more efficiently classification results, specifically comprise the following steps:
Step 2-1, completes the conversion of word bag model BOW feature from higher dimensional space to two dimensional surface of each three-dimensional model by t-SNE (t-DistributedStochasticNeighborEmbedding);
Step 2-2, add the visualization interface of three-dimensional model collection, T-SNE is transformed projection result to be shown thus the visual display completing input three-dimensional model collection, utilize NMF preliminary classification result to carry out not isolabeling to the subpoint of each model simultaneously, complete visual display, for follow-up operation, visualization interface window comprises assistant browsing window and viewable area, and wherein the three-dimensional model of assistant browsing window area and the display point of viewable area carry out one_to_one corresponding display.
Step 3 comprises the following steps:
Step 3-1, design surface can NMF energy equation alternately to three-dimensional model collection, and the minimum value asking for this expression formula obtains optimum solution, can NMF energy equation be alternately:
Wherein V is the input matrix of the NMF that step 1-1-3 obtains, i.e. three-dimensional model collection eigenmatrix; W, H are respectively the center matrix got required by NMF and refer to matrix, and L is the R-matrix relative to W; M
wfor the diagonal matrix of regulating parameter, matrix M
wvalue on diagonal line is between 0 ~ 1, and α is regulating parameter, and its scope is between 0 ~ 1;
Wherein the present invention carrys out the direction convergence of driving N MF optimum solution towards user mainly through amendment L R-matrix, and for this reason by amendment L R-matrix, the present invention devises two kinds of operations: classify and merge.
Step 3-2, divides two generic operations by three-dimensional model collection categories combination and classification and dynamically changes three-dimensional model collection:
Three-dimensional model collection categories combination: after the representative three-dimensional model of user's specified three-dimensional Models Sets classification, its column information in Current central matrix W is located according to this classification represented belonging to three-dimensional model, in union operation process, new column information is obtained according to specified three-dimensional model, the column information do not merged in former center matrix W is directly stored in matrix L, the newer cluster centre need of trying to achieve being merged three-dimensional model joins in matrix L;
Three-dimensional model collection classification divides: user creates brand-new classification on existing classification results basis, namely user is by browsing the classification determining required division, and select correspondingly to represent the new classified information of obtaining three-dimensional model, when dividing, original basis of classification increases new cluster centre, instruct NMF to carry out brand-new calculating, R-matrix L increases the column information representing new cluster centre on original center matrix W basis.
Embodiment 1
As shown in Figure 2, in the present invention: for input three-dimensional model collection, first the three-dimensional model of input as shown in (a) in Fig. 2, then the three-dimensional model HKS thermal map that feature calculation obtains as shown in (b) in Fig. 2 is carried out, finally as Suo Shi (c) in Fig. 2, pass through vector quantization, the distribution calculated on vector space obtains the BOF feature of each three-dimensional model.
Embodiment 2
According to embodiment, each step of the present invention is described below.
Step (1), carries out quick, stable presorting, to instruct subsequent visual display and user operation according to the isomery Models Sets of input;
The Models Sets of the present embodiment process is the isomery three-dimensional model collection of size normalizing, in order to support follow-up visual and mutual NMF operation, needing the BOF of structure model and completing to Models Sets operation of presorting.The present embodiment adopts document BRONSTEINA.M., BRONSTEINM.M., GUIBASL.J., OVSJANIKOVM.:Shapegoogle:Geometricwordsandexpressionsfor invariantshaperetrieval.ACMTransactionsonGraphics (TOG) 30,1 (2011), 1.2,4 methods building three-dimensional model BOF based on HKS proposed.The HKS feature of the employing of the method has the features such as multiple dimensioned, rotational invariance, and the foundation of BOF model can as the input of NMF matrix.Two process: BOF foundation and NMF is mainly divided into presort.
Step (11), three-dimensional model collection BOF sets up:
BOF can be good at representing three-dimensional model, is therefore used for building input non-negative sample matrix V, mainly comprises model HKS feature calculation, and vector quantization and BOF set up three steps.
Step (111) three-dimensional model HKS calculates.
