CN106157375A - A kind of threedimensional model component categories automatic marking method - Google Patents
A kind of threedimensional model component categories automatic marking method Download PDFInfo
- Publication number
- CN106157375A CN106157375A CN201610530040.6A CN201610530040A CN106157375A CN 106157375 A CN106157375 A CN 106157375A CN 201610530040 A CN201610530040 A CN 201610530040A CN 106157375 A CN106157375 A CN 106157375A
- Authority
- CN
- China
- Prior art keywords
- dough sheet
- model
- feature
- limit
- mark
- 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
- 238000000034 method Methods 0.000 title claims abstract description 80
- 238000012549 training Methods 0.000 claims abstract description 134
- 230000008569 process Effects 0.000 claims abstract description 36
- 238000000638 solvent extraction Methods 0.000 claims abstract description 8
- 230000006870 function Effects 0.000 claims description 28
- 210000002569 neuron Anatomy 0.000 claims description 28
- 230000004913 activation Effects 0.000 claims description 27
- 238000005457 optimization Methods 0.000 claims description 27
- 230000011218 segmentation Effects 0.000 claims description 20
- 239000011159 matrix material Substances 0.000 claims description 12
- 238000013528 artificial neural network Methods 0.000 claims description 9
- 239000012141 concentrate Substances 0.000 claims description 9
- 238000000605 extraction Methods 0.000 claims description 9
- 239000000284 extract Substances 0.000 claims description 8
- 238000013507 mapping Methods 0.000 claims description 8
- 238000009499 grossing Methods 0.000 claims description 6
- 238000013459 approach Methods 0.000 claims description 5
- 238000004364 calculation method Methods 0.000 claims description 5
- 230000008859 change Effects 0.000 claims description 3
- 230000001105 regulatory effect Effects 0.000 claims description 3
- 210000004027 cell Anatomy 0.000 claims description 2
- 238000004458 analytical method Methods 0.000 description 4
- 238000002372 labelling Methods 0.000 description 4
- WSFSSNUMVMOOMR-UHFFFAOYSA-N Formaldehyde Chemical compound O=C WSFSSNUMVMOOMR-UHFFFAOYSA-N 0.000 description 3
- 230000015572 biosynthetic process Effects 0.000 description 3
- 230000006872 improvement Effects 0.000 description 3
- 230000008901 benefit Effects 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 230000005764 inhibitory process Effects 0.000 description 2
- 238000001746 injection moulding Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000007781 pre-processing Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000003786 synthesis reaction Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 239000003086 colorant Substances 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 239000004744 fabric Substances 0.000 description 1
- 239000004615 ingredient Substances 0.000 description 1
- 230000002452 interceptive effect Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 210000005036 nerve Anatomy 0.000 description 1
- 238000003672 processing method Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T19/00—Manipulating 3D models or images for computer graphics
- G06T19/20—Editing of 3D images, e.g. changing shapes or colours, aligning objects or positioning parts
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20072—Graph-based image processing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2219/00—Indexing scheme for manipulating 3D models or images for computer graphics
- G06T2219/004—Annotating, labelling
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- Software Systems (AREA)
- Computer Graphics (AREA)
- Architecture (AREA)
- Computer Hardware Design (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of threedimensional model component categories automatic marking method, comprise the following steps: training process carries out Fast Training according to threedimensional model mark training set, and training obtains the quick marking model for the dough sheet of unknown threedimensional model and Grid Edge are classified and marked.Annotation process then utilizes the quick marking model training dough sheet and the Grid Edge obtained to classify dough sheet and the Grid Edge of object module, obtain dough sheet and the distribution of Grid Edge class probability, build graph model, cut by multi-tag figure to optimize and carry out the smooth of partitioning boundary and optimize, thus realize the fast automatic mark to target three-dimensional component categories.
Description
Technical field
The present invention relates to the processing method of a kind of shape analysis, belong to computer graphics techniques field, specifically
A kind of threedimensional model component categories automatic marking method.
Background technology
The understanding constituting threedimensional model parts is that shape understands and the basis of high-level geometric manipulations.In recent years, research
Full-fledged machine learning techniques is progressively incorporated in 3D shape analysis and research by personnel, the geometric model of data-driven
Associating dividing method the most increasingly receives publicity.Utilize " information transmission " this key concept of data-driven, can be in input
The sample shape preassigned or calculate and needs are analyzed and are set up a kind of dependency between the target shape processed, and can will feel
The information of interest is delivered to target shape from sample shape.Therefore, the shape segmentations of data-driven can produce and conforming have
The parts of semantic tagger, and can effectively support the follow-up geometry such as the synthesis of threedimensional model, example modeling and textures synthesis
Process task.
Whether there are available label for labelling data as input information based in study, 3D shape segmentation can be divided
For: supervised segmentation, non-formaldehyde finishing and semi-supervised segmentation.In general, supervised learning method is by given mark number
According to tending to output with it is intended that the result that is close, thus obtain higher segmentation precision.There is supervision shape segmentations usual
Form turns to a classification problem.Such as document 1:KALOGERAKIS E., HERTZMANN A., SINGH K.:Learning 3d
Segmentation is asked by mesh segmentation and labeling.ACM Trans.Graph.29 (2010), 102:1 102:12.
Topic form turns to the optimization problem of condition random field CRF.Its condition random field CRF model, comprises evaluation dough sheet consistent with label
Property unitary item, and the binary item of adjacent dough sheet tag compliance, its unitary item and binary Xiang Junke according to markup information lead to
Cross the study of JointBoost grader to obtain.Document 2:GUO K., Zou D., and Chen X.3D Mesh Labeling
via Deep Convolutional Neural Networks[J].ACM Transactions on Graphics,2015,
35 (1): 3:1-3:12. obtain a kind of effective and the shape representation of robust by deep neural network from data learning, enter
One step improves segmentation and mark performance.But, the training process of said method is the most time-consuming, as only carried out 6 shapes
During training, training process is accomplished by expending several hours.In order to improve training effectiveness, document 3:XIE Z., XU K., LIU L.,
XIONG Y.:3d shape segmentation and labeling via extreme learning
Machine.Computer Graphics Forum 33,5 (2014). then carry out the classification of dough sheet based on ELM grader, training
Process greatly speeds up, the most only need to be less than 1 minutes when training 6 shapes.But, the method only only accounts for according to instruction
The dough sheet markup information practiced in data produces a dough sheet grader, does not utilize the border letter gone out given in labeled data
Breath, but obtain boundary position substantially by target shape being carried out over-segmentation, then be optimized, its final segmentation result
Quality depend heavilys on the result of over-segmentation.Method disclosed by the invention then by the activation primitive of neuron in amendment ELM,
Adjust classifying quality and the classification time of ELM further, and the boundary information in labeled data is also made full use of.According to
Dough sheet classification annotation probability distribution and the distribution of Grid Edge class probability, build graph model, cuts optimization method essence by multi-tag figure
Adjust and obtain the most smooth and accurate segmentation result.
Summary of the invention
Goal of the invention: the technical problem to be solved is for the deficiencies in the prior art, it is provided that a kind of quick
Threedimensional model component categories automatic marking method, carries out automatically splitting and mark rapidly to threedimensional model for support.