The each net point of multi-scale diffusion thermonuclear (HKS) method to each model of input three-dimensional model collection is utilized to carry out feature calculation, to represent the local feature information of each model; Suppose when time t=0, at the heat of a given unit of three-dimensional model gridding point x place, the heat of other points of grid is all 0, then allows heat freely spread on grid, then k
t(x, y) can be interpreted as the amount of heat at time t time point y.The k that HKS descriptor is made up of the t of different size
t(x, x) Sequence composition, i.e. { k
t1(x, x), k
t2(x, x) ..., k
m(x, x), } .HKS descriptor not only has Analysis On Multi-scale Features, and very robust.
Step (112) vector quantization.
Through above-mentioned calculating, obtain the HKS descriptor feature of enormous amount, be necessary that vector quantization builds the word folder of BOF model, so calculate corresponding word list P={p1, p2 by the HKS descriptor feature of each model of k-means cluster ..., pn};
The foundation of step (113) BOF.
Built the probability distribution of the word list of each three-dimensional model by statistics, obtain the BOW of each model, the process of establishing of this process and picture BOW is basically identical.
Detailed process is as follows:
When after the HKS descriptor characteristic sum word list P of all net points calculating model, just start to calculate this model each point probability distribution situation θ (x) in word list={ θ
1(x) ..., θ
v(x) }, what the present invention adopted is gaussian probability distribution, and computing formula is as follows:
Wherein C (x) is regulating parameter, in order to do normalization to formula value, namely || and θ (x) ||
1=1, wherein θ represents probability distribution, the HKS descriptor that p (x) is this point, and pi is a wherein word of the word list that step (112) is calculated.σ is formula variance, and its value decides most suitable value by word list.
Step (12) NMF presorts.
The BOF being completed each model of three-dimensional model collection by step (21) is set up, and also just obtains the input matrix of NMF, finally utilizes NMF to complete calculating of presorting to Models Sets.At this, be first necessary lower NMF is described.
A given size is the non-negative original sample matrix V of n × m, also namely corresponds to input matrix of the present invention, and the n that each row correspond to each sample in m sample ties up nonegative elgenvector, and wherein m is determined by sample number size, and n is determined by the word number of word list.The object of NMF algorithm is exactly find 2 new matrix W and H to carry out approximate original sample matrix V:
In formula, W, H are respectively basis matrix and matrix of coefficients, and in the present invention, W, H are the center matrix got required by NMF and refer to matrix; W is n × r dimension, and H is r × m dimension, and r is the dimension after dimensionality reduction, i.e. the number of base vector, and its selection should guarantee (n+m) r<nm, and the W drawn thus, H just can be considered to the compressed format of data in V.Can obtain W and H by a series of iterative algorithm, wherein each row of W illustrate the cluster of local correlation composition all to a certain extent.In addition this algorithm calculates based on simple iterative process, not time-consuming, and fast convergence rate, and result of calculation is stablized, therefore very applicable for large scale data classification process.
So presort based on the Models Sets of NMF, following form can be expressed as:
p(features|model)＝∑
topicp(features|topic)×p(topic|model)，
The wherein characteristic probability distribution situation of each model in p (features │ model) library representation, p (topic │ model) represents that model carrys out choosing a topic with certain probability, p (features │ topic) to represent in this theme with certain probability to select certain feature, therefore required that theme be final classification situation.After the characteristic probability distribution situation of Models Sets is represented with matrix, just the problems referred to above can be converted into NMF and solve.Wherein input non-negative sample matrix V in p (features │ model) corresponding NMF, basis matrix W in p (features │ topic) corresponding NMF, matrix of coefficients H in p (topic │ model) corresponding NMF, in the present invention, W, H are the center matrix got required by NMF and refer to matrix; .
Based on above-mentioned expression, the BOF vector of each model of three-dimensional model collection is exactly the row of matrix V, also namely determines the input non-negative sample matrix V of Models Sets.The process of presorting is completed, as shown in Figure 3 eventually through NMF.
Step (2), carries out visual presenting to existing category of model result, so that user browse data collection and subsequent modification; The present invention is based on the t-SNE visualization technique of existing classification annotation information to advanced person to improve, and browse Models Sets and follow-up operation for ease of user, design the visual analyzing prototype system achieved for three-dimensional model collection; Mainly be divided into the visual display of t-SNE and auxiliary area models show two parts.
The visual display of step (21) t-SNE
Completed the conversion BOW feature of each three-dimensional model completed from higher dimensional space to two dimensional surface by t-SNE, complete the visual display of input three-dimensional model collection.T-SNE under this is necessary first to introduce.
Hinton etc. are at HintonGE, RoweisST.Stochasticneighborembedding [C] //Advancesinneuralinformationprocessingsystems.2002:833-84 0. proposes to be called that random neighbor embeds (stochasticneighborembedding, SNE) visual Dimension Reduction Analysis method, Euclidean distance between high dimensional data is converted into probability expression-form, and its cost functional builds criterion calls subspace and has identical form of probability with the former input space.Laurens etc. are at VanderMaatenL, HintonG.Visualizingdatausingt-SNE [J] .JournalofMachineLearningResearch, 2008,9 (2579-2605): 85. propose improvement t distribute SNE, t-SNE method adopts the conditional probability form had in the alternative SNE of symmetric joint probability expression, simultaneously, in order to solve the problem of data point " crowded " in SNE method, higher dimensional space adopts gaussian probability distribution, and lower dimensional space employing degree of freedom is the t distribution of 1.This process reduces the attractive force in the lower dimensional space of simulation between mapping point.
This algorithm discloses the classification characteristics of data inherence, and have expressed the similarity between data intuitively by data visualization.The present invention makes full use of t-SNE can disclose the feature of the classification situation of data inherence to instruct the optimum configurations of NMF preliminary classification classification.By concentrating each model to carry out BOF feature calculation to the three-dimensional model of input before, therefore this High Dimensional Data Set can be carried out visual analyzing as the input of t-SNE, and result is presented on two-dimensional screen, as shown in (a) in Fig. 4, user is by this visual display, the rough classification situation in storehouse can be got information about, and the classification situation that can be observed is as priori, instruct the optimum configurations of NMF preliminary classification classification, namely user can arrange the initial category value of NMF calculating with viewed class label.Utilize the method not only can play visual advantage, and the blindness of unsupervised segmentation can be avoided.
By presorting to Models Sets before, therefore inputting data and will have corresponding class label information, in visual procedure for displaying, should as much as possible near similar label data, away from foreign peoples's label data.
Models Sets data after being presorted by NMF are defined as
wherein
represent i-th sample of c class, the sample number of the total classification number of sample to be C, Ni be the i-th class, and N=N1+N2+ ... + NC is total number of samples.After introducing classification information, former space sample similarity be may be defined as following form:
Wherein, x
irepresent the proper vector of i-th model, wherein i, j, k, l, m are also like this, and their dimension is determined by the model sample number inputted, c
irepresent sample x
iaffiliated class label information, λ is the variance parameter of corresponding Gaussian function.Through above-mentioned calculating, namely above formula maintains the symmetry of probability distribution matrix, and similarity probability between homogeneous data and between heterogeneous data and be all 1.
Similar, the Sample Similarity of subspace is calculated by following formula:
Wherein y
ifor sample x
iat the vectorial expression-form of projection subspace.Obtaining similarity p
ijand q
ijafter, for keeping the similarity between same class model as much as possible and reducing the similarity between foreign peoples's sample pattern, the present invention can obtain objective cost function by Kullback-Leibler divergence:
Namely the KL divergence of similar sample is farthest reduced and the KL divergence of increase foreign peoples sample.Utilize gradient descent method to minimize the KL divergence of all data points, obtain best simulation point, just can complete the visual display of input data model collection, as shown in (b) in Fig. 4.
Step (22) auxiliary area models show;
Add the assistant browsing window of model, and the display point of auxiliary area model and viewable area carries out one_to_one corresponding display.Therefore and be not suitable for the visual display of three-dimensional model collection utilize t-SNE to carry out visual display, it just represents the distribution situation of model by different color dot, user intuitively can not understand the particular content of the model representated by each point.The conveniently understanding model hoc scenario of user's profound level, and subsequent operation, original display mode improves, and namely adds the assistant browsing window of model.
User not only can view the entirety classification situation of Models Sets intuitively according to original visualization result, also can be chosen at auxiliary area by region and browse corresponding concrete model situation.In addition, the display point of auxiliary area model and viewable area can also be carried out one_to_one corresponding display, namely choose corresponding model at auxiliary area, just can show corresponding some situation in viewable area.