In order to solve above-mentioned technical problem, the invention discloses a kind of threedimensional model component categories automatic marking method, bag
Include following steps:
Step 1, training threedimensional model mark training set, the threedimensional model in training set is parts in same class model and model
Having provided standard mark, during wherein markup information includes threedimensional model, each patch grids is subordinated to the classification mark of component parts
And every Grid Edge is subordinated to the classification mark of boundary edge, by threedimensional model mark training set is carried out Fast Training, obtain
Must be used for target three-dimensional is split the quick dough sheet marking model with mark and quick Grid Edge marking model, and obtain
Obtain threedimensional model dough sheet classification annotation probability distribution and Grid Edge classification annotation probability distribution, and study obtains figure and cuts the flat of model
Sliding item weight；
Step 2, label target threedimensional model: utilize quick dough sheet marking model and quick Grid Edge mark that step 1 obtains
Target three-dimensional is split and mark by injection molding type respectively, it is thus achieved that each dough sheet of target three-dimensional is subordinated to constituting portion
The classification mark of part and every Grid Edge are subordinated to the classification mark of boundary edge, build graph model, cut excellent by multi-tag figure
Change segmentation and the annotation results obtaining smoothing.
Step 1 comprises the following steps:
Threedimensional models all in training set are carried out pretreatment by step 1-1, extract each patch grids dough sheet feature and
The limit feature of every Grid Edge；
Step 1-2, all dough sheet composing training patch grids collection of threedimensional model in training set, use the limit improved
Habit machine method (ELM, extreme learning machine), the dough sheet to each patch grids that training patch grids is concentrated
Feature and standard mark thereof carry out Fast Training, it is thus achieved that quickly dough sheet marking model；
Step 1-3, all Grid Edge composing training Grid Edge collection of threedimensional model in training set, use the limit improved
Habit machine method carries out Fast Training to limit feature and the standard mark thereof of every Grid Edge that training Grid Edge is concentrated, it is thus achieved that quickly
Grid Edge marking model.
Step 1-4, it is thus achieved that threedimensional model dough sheet classification annotation probability distribution and Grid Edge classification annotation probability distribution, and learn
Acquistion cuts the smooth item weight of model to figure.
Described in step 1-1, threedimensional models all in training set are carried out pretreatment and include model normalized and feature
Extracting, wherein, model normalized comprises the following steps:
Step 1-1-1, moves to zero by the barycenter of the threedimensional model in training set, and the barycenter of threedimensional model passes through
On computation model, the area weighted average on all summits obtains；
Step 1-1-2, the threedimensional model each dough sheet center in training set that calculates, to the Euclidean distance of its barycenter, takes all
The intermediate value of Euclidean distance is as specification item, by the coordinate of each point on the threedimensional model in training set divided by this specification item, completes instruction
Practice the normalized of the threedimensional model concentrated；
Feature extraction comprises the following steps:
Step 1-1-3, by the dough sheet feature of each for the threedimensional model in training set patch grids, by curvature feature, PCA spy
Levy, shape characteristics of diameters, average geodesic distance feature and Shape context feature cascade formed a characteristic vector；
Step 1-1-4, by the limit feature of threedimensional model every the Grid Edge in training set, by limit dihedral angle feature, limit neighbour
The dihedral angle feature on summit, territory, the curvature difference on neighborhood summit, both sides, limit and derivative feature, two adjacent dough sheets of shared Grid Edge
Shape diameter difference feature, the shape image difference of two the adjacent dough sheets sharing limit are levied cascade and are formed a characteristic vector.
Step 1-2 comprises the following steps:
Step 1-2-1, concentrates the dough sheet feature of each dough sheet as Single hidden layer feedforward neural networks training patch grids
Input block value input；
Step 1-2-2, modeling Single hidden layer feedforward neural networks is:
Wherein, j=1,2 ... N, N are the quantity that training set patch grids concentrates dough sheet, xjSpecial for dough sheet corresponding for dough sheet j
Levy vector,Number for hidden layer neuron；wi=(wi1,wi2,…,wiN)TRepresent that hidden layer neuron is with input block even
The weight vector connect, may be configured as the random number between (-1,1)；For offset vector, may be configured as (0,1)
Between random number；G (y) is activation primitive；ljFor the actual output of dough sheet j, β in the training set that obtains according to modelkFor output
The output weight that unit is connected with hidden layer, k=1,2 ... m, m are the dimension of output layer, the number of tags i.e. classified；
Step 1-2-3, selects neuron activation function ReLu (the Rectified Linear Units) function that approximation is biological
As activation primitive, compare traditional sigmoid activation primitive and can have the most unilateral inhibition, it is thus achieved that be the broadest
Excited border, obtains sparse activity, it is hereby achieved that the speed of training faster and more stable training precision.ReLu
Activation primitive is defined as: and g (y)=max (0, y).
Step 1-2-4, when approaching this model with zero error,That is:
Wherein, tjStandard mark corresponding to dough sheet each in training set, g is the neuron activation selected in step 1-2-3
Function, the weight matrix between hidden layer and output layer is represented by:
Wherein,Represent output unit and the weight vector of r hidden layer neuron,
Step 1-2-5, approaches according to model being carried out zero error in step 1-2-4, provides matrix table and is shown as H β=T, its
In, T=(t1 t2 … tj), tjStandard mark corresponding to dough sheet j in training set；G is the nerve selected in step 1-2-3
Activation primitive.By solvingObtain exporting weight.Moore-Penrose for matrix
Generalized inverse.So far the quick marking model of dough sheet has been obtained.
Use ELM that limit feature and the standard thereof of every Grid Edge that training Grid Edge is concentrated are marked described in step 1-3
Carry out Fast Training, it is thus achieved that quickly Grid Edge marking model, comprise the following steps:
Step 1-3-1, concentrates the limit feature input as Single hidden layer feedforward neural networks of each edge using training Grid Edge
Cell value inputs.
Step 1-3-2, modeling Single hidden layer feedforward neural networks is:
Wherein, ej=1,2 ... Ne, NeThe quantity of Grid Edge, x is concentrated for training Grid EdgeejFor the limit that Grid Edge ej is corresponding
Feature,For the number of hidden layer neuron, wei=(wei1,wei2,…,weiN)TRepresent that hidden layer neuron is with input block even
The weight vector connect, may be configured as the random number between (-1,1)；For offset vector, may be configured as (0,
1) random number between；G (y) is activation primitive, lejThe actual output of the Grid Edge ej for obtaining according to model, βekFor output
The output weight that unit is connected with hidden layer, ek=1,2 ... me, meFor the dimension of output layer, the number of tags i.e. classified；
Step 1-3-3, selects neuron activation function ReLu (the Rectified Linear Units) function that approximation is biological
As activation primitive, compare traditional sigmoid activation primitive and can have the most unilateral inhibition, it is thus achieved that be the broadest
Excited border, obtains sparse activity, it is hereby achieved that the speed of training faster and more stable training precision.ReLu
Activation primitive is defined as: and g (y)=max (0, y).
Step 1-3-4, when approaching this model with zero error,That is:
Wherein, tej{ 1,2} is standard mark corresponding to Grid Edge ej in training set to ∈, respectively corresponding non-boundary edge and limit
Limit, boundary, g is the neuron activation function selected in step 1-3-3, and the weight matrix between hidden layer and output layer is represented by:
Wherein,Represent output unit and the weight vector of l hidden layer neuron,For the quantity of hidden layer neuron, meFor output layer unit number, that is tag along sort number.