Step (3), on the basis of existing analytic system, again revises calculating by the semi-supervised NMF of design, and then realizes the Models Sets dynamic cataloging based on user's driving.The method can support that user carries out the operation such as merging, division of Models Sets classification on visual basis.Specifically comprise the semi-supervised NMF design of two large divisions and operation:
Step (31) design surface to three-dimensional model collection can NMF method alternately.
After NMF presorts to Models Sets, matrix centered by the W that decomposition obtains, each row of W represent a cluster centre; H is for referring to matrix, and each row of H represent the generic attribute information of each sample pattern.Carrying out Models Sets merging with splitting operation process, produce new cluster result.Be merge or divide all to provide new reference cluster centre for system, therefore on this basis, new cluster calculation can be carried out and obtain corresponding new classification results.When being abstracted in NMF calculating, the characteristic information just can seen as by interaction models affects the center matrix W of gained of presorting, and re-starts NMF decomposition computation based on the R-matrix of the new cluster centre of gained, to obtain brand-new cluster result.
Wherein, can NMF mathematic(al) representation be alternately:
Wherein V is the input matrix of the NMF that step (12) obtains, i.e. three-dimensional model collection eigenmatrix; The center matrix that W, H get required by NMF with refer to matrix; L is the R-matrix relative to W; M
wfor the diagonal matrix of regulating parameter, the value on diagonal of a matrix is between 0 ~ 1, and α is also regulating parameter, and wherein scope is between 0 ~ 1.The cluster centre that the implication of this objective function can be specified with user for final cluster result that hope obtains is the most close.
Step (32) is based on the operation of semi-supervised NMF.
Under NMF support that can be mutual, devise the dynamic change that the continuous iteration of merge sort two generic operation completes correlation model collection in real time;
Wherein three-dimensional model collection categories combination: the present invention supports that user passes through storage optimization, carries out union operation to respective classes, and Fig. 5 is the schematic diagram of a merging process.User can specify the required representative model merged to carry out corresponding union operation by browsing, as rectangle frame in (a) in Fig. 5 the model that identifies just for the representative model of classification need be merged, note, in schematic diagram of the present invention, represent different models by difformity.When after other representative model of user's specified class, its concrete column information in Current central matrix W can be determined according to generic attribute information.When carrying out union operation process, mainly wishing to obtain new cluster centre according to specified model case, namely regulating and revising R-matrix L.The column information do not merged in former center matrix W, in merging process, is directly stored in matrix L by the present invention, then needs the new cluster centre of pooled model to join in L just can by what try to achieve.The present invention asks for the new column information value in R-matrix L according to the model feature information that user specifies, namely new cluster centre is set, to merge two class models, its detailed process is as follows: first by the feature situation according to user's designated model, KNN (k nearest-neighbors) algorithm is utilized to ask for its model neighbours the most contiguous in former class, then ask for the center vector (as Suo Shi (b) in Fig. 5) of these Model B OF features, add (as Suo Shi (c) in Fig. 5) in R-matrix L to as the cluster centre initial value after merging.In addition, after carrying out KNN calculating, model is remained in addition in specified class, so demand gets the center vector (as Suo Shi (b) in Fig. 5) of residue model BOF feature, adds (as Suo Shi (c) in Fig. 5) in R-matrix L to as another cluster centre initial value after merging.After above-mentioned setting, next, just brand-new cluster result can be tried to achieve according to the guidance of user by the objective function of setting.For the final classification results of (a) gained in Fig. 5 as shown in (d) in Fig. 5.
Wherein three-dimensional model collection classification division: the present invention also supports that user passes through storage optimization, carries out splitting operation to respective classes, and Fig. 6 is the schematic diagram of a fission process.Division refers to that user creates brand-new classification on existing classification results basis, and namely user is by browsing the classification determining required division, and clicks corresponding representative model, as in (a) in Fig. 6 the model that identifies, obtain new classified information.The present invention still completes corresponding operating by reference to matrix L, when dividing, being equivalent to increase new cluster centre on original basis of classification, and then instructing NMF to carry out brand-new calculating.Therefore R-matrix L also will be on original center matrix W basis, increase the column information representing new cluster centre and just can.The present invention asks for the new column information value of R-matrix L according to the model feature information that user specifies, to divide model I, detailed process is as follows: first by the feature situation according to user's designated model, KNN algorithm is utilized to ask for its model neighbours the most contiguous in former class, then ask for the center vector (as Suo Shi (b) in Fig. 6) of these Model B OF features, add (as Suo Shi (c) in Fig. 6) in R-matrix L to as the cluster centre initial value after division.In addition, after carrying out KNN calculating, also demand gets the center vector (as Suo Shi (b) in Fig. 6) of such residue model BOF feature, revises the column vector information (as Suo Shi (c) in Fig. 6) of R-matrix L correspondence position.After above-mentioned setting, next, just brand-new cluster result can be tried to achieve according to the guidance of user by the objective function of setting.For the final classification results of (a) gained in Fig. 6 as shown in (d) in Fig. 6.
Eventually pass multi-pass operations, obtain final three-dimensional model collection classification results as shown in Figure 7.
The invention provides a kind of three-dimensional model sorting technique based on NMF; the method and access of this technical scheme of specific implementation is all few; the above is only the preferred embodiment of the present invention; should be understood that; for those skilled in the art; under the premise without departing from the principles of the invention, can also make some improvements and modifications, these improvements and modifications also should be considered as protection scope of the present invention.The all available prior art of each ingredient not clear and definite in the present embodiment is realized.
Claims (5)
1., based on a three-dimensional model sorting technique of NMF, it is characterized in that, comprise the following steps:
Step 1, input three-dimensional model collection, carry out to the three-dimensional model collection of input the NMF preliminary classification result that feature calculation obtains NMF initial input matrix and three-dimensional model collection, the three-dimensional model that wherein three-dimensional model is concentrated is the triangle grid model comprising the three-dimensional coordinate of net point and the triangle relation of net point;
Step 2, carries out visual presenting to three-dimensional model collection, and wherein visual referring to provides visualization interface window for the follow-up interactive operation of user;
Step 3, design surface can NMF method alternately to three-dimensional model collection, according to energy equation, realizes the three-dimensional model collection dynamic cataloging driven based on user, wherein can NMF method be respectively and operates Abruption and mergence method alternately.
2. a kind of three-dimensional model sorting technique based on NMF according to claim 1, it is characterized in that, step 1 comprises the steps:
Step 1-1, builds the BOW word bag model feature of three-dimensional model collection, obtains NMF initial input matrix;
Step 1-2, utilizes NMF to complete calculating of presorting to three-dimensional model collection, obtains the preliminary classification result of three-dimensional model collection.
3. a kind of three-dimensional model sorting technique based on NMF according to claim 2, it is characterized in that, step 1-1 comprises the steps:
Step 1-1-1, three-dimensional model collection HKS calculate: utilize multi-scale diffusion thermonuclear HKS method to carry out feature calculation to each net point of input three-dimensional model collection, obtain HKS descriptor, to represent the local feature information of three-dimensional model collection;
Step 1-1-2, vector quantization: obtain corresponding word list by the HKS descriptor computation of k-means cluster three-dimensional model collection;
Step 1-1-3, the probability distribution that the corresponding word list of each three-dimensional model concentrated by three-dimensional model is built by statistics, obtain the word bag model BOW feature of each three-dimensional model, utilize the BOW feature calculated to obtain the input matrix V of NMF, i.e. three-dimensional model collection eigenmatrix simultaneously.
4. a kind of three-dimensional model sorting technique based on NMF according to claim 3, it is characterized in that, step 2 comprises the following steps:
Step 2-1, completes the conversion of word bag model BOW feature from higher dimensional space to two dimensional surface of each three-dimensional model by t-SNE;
Step 2-2, add the visualization interface of three-dimensional model collection, t-SNE is transformed projection result to be shown thus the visual display completing input three-dimensional model collection, utilize NMF preliminary classification result to carry out correspondence markings to the subpoint of each model simultaneously, complete visual display, for follow-up interactive operation, visualization interface window comprises assistant browsing window and viewable area, and wherein the three-dimensional model of assistant browsing window area and the display point of viewable area carry out one_to_one corresponding display.