Step 1-3-5, approaches according to model being carried out zero error in step 1-2-4, provides matrix table and be shown as Heβe=Te,
Wherein, Te=(te1 te2 … tej), tejStandard mark corresponding to Grid Edge ej in training set；G is to select in step 1-3-3
The neuron activation function selected.By solvingObtain exporting weight.For matrix
Moore-Penrose generalized inverse.So far the quick marking model of Grid Edge has been obtained.
Step 1-4 comprises the following steps:
Step 1-4-1, uses dough sheet annotation process according to quick dough sheet marking model to the face of threedimensional model in training set
Sheet collection carries out component categories mark, concretely comprises the following steps the dough sheet feature of training set middle mold profile sheet collection as input, according to instruction
The quick marking model of dough sheet got obtains the output valve of model, by output valve by softmax Function Mapping to [0,1],
Thus obtain that there is dough sheet feature vector, XfDough sheet f class probability distribution P (lf,Xf), lfRepresent the mark of patch grids f.
Step 1-4-2, uses limit annotation process according to quick Grid Edge marking model to the limit of threedimensional model in training set
Collection carries out boundary edge classification mark, concretely comprises the following steps the limit feature of training lumped model Grid Edge collection as input, according to instruction
The quick marking model of Grid Edge got obtains the output valve of model, and the output valve that would correspond to boundary edge mark is passed through
Sigmoid Function Mapping is to [0,1], thus obtains having limit feature vector, XeBoundary edge e class probability distribution P (lf≠
lf',Xe), lf'Representing the mark of patch grids f', dough sheet f and dough sheet f' is the adjacent dough sheet with a shared limit；
Step 1-4-3, builds graph model for each threedimensional model in training set, and the node of figure is patch grids, figure
Limit be the shared limit between adjacent dough sheet f and dough sheet f', the negative probability logarithm of the classification annotation of dough sheet cuts the number of optimization as figure
According to item Edata(lf,Xf), Edata(lf,Xf)=-ln (P (lf|Xf)), the negative probability logarithm E of Grid Edge classification annotationweight(lf≠
lf',Xe) × α is as the weight on limit in graph model, for regulating the smoothness of partitioning boundary, uses multi-tag figure to cut optimization and calculates
The mark of each patch grids in target three-dimensional after method calculation optimization.The span of α is [0,10], in this scope
In, arranging step-length is 0.1 to carry out grid search optimization, finds so that in training set, the average segmentation precision of threedimensional model is the highest
α, is designated as figure and cuts the optimal smoothing item weight α of modelbest。
Step 2 comprises the steps:
Step 2-1, carries out pretreatment to target three-dimensional, extracts the dough sheet feature of each patch grids and every grid
The limit feature on limit；
Step 2-2, utilizes quick dough sheet marking model and quick Grid Edge marking model that step 1 obtains, uses dough sheet
Annotation process carries out component categories mark to the dough sheet collection of target three-dimensional, uses limit annotation process to target three-dimensional
Limit collection carries out boundary edge classification mark；
Step 2-3, dough sheet classification annotation probability distribution and Grid Edge classification Marking Probability according to target three-dimensional are divided
Cloth, builds graph model, cuts optimization by multi-tag figure and carries out the smooth of partitioning boundary, the target three-dimensional face after being optimized
Sheet classification annotation results, completes the fast automatic mark to target three-dimensional component categories.
Described in step 2-1, target three-dimensional is carried out pretreatment and include model normalized and feature extraction, its
In, model normalized comprises the following steps:
Step 2-1-1, moves to zero by the barycenter of target three-dimensional, and the barycenter of threedimensional model is by calculating mould
In type, the area weighted average on all summits obtains；
Step 2-1-2, calculates target three-dimensional each dough sheet center to the Euclidean distance of its barycenter, take all European away from
From intermediate value as specification item, the coordinate of each point in target three-dimensional, divided by this specification item, is completed target three-dimensional
Normalized；
Feature extraction comprises the following steps:
Step 2-1-3, by the dough sheet feature of each for target three-dimensional patch grids, by curvature feature, PCA feature, shape
Shape characteristics of diameters, average geodesic distance feature and the cascade of Shape context feature form a characteristic vector；
Step 2-1-4, by the limit feature of target three-dimensional every Grid Edge, by limit dihedral angle feature, neighborhood summit, limit
Dihedral angle feature, the curvature difference on neighborhood summit, both sides, limit and derivative feature, two the proximal surface plate shapes sharing Grid Edge are straight
Footpath difference characteristic, the shape image difference of two the adjacent dough sheets sharing limit are levied cascade and are formed a characteristic vector.
Step 2-2 comprises the following steps:
Step 2-2-1, uses dough sheet annotation process according to the quick dough sheet marking model dough sheet collection to target three-dimensional
Carry out component categories mark, concretely comprise the following steps the dough sheet feature of object module dough sheet collection as input, obtain according to training
The quick marking model of dough sheet obtains the output valve of model, in order to preferably represent that each dough sheet is labeled as the general of different label
Rate, by output valve by softmax Function Mapping to [0,1], thus obtains having dough sheet feature vector, XfDough sheet f classification general
Rate distribution P (lf,Xf), lfRepresent the mark of patch grids f.
Use limit annotation process, according to quick Grid Edge marking model, the limit collection of target three-dimensional is carried out boundary edge class
Do not mark, concretely comprise the following steps the limit feature of object module Grid Edge collection as input, quick according to the Grid Edge that training obtains
Marking model obtains the output valve of model, would correspond to boundary edge mark output valve by sigmoid Function Mapping to [0,
1], thus obtain that there is limit feature vector, XeBoundary edge e class probability distribution P (lf≠lf',Xe), lfRepresent patch grids f
Mark, lf'Representing the mark of patch grids f', dough sheet f and dough sheet f' is for having an adjacent dough sheet in shared limit；
Step 2-2-2, builds graph model for target three-dimensional, and the node of figure is patch grids, and the limit of figure is adjacent
Shared limit between dough sheet f and dough sheet f', the negative probability logarithm of the classification annotation of dough sheet cuts data item E of optimization as figuredata(l,
Xf), the negative probability logarithm E of Grid Edge classification annotationweight(lf≠lf',Xe)×αbestAs the weight on limit in graph model, it is used for
Carry out the smooth optimization of partitioning boundary, each net in the target three-dimensional after using multi-tag figure to cut optimized algorithm calculation optimization
The mark l of lattice dough sheet ff。
Beneficial effect: the invention have the advantages that first, the present invention can be to the mark training of same type of threedimensional model
Collection carries out Fast Training study, it is thus achieved that the quick dough sheet marking model of such threedimensional model and quick Grid Edge marking model；Its
Secondary, the present invention training time greatly reduced compared to the conventional supervised learning time, and the training time can be reduced to from several hours
Several seconds；Finally, the present invention not only make use of the dough sheet markup information of threedimensional model in training set, also further with training
Concentrate the Grid Edge markup information of threedimensional model, thus improve segmentation and mark precision further.
Accompanying drawing explanation
Being the present invention with detailed description of the invention below in conjunction with the accompanying drawings and further illustrate, the present invention's is above-mentioned
And/or otherwise advantage will become apparent.
Fig. 1 is the handling process schematic diagram of the present invention.
Fig. 2 a is the standard mark of dough sheet.
Fig. 2 b is the standard mark of Grid Edge.
Fig. 3 a is a threedimensional model to be marked.
Fig. 3 b is that the target three-dimensional of Fig. 3 a is entered by the quick dough sheet marking model that the present invention is learnt by Fig. 2 a training set
The result of row dough sheet mark.
Fig. 3 c is that the present invention is trained the quick limit marking model the obtained target three-dimensional to Fig. 3 a by Fig. 2 b training set
Carry out the result of Grid Edge mark.
Fig. 3 d is that the dough sheet mark that the present invention is obtained by training set study builds graph model with Grid Edge Marking Probability, carries out
Multi-tag figure cuts the target three-dimensional to Fig. 3 a after optimization and carries out the result of dough sheet mark, the most final segmentation mark knot
Really.
Detailed description of the invention
One disclosed by the invention quick threedimensional model component categories automatic marking method, specifically includes following steps
The flow chart of whole method (Fig. 1 be):
Step one, the Fast Training of threedimensional model mark training set: training has been grouped as in same class model and model parts
Be given standard mark, during wherein markup information includes threedimensional model each patch grids be subordinated to component parts classification mark with
And every Grid Edge is subordinated to the classification mark of boundary edge.As shown in Figure 2 a, the standard for dough sheet marks.Fig. 2 b is Grid Edge
Standard marks.The different labeled class that each model aircraft patch grids is indicated by different colours is labeled, in training set
The standard that is labeled as mark.By threedimensional model mark training set is carried out Fast Training, it is thus achieved that can be used for unknown three-dimensional mould
Type carries out splitting and the quick dough sheet marking model marked and border marking model.And obtain figure cut according to training set data study
The smooth item weight of model.
Step 2, the mark of target three-dimensional: utilize the quick face of training process gained of threedimensional model mark training set
Sheet marking model and quick Grid Edge marking model dough sheet and Grid Edge to target three-dimensional respectively is labeled, thus obtains
Each dough sheet of target three-dimensional is subordinated to the classification mark of component parts and every Grid Edge is subordinated to boundary edge
Classification Marking Probability, builds graph model, and cuts, by multi-tag figure, segmentation and the annotation results that optimization obtains smoothing.
Introduce the main flow of each step in detail below:
1, the Fast Training of threedimensional model mark training set
By threedimensional model mark training set is carried out Fast Training, it is thus achieved that can be used for unknown threedimensional model is labeled
Quick dough sheet marking model and quick Grid Edge marking model.Pre-place is carried out including to threedimensional models all in training set
Reason, uses the ELM method improved to obtain quick dough sheet marking model and three steps of quick Grid Edge marking model.
1.1, threedimensional model pretreatment
Threedimensional model preprocessing process includes model normalized and two steps of feature extraction.
Input: threedimensional model mark training set.
Output: the dough sheet feature of the threedimensional model mark all dough sheets of training set and the limit feature of all Grid Edges.
Step 1 carries out pretreatment to threedimensional models all in training set, obtains normalized threedimensional model collection.Including first
The barycenter of threedimensional model is moved to zero, and the barycenter of model is average by the Area-weighted on all summits on computation model
Value obtains, and the coordinate on the most each summit is by comprising all dough sheet areas on this summit and as weight.Barycenter mV(x, y, z)
It is calculated as:
Wherein, V represents that the vertex set of threedimensional model, the three-dimensional coordinate of summit i are (xi,yi,zi), comprise the surface of summit i
Area is ai, VnumFor the vertex number in vertex set.
Then calculating threedimensional model each dough sheet center Euclidean distance to its barycenter, the intermediate value taking all Euclidean distances is made
For specification item, by the coordinate of each point on threedimensional model divided by this specification item, thus complete the normalized of threedimensional model.
Step 2 extracts the dough sheet feature of each patch grids of threedimensional model, including curvature feature, PCA feature, shape diameter
Feature, average geodesic distance feature and Shape context feature, cascade forms a characteristic vector.Wherein, curvature feature reference
Document 4:Ohtake Y., Belyaev A., Seidel H.-P.Ridge-valley lines on meshes via
Implicit surface fitting.ACM Transactions on Graphics, described in 2004,23 (3): 609-612.
Method calculates.Method described in PCA feature reference literature 1 calculates.Shape diameter Function feature reference literature 5:Shapira L.,
Shalom S.,Shamir A.,Cohen-OrD.,ZhangH.Contextual part analogies in 3D
Objects.InternationalJournal of Computer Vision, 2010,89 (2-3): method described in 309-326.
Calculate.Average geodesic distance feature reference literature 6:Hilaga M., Shinagawa Y., Kohmura T., Kunii
T.L.Topology matching for fully automatic similarity estimationof 3d
shapes.Proceedings of the 28th annual conferenceon Computer graphics and
Interactive techniques (New York, NY, USA, 2001), described in SIGGRAPH ' 01, ACM, pp.203-212.
Method calculates.Shape context feature reference literature 7:Belongie S., Malik J., Puzicha J.Shape
matchingand object recognition using shape contexts.IEEE Transactions
OnPattern Analysis and Machine Intelligence, method described in 2002,24 (4): 509-522. calculates.
Step 3 extracts the limit feature of threedimensional model every Grid Edge, by limit dihedral angle feature, the dihedral angle on neighborhood summit, limit
Feature, the curvature difference on neighborhood summit, both sides, limit and derivative feature, shares two proximal surface plate shape diameter difference features, altogether on limit
Enjoying the shape image difference feature of two adjacent dough sheets on limit, cascade forms a characteristic vector.The extraction reference of limit feature
Method described in document 1 calculates.
1.2, quick dough sheet marking model
Quickly dough sheet marking model acquisition process carries out Fast Classification training to the dough sheet of the threedimensional model in training set, enters
And obtain the data item of graph model.
Input: threedimensional model dough sheet mark training set
Output: the data item of graph model.
Step 1 is according to dough sheet feature X of threedimensional model dough sheetfThe ELM method improved is used to instruct with its standard mark l
Practice.Wherein, mark l ∈ L, L are the mark set of predefined all possible dough sheet sub-categories, e.g., " fuselage ", " wing ",
" vertical tail " and " tailplane ".Dough sheet collection according to threedimensional models all in training set and dough sheet feature Xf, and each
The existing standard of dough sheet marks, and uses document 8:Huang G.-B., Zhou H., Ding X., ZhangR.Extreme
Learning machine for regression and multiclass classification., 2012,42 (2):
513–529.Described ELM sorting technique is trained.During ELM classification based training, the activation primitive of amendment neuron
For ReLu (Rectified Linear Units) function, thus in stability and precision, improve ELM described in original text further
Classifying quality, and reduce computing cost further.
Step 2 exports result for model dough sheet to be marked in the classification of ELM grader, is obtained by softmax function
There is dough sheet feature vector, XfDough sheet, be labeled as l probability distribution P (l | Xf), the corresponding data item of graph model is:
Edata(l,Xf)=-ln (P (l | Xf))。
1.3, quick Grid Edge marking model
Quickly Grid Edge marking model acquisition process carries out Fast Classification instruction to the Grid Edge of the threedimensional model in training set
Practice, and then obtain the weight on limit in graph model.
Input: three-dimensional model gridding limit mark training set
Output: the limit weight of graph model.
Step 1 is according to limit feature X on three-dimensional model gridding limiteFast Training is carried out with its standard mark.For " border
Limit " for, the dough sheet f sharing this Grid Edge is different with the dough sheet of dough sheet f' mark, i.e. lf≠lf'；Otherwise, if sharing this grid
The dough sheet f on limit is identical with the dough sheet of dough sheet f' mark, i.e. lf≠lf', then it is " non-boundary edge ".According to three-dimensionals all in training set
The Grid Edge collection of model and limit feature Xe, and the existing standard in each edge circle limit mark, use the ELM method described in document 7
It is trained.During ELM trains, use ReLu function as the activation primitive of neuron, thus in stability, precision
The ELM classifying quality described in further improvement original text with in efficiency.
Step 2 exports result, by sigmoid function call for model meshes limit to be marked in the test of ELM grader
To having the Grid Edge of limit characteristic vector e, the probability distribution P (l of boundary edgef≠lf'|Xe), graph model corresponding limit weight is:
Eweight(lf≠lf',Xe)=-ln (P (lf≠lf',Xe))
1.4, obtain figure according to training set data study and cut the optimal weight of the smooth item of optimization.
Input: the data item of the figure that training set obtains according to dough sheet marking model and limit weight.
Output: figure cuts the optimal weight optimizing smooth item.
Step 1 builds graph model for each threedimensional model in training set, and the node of figure is patch grids, and the limit of figure is
Shared limit between adjacent dough sheet f and dough sheet f', data item is Edata(lf,Xf), Edata(lf,Xf)=-ln (P (lf|Xf)), smooth
Item Eweight(lf≠lf',Xe) × α is as the weight on limit in graph model, for regulating the smoothness of partitioning boundary, uses many marks
Label figure cuts the mark of each patch grids in the target three-dimensional after optimized algorithm calculation optimization.The span of α be [0,
10], in this range, arranging step-length is 0.1 to carry out grid search optimization, finds so that the average mark of threedimensional model in training set
Cut the α that precision is the highest, be designated as figure and cut the optimal smoothing item weight α of modelbest。
2, the mark of target three-dimensional
Utilize the quick dough sheet marking model of the training process gained of threedimensional model mark training set and quick Grid Edge mark
Injection molding type carries out the mark of dough sheet and Grid Edge to target three-dimensional, builds graph model, cuts optimization by multi-tag figure and obtains
The classification mark of target three-dimensional component parts.This target three-dimensional for training lumped model be subordinated to same kind and
Having same parts to constitute, but not yet split and the threedimensional model of mark, process is as follows:
Input: target three-dimensional.
Output: the dough sheet mark after target three-dimensional segmentation.
Step 1 carries out pretreatment to object module, extracts dough sheet feature and the limit of every Grid Edge of each patch grids
Feature.
Step 2 uses dough sheet annotation process, according to quick dough sheet marking model, the dough sheet collection of target three-dimensional is carried out portion
Part classification marks, and uses limit annotation process, according to quick Grid Edge marking model, the limit collection of target three-dimensional is carried out boundary edge
Classification marks.Wherein, dough sheet classification and the ELM output valve of limit classification, reflected by softmax function and sigmoid function respectively
Be mapped to [0,1], be used for representing dough sheet class probability distribution P (l | Xf) and the probability distribution P (l of boundary edgef≠lf'|Xe)。
Step 3 builds graph model for target three-dimensional, and figure interior joint is patch grids, while be adjacent dough sheet f and f'
Between shared limit, the classification negative probability logarithm of mark of dough sheet cuts data item E of optimization as figuredata(l,Xf), Grid Edge is subordinated to
The mark of boundary edge bears probability logarithm Eweight(lf≠lf',Xe) as the weight on limit in graph model.Use document 9:Boykov
Y.,Veksler O.,Zabih R.Fast Approximate Energy Minimization via Graph
Cuts.IEEE Transactions OnPattern Analysis and Machine Intelligence,2001,23
(11): the multi-tag figure in 1222-1239. cuts optimized algorithm, each patch grids f in the target three-dimensional after calculation optimization
Mark lf。
The target three-dimensional of input is normalized by described preprocessing process, extracts the face of each patch grids
The limit feature of sheet feature and every Grid Edge, process is as follows:
Input: target three-dimensional.
Output: the dough sheet feature of each dough sheet and the limit feature of each edge in target three-dimensional.
The target of described dough sheet annotation process is that the dough sheet in input threedimensional model initially marks, and is labeled as l ∈ L, L
For predefined all possible mark set, e.g., " fuselage ", " wing ", " tailplane " and " vertical tail ", process is such as
Under:
Input: quickly dough sheet marking model and the dough sheet collection of target three-dimensional.
Output: this target three-dimensional dough sheet concentrate each dough sheet class probability distribution P (l | Xf)。
The target of described limit annotation process is that the Grid Edge in input target three-dimensional is labeled.For " boundary edge "
For, the dough sheet f sharing this Grid Edge is different with the dough sheet of dough sheet f' mark, i.e. lf≠lf'；Conversely, for " non-boundary edge "
Speech, the dough sheet f sharing this Grid Edge is identical with the dough sheet of dough sheet f' mark, i.e. lf=lf'.Process is as follows:
Input: quickly border marking model and the Grid Edge collection of target three-dimensional.
Output: this target three-dimensional Grid Edge concentrates the probability distribution P (l of boundary edgef≠lf'|Xe)
The target of described structure graph model process is initial dough sheet mark to carry out border smooth and optimize, thus obtains
Being labeled the classification of the component parts of target three-dimensional, process is as follows:
Input: graph model, including data item E of figure interior jointdata(l,Xf)=-ln (P (l | Xf)), smooth item EsmoothWith
And between figure interior joint, connect weight E on limitweight(lf≠lf',Xe)×αbest=-ln (P (lf≠lf',Xe))×αbest。
Output: in target three-dimensional, each dough sheet f is subordinated to the mark l of component parts classificationf。
Step 1 builds a figure, and the node of figure is the dough sheet of threedimensional model, there is Grid Edge as figure between adjacent dough sheet
Limit, dough sheet f and dough sheet f' are noted as marking lfWith mark lf'Time, limit { weight E of f, f'}weight(lf≠lf',Xe) it is:
Eweight(lf≠lf',Xe)×αbest=-ln (P (lf≠lf',Xe))×αbest
Step 2 calculates data item Edata(l,Xf):
Edata(l,Xf)=-ln (P (l | Xf))
Step 3 arranges smooth item EsmoothBeing 0 for leading diagonal, other element is the L of 1numRank square formation, is used for representing dough sheet
Distance metric between classification mark label.Wherein, LnumStandard for dough sheet marks classification number.
Step 4 uses the multi-tag figure in document 9 to cut optimized algorithm, by the way of calculating this figure minimal cut, calculates three
The optimal label l of each dough sheet on dimension module grid, optimizes graph model, and then tries to achieve after optimization each net in target three-dimensional
The mark l of lattice dough sheet ff。
In the present invention, be illustrated in figure 2 input threedimensional model mark training set example, by of the present invention quickly
Threedimensional model component categories automatic marking method, can split model in Fig. 3 a, obtains initial dough sheet classification chart 3b, limit
Class probability Fig. 3 c, builds graph model and carries out the dough sheet after figure cuts optimization and be labeled as shown in Figure 3 d.
The invention provides a kind of threedimensional model component categories automatic marking method, the method implementing this technical scheme
A lot of with approach, the above is only the preferred embodiment of the present invention, it is noted that for the ordinary skill of the art
For personnel, under the premise without departing from the principles of the invention, it is also possible to make some improvements and modifications, these improvements and modifications are also
Should be regarded as protection scope of the present invention.Each ingredient the clearest and the most definite in the present embodiment all can use prior art to be realized.
Claims (9)
1. a quick threedimensional model component categories automatic marking method, it is characterised in that comprise the following steps:
Step 1, training threedimensional model mark training set, the threedimensional model in training set is for give with parts in class model and model
The quasi-markup information of bid, during wherein markup information includes threedimensional model, each patch grids is subordinated to the classification mark of component parts
And every Grid Edge is subordinated to the classification mark of boundary edge, by threedimensional model mark training set is carried out Fast Training, obtain
Must be used for target three-dimensional is split the quick dough sheet marking model with mark and quick Grid Edge marking model, it is thus achieved that
Threedimensional model dough sheet classification annotation probability distribution and Grid Edge classification annotation probability distribution, and learn to obtain figure and cut the smooth of model
Item weight；
Step 2, label target threedimensional model: utilize quick dough sheet marking model and quick Grid Edge mark mould that step 1 obtains
Target three-dimensional is split and mark by type respectively, it is thus achieved that each dough sheet of target three-dimensional is subordinated to component parts
Classification mark and every Grid Edge are subordinated to the classification mark of boundary edge, build graph model, are cut by multi-tag figure and optimize
To smooth segmentation and annotation results.
Method the most according to claim 1, it is characterised in that step 1 comprises the following steps:
Threedimensional models all in training set are carried out pretreatment by step 1-1, extract the dough sheet feature of each patch grids and every
The limit feature of Grid Edge；
Step 1-2, all dough sheet composing training patch grids collection of threedimensional model in training set, use the extreme learning machine improved
Method, dough sheet feature and standard mark thereof to each patch grids that training patch grids is concentrated carry out Fast Training, it is thus achieved that
Quickly dough sheet marking model；
Step 1-3, all Grid Edge composing training Grid Edge collection of threedimensional model in training set, use the extreme learning machine improved
Method carries out Fast Training to limit feature and the standard mark thereof of every Grid Edge that training Grid Edge is concentrated, it is thus achieved that quickly grid
Limit marking model；
Step 1-4, it is thus achieved that threedimensional model dough sheet classification annotation probability distribution and Grid Edge classification annotation probability distribution, and learn
The optimal smoothing item weight of model is cut to figure.
Method the most according to claim 2, it is characterised in that in step 1-1, described to threedimensional models all in training set
Carrying out pretreatment and include model normalized and feature extraction, wherein, model normalized comprises the following steps:
Step 1-1-1, moves to zero by the barycenter of the threedimensional model in training set, and the barycenter of threedimensional model is by calculating
On model, the area weighted average on all summits obtains；
Step 1-1-2, the threedimensional model each dough sheet center in training set that calculates, to the Euclidean distance of its barycenter, takes all European
The coordinate of each point on threedimensional model in training set, as specification item, divided by this specification item, is completed training set by the intermediate value of distance
In the normalized of threedimensional model；
Feature extraction comprises the following steps:
Step 1-1-3, by the dough sheet feature of each for the threedimensional model in training set patch grids, by curvature feature, PCA feature,
Shape characteristics of diameters, average geodesic distance feature and the cascade of Shape context feature form a characteristic vector；
Step 1-1-4, by the limit feature of threedimensional model every the Grid Edge in training set, by limit dihedral angle feature, neighborhood top, limit
The dihedral angle feature of point, the curvature difference on neighborhood summit, both sides, limit and derivative feature, two proximal surface plate shapes of shared Grid Edge
Diameter difference feature, the shape image difference of two the adjacent dough sheets sharing limit are levied cascade and are formed a characteristic vector.
Method the most according to claim 3, it is characterised in that step 1-2 comprises the steps:
Step 1-2-1, defeated using train patch grids to concentrate the dough sheet feature of each dough sheet as Single hidden layer feedforward neural networks
Enter cell value to input；
Step 1-2-2, modeling Single hidden layer feedforward neural networks is:
Wherein, j=1,2 ... N, N are the quantity that training set patch grids concentrates dough sheet, xjFor dough sheet feature corresponding for dough sheet j to
Amount,For the number of hidden layer neuron, wi=(wi1,wi2,…,wiN)TRepresent what hidden layer neuron was connected with input block
Weight vector,For offset vector, g (y) is activation primitive；ljFor face in the training set that obtains according to model
Sheet j actual output, βkThe output weight being connected with hidden layer for output unit, k=1,2 ... m, m are the dimension of output layer
Degree, the number of tags i.e. classified；
Step 1-2-3, selects ReLu neuron activation function to be defined as activation primitive, ReLu neuron activation function: g (y)=
max(0,y)；
Step 1-2-4, when approaching this model with zero error,That is:
Wherein, tjStandard mark corresponding to dough sheet j in training set, g is the neuron activation function selected in step 1-2-3, hidden
The weight matrix hidden between layer and output layer is expressed as:
Wherein,Represent output unit and the weight vector of r hidden layer neuron,
Step 1-2-5, approaches according to model being carried out zero error in step 1-2-4, provides matrix table and is shown as H β=T, wherein, T
=(t1 t2 … tj),By solvingObtain exporting weight, for the Moore-of matrix
Penrose generalized inverse, thus obtain quick dough sheet marking model.
Method the most according to claim 4, it is characterised in that step 1-3 comprises the steps:
Step 1-3-1, concentrates the limit feature input block as Single hidden layer feedforward neural networks of each edge using training Grid Edge
Value inputs；
Step 1-3-2, modeling Single hidden layer feedforward neural networks is:
Wherein, ej=1,2 ... Ne, NeThe quantity of Grid Edge, x is concentrated for training Grid EdgeejFor limit feature corresponding for Grid Edge ej,For the number of hidden layer neuron, wei=(wei1,wei2,…,weiN)TRepresent what hidden layer neuron was connected with input block
Weight vector,For offset vector, g (y) is activation primitive, lejFor the Grid Edge ej's that obtains according to model
Actual output, βekThe output weight being connected with hidden layer for output unit, ek=1,2 ... me, meFor the dimension of output layer, i.e.
The number of tags of classification；
Step 1-3-3, selects ReLu neuron activation function to be defined as activation primitive, ReLu neuron activation function: g (y)=
max(0,y)；
Step 1-3-4, when approaching this model with zero error,That is:
Wherein, tej∈ 1,2} be in training set corresponding to Grid Edge ej standard mark, correspond respectively to non-boundary edge and border
Limit, g is the neuron activation function selected in step 1-3-3, and the weight matrix between hidden layer and output layer is expressed as:
Wherein,Represent output unit and the weight vector of l hidden layer neuron, For the quantity of hidden layer neuron, meFor output layer unit number, that is tag along sort number；
Step 1-3-5, approaches according to model being carried out zero error in step 1-2-4, provides matrix table and be shown as Heβe=Te, wherein,
Te=(te1 te2 … tej),By solvingObtain exporting weight, for matrix
Moore-Penrose generalized inverse, thus obtain quick Grid Edge marking model.
Method the most according to claim 5, it is characterised in that step 1-4 comprises the steps:
Step 1-4-1, uses dough sheet annotation process according to quick dough sheet marking model to the dough sheet collection of threedimensional model in training set
Carry out component categories mark, concretely comprise the following steps the dough sheet feature of training set middle mold profile sheet collection as input, according to training
To the quick marking model of dough sheet obtain the output valve of model, by output valve by softmax Function Mapping to [0,1], thus
Obtain that there is dough sheet characteristic vector ΧfDough sheet f class probability distribution P (lf,Χf), lfRepresent the mark of patch grids f；
Step 1-4-2, uses limit annotation process to enter the limit collection of threedimensional model in training set according to quick Grid Edge marking model
Row bound limit classification mark, concretely comprises the following steps the limit feature of training lumped model Grid Edge collection as input, according to training
To the quick marking model of Grid Edge obtain the output valve of model, the output valve that would correspond to boundary edge mark passes through sigmoid
Function Mapping is to [0,1], thus obtains having limit characteristic vector ΧeBoundary edge e class probability distribution P (lf≠lf',Χe),
lf'Representing the mark of patch grids f', dough sheet f and dough sheet f' is the adjacent dough sheet with a shared limit；
Step 1-4-3, builds graph model for each threedimensional model in training set, and the node of figure is patch grids, the limit of figure
For the shared limit between adjacent dough sheet f and dough sheet f', the negative probability logarithm of the classification annotation of dough sheet cuts the data item of optimization as figure
Edata(lf,Χf), Edata(lf,Χf)=-ln (P (lf|Χf)), the negative probability logarithm E of Grid Edge classification annotationweight(lf≠
lf',Χe) × α is as the weight on limit in graph model, for regulating the smoothness of partitioning boundary, uses multi-tag figure to cut optimization
The mark of each patch grids in target three-dimensional after algorithm calculation optimization, the span of weight α is [0,10], at this
In the range of, arranging step-length is 0.1 to carry out grid search optimization, finds so that in training set, the average segmentation precision of threedimensional model is
High weight α, is designated as figure and cuts the optimal smoothing item weight α of modelbest。
Method the most according to claim 6, it is characterised in that step 2 comprises the steps:
Step 2-1, carries out pretreatment to target three-dimensional, extracts the dough sheet feature of each patch grids and every Grid Edge
Limit feature；
Step 2-2, utilizes quick dough sheet marking model and quick Grid Edge marking model that step 1 obtains, uses dough sheet mark
Process carries out component categories mark to the dough sheet collection of target three-dimensional, uses the limit annotation process limit collection to target three-dimensional
Carry out boundary edge classification mark；
Step 2-3, according to dough sheet classification annotation probability distribution and the distribution of Grid Edge classification Marking Probability, the structure of target three-dimensional
Build graph model, cut optimization by multi-tag figure and carry out the smooth of partitioning boundary, the target three-dimensional dough sheet class after being optimized
Other annotation results, completes the fast automatic mark to target three-dimensional component categories.
Method the most according to claim 7, it is characterised in that described in step 2-1, target three-dimensional is carried out pre-place
Reason includes model normalized and feature extraction, and wherein, model normalized comprises the following steps:
Step 2-1-1, moves to zero by the barycenter of target three-dimensional, and the barycenter of threedimensional model is by computation model
The area weighted average on all summits obtains；
Step 2-1-2, calculating target three-dimensional each dough sheet center, to the Euclidean distance of its barycenter, takes all Euclidean distances
Intermediate value, as specification item, by the coordinate of each point in target three-dimensional divided by this specification item, completes the normalizing of target three-dimensional
Change processes；
Feature extraction comprises the following steps:
Step 2-1-3 is by the dough sheet feature of each for target three-dimensional patch grids, straight by curvature feature, PCA feature, shape
Footpath feature, average geodesic distance feature and the cascade of Shape context feature form a characteristic vector；
Step 2-1-4, by the limit feature of target three-dimensional every Grid Edge, by limit dihedral angle feature, the two of neighborhood summit, limit
Face corner characteristics, the curvature difference on neighborhood summit, both sides, limit and derivative feature, two proximal surface plate shape diameter difference of shared Grid Edge
Different feature, the shape image difference of two the adjacent dough sheets sharing limit are levied cascade and are formed a characteristic vector.
Method the most according to claim 8, it is characterised in that step 2-2 comprises the following steps:
Step 2-2-1, uses dough sheet annotation process to carry out the dough sheet collection of target three-dimensional according to quick dough sheet marking model
Component categories marks, and concretely comprises the following steps the dough sheet feature of object module dough sheet collection as input, the dough sheet obtained according to training
Quickly marking model obtains the output valve of model, by output valve by softmax Function Mapping to [0,1], thus is had
Dough sheet characteristic vector ΧfDough sheet f class probability distribution P (lf,Χf), lfRepresent the mark of patch grids f；
Use limit annotation process, according to quick Grid Edge marking model, the limit collection of target three-dimensional is carried out boundary edge classification mark
Note, concretely comprises the following steps the limit feature of object module Grid Edge collection as input, quickly marks according to the Grid Edge that training obtains
Model obtains the output valve of model, would correspond to the output valve of boundary edge mark by sigmoid Function Mapping to [0,1], from
And obtain that there is limit characteristic vector ΧeBoundary edge e class probability distribution P (lf≠lf',Χe), lf'Represent patch grids f'
Mark, dough sheet f and dough sheet f' is the adjacent dough sheet with a shared limit；
Step 2-2-2, builds graph model for target three-dimensional, and the node of figure is patch grids, and the limit of figure is adjacent dough sheet f
And the shared limit between dough sheet f', the negative probability logarithm of the classification annotation of dough sheet cuts data item E of optimization as figuredata(lf,Χf),
Edata(lf,Χf)=-ln (P (lf|Χf)), the negative probability logarithm E of Grid Edge classification annotationweight(lf≠lf',Χe)×αbest
As the weight on limit in graph model, for carrying out the smooth optimization of partitioning boundary, use multi-tag figure to cut optimized algorithm and calculate excellent
The mark of each patch grids in target three-dimensional after change.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610530040.6A CN106157375B (en) | 2016-07-06 | 2016-07-06 | A kind of threedimensional model component categories automatic marking method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610530040.6A CN106157375B (en) | 2016-07-06 | 2016-07-06 | A kind of threedimensional model component categories automatic marking method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN106157375A true CN106157375A (en) | 2016-11-23 |
| CN106157375B CN106157375B (en) | 2018-11-30 |
Family
ID=58062820
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610530040.6A Active CN106157375B (en) | 2016-07-06 | 2016-07-06 | A kind of threedimensional model component categories automatic marking method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106157375B (en) |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107833209A (en) * | 2017-10-27 | 2018-03-23 | 浙江大华技术股份有限公司 | A kind of x-ray image detection method, device, electronic equipment and storage medium |
| CN108304876A (en) * | 2018-01-31 | 2018-07-20 | 国信优易数据有限公司 | Disaggregated model training method, device and sorting technique and device |
| CN108427970A (en) * | 2018-03-29 | 2018-08-21 | 厦门美图之家科技有限公司 | Picture mask method and device |
| CN108460415A (en) * | 2018-02-28 | 2018-08-28 | 国信优易数据有限公司 | Pseudo label generates model training method and pseudo label generation method |
| CN108898679A (en) * | 2018-04-10 | 2018-11-27 | 宁波大红鹰学院 | A kind of method of component serial number automatic marking |
| CN109199604A (en) * | 2018-08-31 | 2019-01-15 | 浙江大学宁波理工学院 | A kind of pedicle screw based on multiple features most preferably enters independent positioning method |
| CN109255807A (en) * | 2017-07-13 | 2019-01-22 | 腾讯科技（深圳）有限公司 | A kind of image information processing method and server, computer storage medium |
| CN109491327A (en) * | 2017-09-13 | 2019-03-19 | 发那科株式会社 | Threedimensional model producing device |
| CN109815830A (en) * | 2018-12-28 | 2019-05-28 | 梦多科技有限公司 | A method of obtaining foot information in the slave photo based on machine learning |
| CN110533781A (en) * | 2019-08-28 | 2019-12-03 | 南京信息职业技术学院 | A kind of multi-class threedimensional model component automatic marking method |
| CN110543625A (en) * | 2019-07-29 | 2019-12-06 | 北京航空航天大学 | rapid marking editing and displaying method for 3D virtual training courseware information |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103021029A (en) * | 2013-01-18 | 2013-04-03 | 南京大学 | Automatic labeling method for three-dimensional model component categories |
| CN103473813A (en) * | 2013-09-18 | 2013-12-25 | 南京大学 | Automatic extracting method for three-dimensional model members |
| CN103971415A (en) * | 2014-05-23 | 2014-08-06 | 南京大学 | Online marking method for three-dimensional model component |
-
2016
- 2016-07-06 CN CN201610530040.6A patent/CN106157375B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103021029A (en) * | 2013-01-18 | 2013-04-03 | 南京大学 | Automatic labeling method for three-dimensional model component categories |
| CN103473813A (en) * | 2013-09-18 | 2013-12-25 | 南京大学 | Automatic extracting method for three-dimensional model members |
| CN103971415A (en) * | 2014-05-23 | 2014-08-06 | 南京大学 | Online marking method for three-dimensional model component |
Non-Patent Citations (3)
| Title |
|---|
| FEIQIAN ZHANG ET AL: ""Progressive 3D shape segmentation using online learning"", 《COMPUTER-AIDED DESIGN》 * |
| XIUPING LIU ET AL: ""Low-rank 3D mesh segmentation and labeling with structure guiding"", 《COMPUTERS&GRAPHICS》 * |
| ZHIGE XIE ET AL: ""3D Shape Segmentation and Labeling via Extreme Learning Machine"", 《EUROGRAPHICS SYMPOSIUM ON GEOMETRY PROCESSSING 2014》 * |
Cited By (18)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109255807B (en) * | 2017-07-13 | 2023-02-03 | 腾讯科技（深圳）有限公司 | Image information processing method, server and computer storage medium |
| CN109255807A (en) * | 2017-07-13 | 2019-01-22 | 腾讯科技（深圳）有限公司 | A kind of image information processing method and server, computer storage medium |
| CN109491327A (en) * | 2017-09-13 | 2019-03-19 | 发那科株式会社 | Threedimensional model producing device |
| CN109491327B (en) * | 2017-09-13 | 2023-04-07 | 发那科株式会社 | Three-dimensional model making device |
| CN107833209B (en) * | 2017-10-27 | 2020-05-26 | 浙江大华技术股份有限公司 | X-ray image detection method and device, electronic equipment and storage medium |
| CN107833209A (en) * | 2017-10-27 | 2018-03-23 | 浙江大华技术股份有限公司 | A kind of x-ray image detection method, device, electronic equipment and storage medium |
| CN108304876A (en) * | 2018-01-31 | 2018-07-20 | 国信优易数据有限公司 | Disaggregated model training method, device and sorting technique and device |
| CN108304876B (en) * | 2018-01-31 | 2021-07-06 | 国信优易数据股份有限公司 | Classification model training method and device and classification method and device |
| CN108460415A (en) * | 2018-02-28 | 2018-08-28 | 国信优易数据有限公司 | Pseudo label generates model training method and pseudo label generation method |
| CN108460415B (en) * | 2018-02-28 | 2021-06-15 | 国信优易数据股份有限公司 | Language identification method |
| CN108427970A (en) * | 2018-03-29 | 2018-08-21 | 厦门美图之家科技有限公司 | Picture mask method and device |
| CN108898679A (en) * | 2018-04-10 | 2018-11-27 | 宁波大红鹰学院 | A kind of method of component serial number automatic marking |
| CN108898679B (en) * | 2018-04-10 | 2023-07-21 | 宁波财经学院 | Automatic labeling method for serial numbers of parts |
| CN109199604B (en) * | 2018-08-31 | 2020-12-01 | 浙江大学宁波理工学院 | Multi-feature-based pedicle screw optimal entry point positioning method |
| CN109199604A (en) * | 2018-08-31 | 2019-01-15 | 浙江大学宁波理工学院 | A kind of pedicle screw based on multiple features most preferably enters independent positioning method |
| CN109815830A (en) * | 2018-12-28 | 2019-05-28 | 梦多科技有限公司 | A method of obtaining foot information in the slave photo based on machine learning |
| CN110543625A (en) * | 2019-07-29 | 2019-12-06 | 北京航空航天大学 | rapid marking editing and displaying method for 3D virtual training courseware information |
| CN110533781A (en) * | 2019-08-28 | 2019-12-03 | 南京信息职业技术学院 | A kind of multi-class threedimensional model component automatic marking method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN106157375B (en) | 2018-11-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106157375B (en) | A kind of threedimensional model component categories automatic marking method | |
| Jiao et al. | AF-RCNN: An anchor-free convolutional neural network for multi-categories agricultural pest detection | |
| CN104537676B (en) | Gradual image segmentation method based on online learning | |
| CN109344825A (en) | A kind of licence plate recognition method based on convolutional neural networks | |
| CN102542302B (en) | Automatic complicated target identification method based on hierarchical object semantic graph | |
| CN105825502B (en) | A kind of Weakly supervised method for analyzing image of the dictionary study based on conspicuousness guidance | |
| CN105447473A (en) | PCANet-CNN-based arbitrary attitude facial expression recognition method | |
| CN105117739A (en) | Clothes classifying method based on convolutional neural network | |
| CN103886330A (en) | Classification method based on semi-supervised SVM ensemble learning | |
| CN108629367A (en) | A method of clothes Attribute Recognition precision is enhanced based on depth network | |
| CN104504366A (en) | System and method for smiling face recognition based on optical flow features | |
| CN104834941A (en) | Offline handwriting recognition method of sparse autoencoder based on computer input | |
| CN103971415A (en) | Online marking method for three-dimensional model component | |
| CN103838836A (en) | Multi-modal data fusion method and system based on discriminant multi-modal deep confidence network | |
| CN104636732A (en) | Sequence deeply convinced network-based pedestrian identifying method | |
| CN105005794A (en) | Image pixel semantic annotation method with combination of multi-granularity context information | |
| CN101799875B (en) | Target detection method | |
| CN103065158A (en) | Action identification method of independent subspace analysis (ISA) model based on relative gradient | |
| CN108230330B (en) | Method for quickly segmenting highway pavement and positioning camera | |
| CN106780376A (en) | The background image dividing method of partitioning algorithm is detected and combined based on conspicuousness | |
| CN104050460B (en) | The pedestrian detection method of multiple features fusion | |
| CN106874942A (en) | A kind of object module fast construction method semantic based on regular expression | |
| CN103473813A (en) | Automatic extracting method for three-dimensional model members | |
| Guan et al. | An Object Detection Framework Based on Deep Features and High-Quality Object Locations. | |
| Zhigang et al. | Vehicle target detection based on R-FCN |
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 |