5. a kind of three-dimensional model sorting technique based on NMF according to claim 4, it is characterized in that, step 3 comprises the following steps:
Step 3-1, design surface can NMF energy equation alternately to three-dimensional model collection, and the minimum value asking for this expression formula obtains optimum solution, can NMF energy equation be alternately:
equation passes through input parameter V, L, M
wask for best W, H value, wherein min
w, H>=0represent the W under asking for formula value minimizes, H value, V is the input matrix of the NMF that step 1-1-3 obtains, i.e. three-dimensional model collection eigenmatrix; W, H are respectively the center matrix got required by NMF and refer to matrix, and L is the R-matrix relative to W; M
wfor the diagonal matrix of regulating parameter, matrix M
wvalue on diagonal line is between 0 ~ 1, and α is regulating parameter, and regulate the weight in equation result convergence direction, its scope is between 0 ~ 1;
Step 3-2, divides two generic operations by three-dimensional model collection categories combination and classification and dynamically changes three-dimensional model collection:
Three-dimensional model collection categories combination: after the representative three-dimensional model of user's specified three-dimensional Models Sets classification, its column information in Current central matrix W is located according to this classification represented belonging to three-dimensional model, in union operation process, new column information is obtained according to specified three-dimensional model, the column information do not merged in former center matrix W is directly stored in matrix L, the newer cluster centre need of trying to achieve being merged three-dimensional model joins in matrix L;
Three-dimensional model collection classification divides: user creates brand-new classification on existing classification results basis, namely user is by browsing the classification determining required division, and select correspondingly to represent the new classified information of obtaining three-dimensional model, when dividing, original basis of classification increases new cluster centre, instruct NMF to calculate, R-matrix L increases the column information representing new cluster centre on original center matrix W basis.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510695635.2A CN105354593B (en) | 2015-10-22 | 2015-10-22 | A kind of threedimensional model sorting technique based on NMF |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510695635.2A CN105354593B (en) | 2015-10-22 | 2015-10-22 | A kind of threedimensional model sorting technique based on NMF |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN105354593A true CN105354593A (en) | 2016-02-24 |
| CN105354593B CN105354593B (en) | 2018-10-30 |
Family
ID=55330560
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510695635.2A Active CN105354593B (en) | 2015-10-22 | 2015-10-22 | A kind of threedimensional model sorting technique based on NMF |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105354593B (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106650747A (en) * | 2016-11-16 | 2017-05-10 | 佛山科学技术学院 | 3D model feature extraction method based on compressed sensing |
| CN106878073A (en) * | 2017-02-14 | 2017-06-20 | 南京邮电大学 | Network multimedia business semisupervised classification method based on t Distribution Mixed Models |
| CN107909099A (en) * | 2017-11-10 | 2018-04-13 | 佛山科学技术学院 | A kind of threedimensional model identification and search method based on thermonuclear |
| CN109034270A (en) * | 2018-08-23 | 2018-12-18 | 西安交通大学 | A kind of visualization feature selection method based on the classification Non-negative Matrix Factorization of failure two |
| CN109492663A (en) * | 2018-09-28 | 2019-03-19 | 西安交通大学 | More characteristic of division selection methods based on Non-negative Matrix Factorization and visualization technique |
| CN109829505A (en) * | 2019-02-15 | 2019-05-31 | 北京工业大学 | A kind of threedimensional model classification method based on various features descriptor |
| CN110704624A (en) * | 2019-09-30 | 2020-01-17 | 武汉大学 | Geographic information service metadata text multi-level multi-label classification method |
| CN113393145A (en) * | 2021-06-25 | 2021-09-14 | 广东利元亨智能装备股份有限公司 | Model similarity obtaining method and device, electronic equipment and storage medium |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101201845A (en) * | 2007-09-24 | 2008-06-18 | 覃征 | Method for searching three-dimensional model based on virgula point-set delamination helix information |
| CN101320487A (en) * | 2008-07-07 | 2008-12-10 | 中国科学院计算技术研究所 | Scene pretreatment method for fire disaster simulation |
| KR101221757B1 (en) * | 2011-01-25 | 2013-01-11 | 세종대학교산학협력단 | Method for generating of 3D rule model using decision tree in OLAP and system thereof |
| CN103021029A (en) * | 2013-01-18 | 2013-04-03 | 南京大学 | Automatic labeling method for three-dimensional model component categories |
-
2015
- 2015-10-22 CN CN201510695635.2A patent/CN105354593B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101201845A (en) * | 2007-09-24 | 2008-06-18 | 覃征 | Method for searching three-dimensional model based on virgula point-set delamination helix information |
| CN101320487A (en) * | 2008-07-07 | 2008-12-10 | 中国科学院计算技术研究所 | Scene pretreatment method for fire disaster simulation |
| KR101221757B1 (en) * | 2011-01-25 | 2013-01-11 | 세종대학교산학협력단 | Method for generating of 3D rule model using decision tree in OLAP and system thereof |
| CN103021029A (en) * | 2013-01-18 | 2013-04-03 | 南京大学 | Automatic labeling method for three-dimensional model component categories |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106650747A (en) * | 2016-11-16 | 2017-05-10 | 佛山科学技术学院 | 3D model feature extraction method based on compressed sensing |
| CN106650747B (en) * | 2016-11-16 | 2019-08-02 | 佛山科学技术学院 | A kind of compressed sensing based method for extracting characteristics of three-dimensional model |
| CN106878073A (en) * | 2017-02-14 | 2017-06-20 | 南京邮电大学 | Network multimedia business semisupervised classification method based on t Distribution Mixed Models |
| CN106878073B (en) * | 2017-02-14 | 2019-10-08 | 南京邮电大学 | Network multimedia business semisupervised classification method based on t Distribution Mixed Model |
| CN107909099A (en) * | 2017-11-10 | 2018-04-13 | 佛山科学技术学院 | A kind of threedimensional model identification and search method based on thermonuclear |
| CN109034270A (en) * | 2018-08-23 | 2018-12-18 | 西安交通大学 | A kind of visualization feature selection method based on the classification Non-negative Matrix Factorization of failure two |
| CN109492663A (en) * | 2018-09-28 | 2019-03-19 | 西安交通大学 | More characteristic of division selection methods based on Non-negative Matrix Factorization and visualization technique |
| CN109829505A (en) * | 2019-02-15 | 2019-05-31 | 北京工业大学 | A kind of threedimensional model classification method based on various features descriptor |
| CN110704624A (en) * | 2019-09-30 | 2020-01-17 | 武汉大学 | Geographic information service metadata text multi-level multi-label classification method |
| CN113393145A (en) * | 2021-06-25 | 2021-09-14 | 广东利元亨智能装备股份有限公司 | Model similarity obtaining method and device, electronic equipment and storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN105354593B (en) | 2018-10-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105354593A (en) | NMF (Non-negative Matrix Factorization)-based three-dimensional model classification method | |
| Li et al. | A comparison of 3D shape retrieval methods based on a large-scale benchmark supporting multimodal queries | |
| CN103345645B (en) | Commodity image class prediction method towards net purchase platform | |
| CN108875076B (en) | Rapid trademark image retrieval method based on Attention mechanism and convolutional neural network | |
| CN108537269B (en) | Weak interactive object detection deep learning method and system thereof | |
| CN110428428A (en) | A kind of image, semantic dividing method, electronic equipment and readable storage medium storing program for executing | |
| US20140328540A1 (en) | Sketch Segmentation | |
| CN104200240A (en) | Sketch retrieval method based on content adaptive Hash encoding | |
| CN103838864A (en) | Visual saliency and visual phrase combined image retrieval method | |
| CN102663015A (en) | Video semantic labeling method based on characteristics bag models and supervised learning | |
| CN105493078A (en) | Color sketch image searching | |
| CN110210431A (en) | A kind of point cloud classifications method based on cloud semantic tagger and optimization | |
| CN109766822A (en) | Gesture identification method neural network based and system | |
| Fan et al. | Structure completion for facade layouts. | |
| CN106250909A (en) | A kind of based on the image classification method improving visual word bag model | |
| Zhou et al. | Enhance the recognition ability to occlusions and small objects with Robust Faster R-CNN | |
| CN103295032A (en) | Spatial Fisher vector based image classification method | |
| CN103617609A (en) | A k-means nonlinear manifold clustering and representative point selecting method based on a graph theory | |
| Xu et al. | Large-margin multi-view Gaussian process for image classification | |
| CN106971005A (en) | Distributed parallel Text Clustering Method based on MapReduce under a kind of cloud computing environment | |
| Liang et al. | PointFusionNet: Point feature fusion network for 3D point clouds analysis | |
| Mei et al. | Proximity-based k-partitions clustering with ranking for document categorization and analysis | |
| CN103295026A (en) | Spatial local clustering description vector based image classification method | |
| CN104965928A (en) | Chinese character image retrieval method based on shape matching | |
| Wu et al. | Deep texture exemplar extraction based on trimmed T-CNN |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |