CN109063271B - Three-dimensional CAD model segmentation method and device based on ultralimit learning machine - Google Patents
Three-dimensional CAD model segmentation method and device based on ultralimit learning machine Download PDFInfo
- Publication number
- CN109063271B CN109063271B CN201810758421.9A CN201810758421A CN109063271B CN 109063271 B CN109063271 B CN 109063271B CN 201810758421 A CN201810758421 A CN 201810758421A CN 109063271 B CN109063271 B CN 109063271B
- Authority
- CN
- China
- Prior art keywords
- attribute
- cad model
- dimensional cad
- graph
- learning machine
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Computer Hardware Design (AREA)
- Evolutionary Computation (AREA)
- Geometry (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a three-dimensional CAD model segmentation method and a device based on an ultralimit learning machine, which are used for calculating a feature description operator corresponding to each surface of a three-dimensional CAD model; training and testing an overrun learning machine based on feature description operators of all the surfaces; classifying and labeling each surface of the three-dimensional CAD model by using an overrun learning machine; constructing an attribute adjacency marking chart of the three-dimensional CAD model based on the classification result; segmenting the attribute adjacency label graph; and taking the maximum cohesion degree of the attribute adjacent mark graph segmentation as an objective function to carry out merging optimization on the segmented attribute adjacent mark graph to obtain a plurality of local areas. The invention classifies the plane, the concave surface and the convex surface of the three-dimensional CAD model by using an overrun learning machine, expresses the three-dimensional CAD model by using attribute adjacent mark drawings, and then performs segmentation and optimization according to the attribute adjacent mark drawings corresponding to the three-dimensional CAD model.
Description
Technical Field
The invention relates to the field of three-dimensional CAD model segmentation, in particular to a three-dimensional CAD model segmentation method and device based on an ultralimit learning machine.
Background
The division and marking of the CAD model are fundamental subjects of the feature understanding of the three-dimensional model. Therefore, CAD model segmentation is a key work before the subsequent processing of three-dimensional model features, the traditional surface segmentation method mainly aims at a grid model, and the expression of the three-dimensional CAD model is usually realized by adopting a B-rep boundary representation method. How to divide the CAD model into local areas with certain engineering significance is integrated into a problem to be solved urgently.
The conventional CAD model segmentation adopts a manual mode, and the working efficiency and the precision of the segmentation of the CAD model by the manual mode are low. With the rapid pace development of the field of machine learning, some new classification methods based on machine learning are proposed. At present, the method for segmenting the CAD model by using machine learning mainly includes: supervised learning segmentation methods and unsupervised learning segmentation methods. The best result of model segmentation based on supervised learning in the model segmentation evaluation set of university of princeton is 94%, but the training speed of the supervised learning segmentation method is low. In the unsupervised learning segmentation method, researchers adopt different clustering methods to perform clustering, and then perform segmentation and labeling. In order to accelerate the calculation of clustering, the method usually performs over-segmentation on the three-dimensional model, and then performs feature extraction and further clustering, so that the speed of the method can be greatly improved, but the final segmentation result seriously depends on the effect of over-segmentation.
In summary, the prior art still lacks an effective solution to the problems of too long training time, low generalization accuracy and easy falling into local minimum.
Disclosure of Invention
In order to overcome the defects of the prior art, the invention provides a three-dimensional CAD model segmentation method and a three-dimensional CAD model segmentation device based on an overrun learning machine.
The technical scheme adopted by the invention is as follows:
the invention aims to provide a three-dimensional CAD model segmentation method based on an overrun learning machine classifier, which comprises the following steps:
calculating a feature description operator corresponding to each surface of the three-dimensional CAD model;
training and testing an overrun learning machine based on feature description operators of all faces;
classifying and labeling each surface of the three-dimensional CAD model by using an overrun learning machine;
constructing an attribute adjacency marking chart of the three-dimensional CAD model based on the classification result;
segmenting the attribute adjacency label graph;
and taking the maximum cohesion of the attribute adjacent label graph segmentation as an objective function to carry out merging optimization on the segmented attribute adjacent label graph to obtain a plurality of local areas.
Further, the feature descriptors corresponding to each face of the three-dimensional CAD model comprise features based on principal component analysis, face curvature features and shape diameter features.
Further, the training and testing method of the ultralimit learning machine comprises the following steps:
normalizing the feature based on principal component analysis, the surface curvature feature and the shape diameter feature of all the surfaces to obtain a vector, using the vector as an input feature vector for the training of the classifier of the ultralimit learning machine, and inputting the vector into the ultralimit learning machine for training;
and selecting the number of hidden layer nodes, the number of training models and the neuron emergency function of the ultralimit learning machine, and removing the original weight items in the ultralimit learning machine.
Further, the step of classifying and labeling each face of the three-dimensional CAD model using the ultralimit learning machine includes:
calculating a plurality of neuron probability values of each surface of the three-dimensional CAD model by using an ultralimit learning machine;
normalizing the probability values of a plurality of neurons of each surface to obtain a label probability value of the surface;
and classifying each face of the three-dimensional CAD model by using the label probability value of each face, and distinguishing the plane, the convex surface and the concave surface of the three-dimensional CAD model.
Further, the method for constructing the attribute adjacency marking graph of the three-dimensional CAD model comprises the following steps:
defining a data structure of the attribute adjacency marking graph, including adjacency and concavity and convexity;
traversing each face of the three-dimensional model, extracting all attributes of each face, and creating corresponding nodes of attribute adjacent label graphs;
the adjacency relation between each face of the three-dimensional model is recognized, and the edge of the attribute adjacency label graph is created.
Further, the method for segmenting the attribute adjacency label graph comprises the following steps:
segmenting the attribute adjacent marker graph G according to the attributes of the nodes and the connecting lines in the attribute adjacent marker graph to obtain a plurality of local region sub-graphs to form a local region set S;
deleting each node in the local area set S and the connecting lines between the nodes from the attribute adjacent tag graph G to obtain a new attribute adjacent tag graph G';
if the new attribute connection tag graph G' is empty, the segmentation of all nodes and connecting lines in the attribute adjacent tag graph is finished;
if the new attribute connection mark graph G 'contains subgraphs of mixed nodes, the attribute connection mark graph G' containing the mixed nodes is divided again according to the principle that the divided concave subgraph is firstly identified and then the convex subgraph is divided until the new attribute adjacent mark graph is obtained to be empty.
Further, the method also comprises the step of setting the regional cohesion degree, the cohesion degree of attribute adjacent label graph segmentation and the regional coupling degree, wherein the regional cohesion degree is the average degree of each node in the attribute adjacent label graph; the cohesiveness of the attribute connection marking graph is the average value of the cohesiveness of each local area corresponding to the three-dimensional CAD model; and the regional coupling degree is a connecting line of nodes in any two local region sub-graphs in the local region set S in the attribute connection marked graph corresponding to the three-dimensional CAD model.
Further, the method for merging and optimizing the segmented attribute adjacent label graph comprises the following steps:
calculating the cohesion degree of the local region set S obtained after segmentation according to the cohesion expression of attribute adjacency marking graph segmentation, and obtaining a new local region set S' after updating;
for each local region in the local region set S', a sub-graph G is generated i Analyzing to obtain a plurality of alternative merged subgraphs to form an alternative set A;
selecting and G from the alternative set A i Combining the subgraphs with the maximum coupling degree;
and calculating the cohesion of the segmented local region set S 'according to the cohesion of the attribute adjacent label graph, comparing the cohesion of the local region set S before segmentation with the cohesion of the segmented local region set S', and outputting the local region set S 'if the cohesion of the local region set S' is smaller than the cohesion of the local region set S.
Further, each local region sub-graph G in the pair of local region sets S i ' the method of performing the assay is:
each local region in the local region set SSubfigure G i ' analyzing the attribute;
judging whether subgraph G with partial area and convex-concave property of nodes and connecting lines of subgraph exists i ' consistent;
subgraph G if there are nodes of subgraph and connected convexity and concavity of subgraph i ' same, then the subgraph is taken as an alternative merged subgraph.
A second object of the present invention is to provide an apparatus for three-dimensional CAD model segmentation based on an ultralimit learning machine classifier, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the following steps when executing the program, including:
calculating a feature description operator corresponding to each surface of the three-dimensional CAD model;
training and testing an overrun learning machine based on feature description operators of all the surfaces;
classifying and labeling each surface of the three-dimensional CAD model by using an overrun learning machine;
constructing an attribute adjacency marking chart of the three-dimensional CAD model based on the classification result;
segmenting the attribute adjacency label graph;
and taking the maximum cohesion degree of the attribute adjacent mark graph segmentation as an objective function to carry out merging optimization on the segmented attribute adjacent mark graph to obtain a plurality of local areas.
Compared with the prior art, the invention has the beneficial effects that:
(1) The method extracts features for each surface of the three-dimensional CAD model, trains the overrun learning machine, predicts the segmentation and marking of the input three-dimensional CAD model through the overrun learning machine, and adopts the overrun learning machine as a classifier, so that the rapid training and testing speed can be obtained;
(2) Compared with the traditional BP neural network applying an error gradient descent learning strategy, the ultralimit learning machine adopted by the method has the advantages of high learning speed, high generalization precision, no falling into a local minimum value and capability of adopting various excitation functions.
Drawings
The accompanying drawings, which are incorporated in and constitute a part of this application, are included to provide a further understanding of the application, and the description of the exemplary embodiments and illustrations of the application are intended to explain the application and are not intended to limit the application.
FIG. 1 is a first flowchart of a three-dimensional CAD model segmentation method based on an overrun learning machine;
FIG. 2 is a flow chart of a three-dimensional CAD model segmentation method based on an overrun learning machine;
FIG. 3 is a schematic diagram of a three-dimensional CAD model;
FIG. 4 is a pictorial view of a property adjacency mark of a three-dimensional CAD model;
fig. 5 is a diagram showing an example of attribute adjacency label graph division.
Detailed Description
The invention is further described with reference to the following figures and examples.
It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the disclosure. Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.
It is noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of example embodiments according to the present application. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, and it should be understood that when the terms "comprises" and/or "comprising" are used in this specification, they specify the presence of stated features, steps, operations, devices, components, and/or combinations thereof, unless the context clearly indicates otherwise.
As described in the background art, in order to solve the above technical problems, the present application provides a three-dimensional CAD model segmentation method based on an ultralimit learning machine, which has the disadvantages of too long training time, low generalization accuracy, and easy falling into a local minimum in the prior art.
Example 1
In an exemplary embodiment of the present application, as shown in fig. 1, there is provided an overrun learning machine-based three-dimensional CAD model segmentation method, including the following steps:
step 101: calculating a feature description operator corresponding to each surface of the three-dimensional CAD model;
step 102: training and testing an overrun learning machine based on feature description operators of all the surfaces;
step 103: classifying and labeling each surface of the three-dimensional CAD model by using an ultralimit learning machine;
step 104: constructing an attribute adjacency marking chart of the three-dimensional CAD model based on the classification result;
step 105: segmenting the attribute adjacency label graph;
step 106: and defining the regional cohesion degree, the divided cohesion degree of the attribute adjacent label graph and the regional coupling degree, and combining and optimizing the divided attribute adjacent label graph by taking the maximum cohesion degree of the attribute adjacent label graph as a target function to obtain a plurality of local regions.
The embodiment of the invention provides a three-dimensional CAD model segmentation method based on an overrun learning machine, which is used for extracting features of each surface of the three-dimensional CAD model, then training the overrun learning machine, predicting the segmentation and labeling of the input three-dimensional CAD model through the overrun learning machine, and obtaining rapid training and testing speed by adopting the overrun learning machine as a classifier.
Example 2
In order to make those skilled in the art better understand the present invention, a more detailed embodiment is listed below, and as shown in fig. 2, an embodiment of the present invention provides a method for segmenting a three-dimensional CAD model based on an ultralimit learning machine, which includes the following steps:
step 201: and calculating a feature description operator of each three-dimensional CAD model.
And calculating corresponding feature descriptors for each surface of the selected three-dimensional CAD model, wherein the feature descriptors can fully reflect the attribute of each surface, and the surfaces (plane, convex surface and concave surface) corresponding to the three-dimensional CAD model can be distinguished by combining the features by using an ultralimit learning machine.
The feature descriptors corresponding to each face include principal component analysis-based features, face curvature features (principal curvature, minimum curvature, maximum curvature), shape diameter features.
The calculation method of the feature based on the principal component analysis comprises the following steps:
Fp＝a 1i *ZX 1 +a 2i *ZX 2 +……+a pi *ZX p
wherein, a 1i 、a 2i 、……、a pi (i =1, \8230;, m) is a feature vector corresponding to a feature value of the covariance matrix Σ of X, ZX 1 、ZX 2 、……、ZX p The method is characterized in that the method is a value of an original variable subjected to standardization processing, m is the number of principal components, and p is the number of corresponding eigenvectors; in the calculation process, the indexes are different in dimension, so that the influence of the dimension is eliminated before calculation, and the original data is standardized.
A＝(a ij )p×m＝(a 1 ，a 2 ，…a m )；
Ra i ＝λ i a i ，
A is a feature vector corresponding to the normalized A; r is a matrix of correlation coefficients, λ i 、a i Are the corresponding eigenvalues and unit eigenvectors, where 1 ≥λ 2 ≥…≥λ p ≥0。
The surface curvature characteristic calculation method comprises the following steps:
the curved surface can be expressed by a parametric equation
P(u,v)＝(x(u,v),y(u,v),z(u,v))
Wherein x (u, v) is the component of x in the u, v direction; y (u, v) is the component of y in the u, v direction; z (u, v) is the component of z in the u, v direction; p (u, v) represents a parametric equation for the surface, and the surface is represented by the parametric equation.
The first order directional derivatives in the u and v directions are as follows:
P u (u,v)＝(x u (u,v),y u (u,v),z u (u,v))
P v (u,v)＝(x v (u,v),y v (u,v),z v (u,v))
suppose that:
E＝p u ·p u
F＝p v ·p u ＝p u ·p v
G＝p v ·p v
a symmetrical determinant can be obtained:
P u and P v The first partial derivative combination of (a) can be expressed as:
ξp u +ηp v
the quadratic product of the combination is:
(ξp u +ηp v )·(ξp u +ηp v )＝Eξ 2 +2Fξη+Gη 2
wherein xi is the expression p u The coefficient of (a); eta is expression p v The coefficient of (c). p is a radical of u And p v The first partial derivative combination of (a) may be expressed as:
ξp u +ηp v
the invention can fully reflect the attribute of each surface through the feature description operators, and can distinguish the surfaces (plane, convex surface and concave surface) corresponding to the three-dimensional CAD model by combining the features through an ultralimit learning machine.
Step 202: the overrun learning machine phase is trained and tested.
The training and testing method of the ultralimit learning machine comprises the following steps:
normalizing the feature based on principal component analysis, the surface curvature feature and the shape diameter feature of each surface of the selected three-dimensional CAD model, forming operators into a vector, using the vector as an input feature vector for the training of the ultralimit learning machine classifier, and inputting the vector into the ultralimit learning machine for training;
and selecting parameters such as the number of nodes of the hidden layer, the number of training models, a neuron emergency function and the like of the ultralimit learning machine.
The over-limit learning machine can automatically learn the weight corresponding to the neuron, so that the over-limit learning machine can automatically select the attribute characteristics of each surface of the input model.
The invention adopts the ultralimit learning machine as the classifier, and can obtain fast training and testing speed.
Step 203: and classifying and labeling the plane, the concave surface and the convex surface of each surface of the three-dimensional CAD model by using a neural network trained by an ultralimit learning machine.
And calculating the label probability value of each surface of the three-dimensional CAD model by using an overrun learning machine, and classifying each surface of the three-dimensional CAD model by using the label probability value. The specific implementation method comprises the following steps:
the three-dimensional model to be segmented is regarded as a graph and is marked as G = { V, E }, a node V in the graph corresponds to each face of the model, and arcs { u, V } in the graph represent that the faces u and V are in an adjacent relation in the model. The number n of the segmentation labels of the model is recorded. Through the training and testing process of the ELM classifier, each face u in the model to be segmented can obtain n neuron probability values. According to ELM learning theory, surface label l u Is obtained from the label corresponding to the maximum value of the n neuron probability values. The probability values of n neurons facing u are normalized to obtain the label probability value p (l) of the probability values u U) to better classify the three-dimensional model surface.
And classifying each face of the selected three-dimensional CAD model by using the label probability value of each face of the three-dimensional CAD model to distinguish the plane, convex surface and concave surface of the CAD model.
The invention trains a neural network to classify the plane, the concave surface and the convex surface of the three-dimensional CAD model by adopting an ultralimit learning machine method without manual classification, thereby accelerating the efficiency and the accuracy of the classification of the CAD surface by manual classification and simplifying the step of CAD model segmentation.
Step 204: and establishing an attribute adjacency marking chart of the selected three-dimensional CAD model.
The method for constructing the attribute adjacency marking map of the three-dimensional CAD model comprises the following steps:
(1) Defining a data structure of an attribute adjacency marking (AALG) graph, including adjacency, ruggedness, and the like;
(2) Traversing each surface of the three-dimensional CAD model, extracting each attribute of each surface, and creating corresponding nodes of attribute adjacent label graphs;
(3) The adjacency relationships between the faces of the three-dimensional CAD model are identified, and edges of the attribute adjacency label graph are created.
The three-dimensional CAD model is shown in FIG. 3, and the attribute adjacency label graph of the three-dimensional CAD model is shown in FIG. 4, wherein V represents a collection of nodes, and each node has a unique face in the model corresponding to the node; e represents a set of connecting lines, and each connecting line has a unique edge in the model corresponding to the connecting line; CV = {0,1,2} is a set of convex-concave labels for nodes, representing convex-concave for corresponding faces, where 0 represents convexity, 1 represents concavity, and 2 represents mixedness; LE = { +, -,0} is a mark set of connecting lines, which represents the convexity and concavity of the corresponding edge, wherein + represents convex edge, -represents concave edge, and 0 represents cut edge.
The connecting edge between the surfaces can be distinguished according to the outer included angle between the curved surfaces and is divided into a convex edge, a concave edge and an edge cutting. If a surface belongs to a convex curved surface or a plane and the connecting edge of the surface does not belong to a concave edge and a concave cut edge, the node becomes a convex node of the graph, and if a surface belongs to a concave curved surface or a plane and the connecting edge of the surface does not belong to a convex edge and a convex cut edge, the node becomes a concave node of the graph. If a node satisfies the convex (concave) surface of the surface it represents, and its adjoining edges have concave (convex) edges or concave (convex) cut edges.
The invention establishes the attribute adjacency mark maps of the three-dimensional CAD model, each attribute adjacency mark map corresponds to a local area with continuous convexoconcave, and the attribute adjacency maps are utilized to divide the three-dimensional model, thereby more intuitively dividing the local area with better cohesion and more engineering significance.
Step 204: and segmenting the three-dimensional CAD model.
The method comprises the steps of firstly segmenting an Attribute Adjacency Label Graph (AALG) according to attributes such as the convexity and concavity of nodes and connecting lines in the Attribute Adjacency Label Graph (AALG), then respectively defining the regional cohesion degree, the cohesion degree of segmentation of the Attribute Adjacency Label Graph (AALG) and the regional coupling degree, carrying out merging optimization by taking the maximum cohesion degree of segmentation of the Attribute Adjacency Label Graph (AALG) as a target, and segmenting a plurality of local regions which meet requirements.
Step 2041: the attribute adjacency label graph is segmented.
The segmentation method of the attribute adjacency label graph comprises the following steps:
dividing the attribute adjacent mark map G with determined concave-convex property, and storing the division result in a local area set S; then deleting each node in the local area set S and the connecting lines between the nodes from the attribute adjacent tag graph G, and assigning a value to a new attribute adjacent tag graph G'; if the new attribute adjacency label graph G 'is empty at this time, it indicates that the segmentation of all nodes and links in the attribute adjacency label graph G has been completed, otherwise, the new attribute adjacency label graph G' contains mixed nodes.
And further segmenting the subgraph G' containing the mixed nodes according to the principle of firstly identifying and segmenting the concave subgraph and then segmenting the convex subgraph until all the attribute adjacent marker graphs G have clear concave-convex characteristics.
Step 2042: defining a regional cohesion, a cohesion of Attribute Adjacency Label Graph (AALG) segmentation, and a regional coupling.
The attribute adjacency marking graph (AALG) segmentation method is easy to generate a considerable number of 'isolated points' or subgraphs with few nodes, so that the cohesion degree between nodes in a region is too high, and the model region segmentation effect is not ideal.
The degree of association between each node in the Attribute Adjacency Label Graph (AALG) corresponding to the region, and the related association between the nodes are represented by connecting lines, in this case, the regional cohesion can be defined by the average degree of each node in the subgraph.
Cohesion of attribute adjacency marking map (AALG) segmentation: and averaging the cohesion degrees of the local areas corresponding to the three-dimensional CAD model.
The area coupling degree: and setting any two local region sub-graphs in the local region set S after the attribute adjacency marking graph (AALG) is segmented, wherein the connection line of the nodes in the two local region sub-graphs in the attribute adjacency marking graph corresponding to the three-dimensional CAD model reflects the coupling degree between the two sub-graphs.
Step 2043: and (6) merging and optimizing.
Conditions that can be combined are as follows: for each subgraph of the three-dimensional CAD model, if the merging of adjacent subgraphs is completed and still a regional subgraph, the two subgraphs can be merged as shown in FIG. 5.
The AALG segmentation optimization merging method based on the cohesion degree comprises the following steps:
inputting: preliminary segmentation result S of AALG.
And (3) outputting: the segmented optimization merges the result S'.
(1) And S '← S, calculating the degree of cohesion of S based on the expression of the degree of cohesion divided by AALG, and updating the calculation result to obtain S'.
The cohesion expression for AALG segmentation is:
in the formula, G i 'induced subgraph of attribute adjacency label graph S'; v is a node in the corresponding subgraph; v. of j Is a node j; | G i ′ v j I is subgraph G i The number of nodes of'.
(2) For each sub-graph G of S i ', each subgraph G in S i ' analyze, if there is a sub-graph, the convex-concave nature of the nodes and connecting lines of this sub-graph and G i ' then regard this subgraph as the alternative and amalgamate subgraph, call all alternative subgraph sets together, alternative set A; then select from the alternative set A and G i ' the subgraph with the largest coupling degree is combined.
And finishing the segmentation operation of the AALG and the evaluation of the regional cohesion and coupling degree on each subgraph of the processed S'.
Definition of the degree of area coupling: let any two regions G in AALG-split S i ′，G j ' the connecting line of the nodes in the two subgraphs in the attribute adjacency label graph G reflects the degree of coupling between the two subgraphs.
Expression of the degree of area coupling: f (G) i ′,G j ′)＝ink(G i ′,G j ′)
Link () represents the number of links in G between nodes in two subgraphs, and the greater the number of links, the higher the coupling degree between the subgraphs of the region.
When subgraph merging is carried out, if a plurality of alternative merging subgraphs meeting the convex-concave continuity condition exist, the subgraph with the largest regional coupling degree is preferably considered to be merged, because the greater the coupling degree between the subgraphs is, the closer the regional association is, and the greater the possibility that the subgraphs belong to the same region is.
(3) And (3) calculating the cohesion degree of the S 'after segmentation according to the cohesion degree of the defined AALG segmentation, comparing the cohesion degrees of the S and the S', if the cohesion degree of the S 'is greater than the cohesion degree of the S, turning to the step (1), otherwise, outputting S', and finishing the algorithm.
In order to further improve the effect of the CAD model segmentation, the invention converts the problem of the three-dimensional CAD model segmentation into attributes according to the nodes and the connecting edges of the graph, extracts the attributes of the three-dimensional CAD model in a B-rep form in order to improve the effect of the three-dimensional CAD model segmentation, and converts the model into an attribute adjacent marked graph; then, the model is preliminarily divided into local areas with continuous convexoconcave according to the convexoconcave of nodes and connecting lines in the graph, and then the selection and the optimized combination of combinable areas are carried out based on the regional coupling degree with the maximum dividing cohesion degree as a target; and (3) segmenting the attribute adjacent mark graph corresponding to the three-dimensional CAD model into a plurality of subgraphs by using the attribute and relationship problem of the nodes and the connecting edges of the graph, wherein the subgraphs are induced subgraphs which are continuous in concave-convex and not mutually intersected.
The embodiment of the invention provides a three-dimensional CAD model segmentation method based on an ultralimit learning machine, which comprises the steps of calculating a feature description operator of each three-dimensional CAD model, forming the operators into a vector to be used as an input feature vector for training an ultralimit learning machine classifier, then training the ultralimit learning machine, classifying and marking the input model by using the trained ultralimit learning machine classifier, representing the three-dimensional CAD model by using a property adjacent label graph by using B-rep of the CAD model as an information input source, and representing the three-dimensional CAD model by using a property adjacent label graph; and then, carrying out segmentation set optimization merging on the CAD model by taking the maximum cohesion degree of the segmented regions and the like as an objective function according to the attribute adjacent marked graph.
Although the embodiments of the present invention have been described with reference to the accompanying drawings, it is not intended to limit the scope of the present invention, and it should be understood by those skilled in the art that various modifications and variations can be made without inventive efforts by those skilled in the art based on the technical solution of the present invention.
Claims (4)
1. A three-dimensional CAD model segmentation method based on an overrun learning machine classifier is characterized by comprising the following steps:
calculating a feature description operator corresponding to each surface of the three-dimensional CAD model;
training and testing an overrun learning machine based on feature description operators of all faces;
classifying and labeling each surface of the three-dimensional CAD model by using an overrun learning machine;
constructing an attribute adjacency marking chart of the three-dimensional CAD model based on the classification result;
segmenting the attribute adjacency label graph;
merging and optimizing the segmented attribute adjacent label graphs by taking the maximum cohesion of the attribute adjacent label graph segmentation as a target function to obtain a plurality of local areas;
the feature description operator corresponding to each surface of the three-dimensional CAD model comprises features based on principal component analysis, surface curvature features and shape diameter features;
the training and testing method of the ultralimit learning machine comprises the following steps: normalizing the characteristics of all surfaces based on principal component analysis, surface curvature characteristics and shape diameter characteristics to obtain a vector which is used as an input characteristic vector for the training of an ultralimit learning machine classifier and is input into an ultralimit learning machine for training, selecting the number of hidden layer nodes, the number of training models and a neuron emergency function of the ultralimit learning machine, and removing original weight items in the ultralimit learning machine;
the step of classifying and labeling each face of the three-dimensional CAD model by using the ultralimit learning machine comprises the following steps: calculating a plurality of neuron probability values of each face of the three-dimensional CAD model by using an overrun learning machine, carrying out normalization processing on the neuron probability values of each face to obtain a label probability value of the face, classifying each face of the three-dimensional CAD model by using the label probability value of each face, and distinguishing a plane, a convex surface and a concave surface of the three-dimensional CAD model;
the method for constructing the attribute adjacency marking map of the three-dimensional CAD model comprises the following steps: defining a data structure of the attribute adjacent label graph, wherein the data structure comprises adjacency and concave-convex, traversing each face of the three-dimensional model, extracting all attributes of each face, creating corresponding nodes of the attribute adjacent label graph, identifying the adjacent relation between each face of the three-dimensional model, and creating edges of the attribute adjacent label graph;
the segmentation method of the attribute adjacent label graph comprises the following steps: dividing the attribute adjacent label graph G according to the attributes of nodes and connecting lines in the attribute adjacent label graph to obtain a plurality of local area subgraphs to form a local area set S, deleting each node in the local area set S and the connecting lines between the nodes from the attribute adjacent label graph G to obtain a new attribute adjacent label graph G ', if the new attribute adjacent label graph G' is empty, indicating that the division of all the nodes and the connecting lines in the attribute adjacent label graph is finished, and if the new attribute adjacent label graph G 'contains subgraphs of mixed nodes, dividing the attribute connected label graph G' containing the mixed nodes again according to the principle of firstly identifying the divided concave subgraph and then dividing the convex subgraph until the new attribute adjacent label graph is empty;
the method for merging and optimizing the segmented attribute adjacent label graph comprises the following steps: calculating the local part obtained after segmentation according to the cohesive expression of attribute adjacent label graph segmentationUpdating the cohesion of the region set S to obtain a new local region set S ', and determining each local region sub-graph G ' in the local region set S ' i Analyzing to obtain several optional merged subgraphs to form optional set A, and selecting G 'from optional set A' i Merging the subgraphs with the maximum coupling degree, calculating the cohesion degree of the local area set S 'after segmentation according to the cohesion degree of the attribute adjacent marked graph, comparing the cohesion degree of the local area set S before segmentation with the cohesion degree of the local area set S' after segmentation, and outputting the local area set S 'if the cohesion degree of the local area set S' is less than the cohesion degree of the local area set S.
2. The method for segmenting the three-dimensional CAD model based on the ultralimit learning machine classifier as claimed in claim 1, further comprising the steps of setting a region cohesion degree, a cohesion degree of segmentation of the attribute adjacent mark map and a region coupling degree, wherein the region cohesion degree is an average degree of each node in the attribute adjacent mark map; the cohesion degree of the attribute connection marking graph is the average value of the cohesion degree of each local area corresponding to the three-dimensional CAD model; the region coupling degree is a connecting line of nodes in any two local region subgraphs in the local region set S in the attribute connection marked graph corresponding to the three-dimensional CAD model.
3. The three-dimensional CAD model segmentation method based on ultralimit learning machine classifier according to claim 1, wherein each local region sub-graph G ' in the local region set S ' is subjected to ' i The method of performing the analysis was:
g 'is sub-mapped to each local region in the local region set S' i Analyzing the attributes of (1);
judging whether the convex-concave property of the nodes and the connecting lines of the subgraph and the local area subgraph G 'exist or not' i The consistency is achieved;
subgraph G 'if there are nodes of subgraph and connected concavity and local region subgraph G' i And if the sub-graphs are the same, taking the sub-graph as an alternative combined sub-graph.
4. A three-dimensional CAD model segmentation device based on an overrun learning machine classifier is characterized by comprising a memory, a processor and a computer program which is stored on the memory and can run on the processor, wherein the processor executes the program to realize the following steps, and the three-dimensional CAD model segmentation device comprises:
calculating a feature description operator corresponding to each surface of the three-dimensional CAD model;
training and testing an overrun learning machine based on feature description operators of all the surfaces;
classifying and labeling each surface of the three-dimensional CAD model by using an overrun learning machine;
constructing an attribute adjacency marking chart of the three-dimensional CAD model based on the classification result;
segmenting the attribute adjacency label graph;
merging and optimizing the segmented attribute adjacent label graphs by taking the maximum cohesion of the attribute adjacent label graph segmentation as a target function to obtain a plurality of local areas;
the feature description operator corresponding to each face of the three-dimensional CAD model comprises features based on principal component analysis, face curvature features and shape diameter features;
the training and testing method of the ultralimit learning machine comprises the following steps: normalizing the principal component analysis-based features, the surface curvature features and the shape diameter features of all surfaces to obtain a vector which is used as an input feature vector for the training of the classifier of the ultralimit learning machine and is input into the ultralimit learning machine for training, selecting the number of hidden layer nodes, the number of training models and a neuron emergency function of the ultralimit learning machine, and removing original weight terms in the ultralimit learning machine;
the step of classifying and labeling each face of the three-dimensional CAD model by using the ultralimit learning machine comprises the following steps: calculating a plurality of neuron probability values of each face of the three-dimensional CAD model by using an overrun learning machine, carrying out normalization processing on the neuron probability values of each face to obtain a label probability value of the face, classifying each face of the three-dimensional CAD model by using the label probability value of each face, and distinguishing a plane, a convex surface and a concave surface of the three-dimensional CAD model;
the method for constructing the attribute adjacency marking map of the three-dimensional CAD model comprises the following steps: defining a data structure of the attribute adjacent label graph, wherein the data structure comprises adjacency and concave-convex, traversing each face of the three-dimensional model, extracting all attributes of each face, creating corresponding nodes of the attribute adjacent label graph, identifying the adjacent relation between each face of the three-dimensional model, and creating edges of the attribute adjacent label graph;
the segmentation method of the attribute adjacent label graph comprises the following steps: dividing the attribute adjacent label graph G according to the attributes of nodes and connecting lines in the attribute adjacent label graph to obtain a plurality of local area subgraphs to form a local area set S, deleting each node in the local area set S and the connecting lines between the nodes from the attribute adjacent label graph G to obtain a new attribute adjacent label graph G ', if the new attribute adjacent label graph G' is empty, indicating that the division of all the nodes and the connecting lines in the attribute adjacent label graph is finished, and if the new attribute adjacent label graph G 'contains subgraphs of mixed nodes, dividing the attribute connected label graph G' containing the mixed nodes again according to the principle of firstly identifying the divided concave subgraph and then dividing the convex subgraph until the new attribute adjacent label graph is empty;
the method for merging and optimizing the segmented attribute adjacent label graph comprises the following steps: calculating the cohesion degree of the local region set S obtained after the division according to the cohesion expression obtained by dividing the attribute adjacent label graph, and obtaining a new local region set S 'after updating' i Analyzing to obtain several optional merged subgraphs to form optional set A, and selecting G 'from optional set A' i Merging the subgraphs with the maximum coupling degree, calculating the cohesion degree of the local area set S 'after segmentation according to the cohesion degree of the attribute adjacent marked graph, comparing the cohesion degree of the local area set S before segmentation with the cohesion degree of the local area set S' after segmentation, and outputting the local area set S 'if the cohesion degree of the local area set S' is less than the cohesion degree of the local area set S.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810758421.9A CN109063271B (en) | 2018-07-11 | 2018-07-11 | Three-dimensional CAD model segmentation method and device based on ultralimit learning machine |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810758421.9A CN109063271B (en) | 2018-07-11 | 2018-07-11 | Three-dimensional CAD model segmentation method and device based on ultralimit learning machine |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109063271A CN109063271A (en) | 2018-12-21 |
| CN109063271B true CN109063271B (en) | 2022-12-20 |
Family
ID=64815988
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810758421.9A Active CN109063271B (en) | 2018-07-11 | 2018-07-11 | Three-dimensional CAD model segmentation method and device based on ultralimit learning machine |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109063271B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111079208B (en) * | 2019-11-20 | 2024-01-23 | 杭州电子科技大学 | Particle swarm algorithm-based CAD model surface corresponding relation identification method |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2013015958A (en) * | 2011-07-01 | 2013-01-24 | Hitachi Ltd | Dependence evaluation device among software development deliverable and evaluation method |
| CN106327506A (en) * | 2016-08-05 | 2017-01-11 | 北京三体高创科技有限公司 | Probability-partition-merging-based three-dimensional model segmentation method |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20090092299A1 (en) * | 2007-10-03 | 2009-04-09 | Siemens Medical Solutions Usa, Inc. | System and Method for Joint Classification Using Feature Space Cluster Labels |
-
2018
- 2018-07-11 CN CN201810758421.9A patent/CN109063271B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2013015958A (en) * | 2011-07-01 | 2013-01-24 | Hitachi Ltd | Dependence evaluation device among software development deliverable and evaluation method |
| CN106327506A (en) * | 2016-08-05 | 2017-01-11 | 北京三体高创科技有限公司 | Probability-partition-merging-based three-dimensional model segmentation method |
Non-Patent Citations (2)
| Title |
|---|
| 三维形状分割和标注的快速学习方法;李红岩;《计算机工程与应用》;20170601(第11期);全文 * |
| 面向三维CAD模型检索的模型分割方法;皇甫中民等;《机械科学与技术》(第05期);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109063271A (en) | 2018-12-21 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US11853903B2 (en) | SGCNN: structural graph convolutional neural network | |
| CN105741355B (en) | A kind of block dividing method of triangle grid model | |
| CN109002834B (en) | Fine-grained image classification method based on multi-modal representation | |
| CN105975913B (en) | Road network extraction method based on adaptive cluster learning | |
| US8238660B2 (en) | Hybrid graph model for unsupervised object segmentation | |
| CN108345587B (en) | Method and system for detecting authenticity of comments | |
| CN108805170A (en) | Form the data set for the study of full supervised | |
| CN104685540B (en) | Image semantic segmentation method and apparatus | |
| CN110021028B (en) | Automatic clothing making method based on clothing style drawing | |
| CN108595558B (en) | Image annotation method based on data equalization strategy and multi-feature fusion | |
| CN110610472A (en) | Computer device and method for realizing classification detection of lung nodule images | |
| Fukuma et al. | A study on feature extraction and disease stage classification for glioma pathology images | |
| CN108595631B (en) | Three-dimensional CAD model double-layer retrieval method based on graph theory | |
| CN111860656B (en) | Classifier training method, device, equipment and storage medium | |
| Shalom et al. | Part Analogies in Sets of Objects. | |
| CN112784905A (en) | Data sample expansion method and device and electronic equipment | |
| Zafari et al. | Resolving overlapping convex objects in silhouette images by concavity analysis and Gaussian process | |
| CN109063271B (en) | Three-dimensional CAD model segmentation method and device based on ultralimit learning machine | |
| US20210150078A1 (en) | Reconstructing an object | |
| Wu et al. | AAGNet: A graph neural network towards multi-task machining feature recognition | |
| CN112685374B (en) | Log classification method and device and electronic equipment | |
| CN111611919A (en) | Road scene layout analysis method based on structured learning | |
| KR102301635B1 (en) | Method of inferring bounding box using artificial intelligence model and computer apparatus of inferring bounding box | |
| Denk et al. | Feature line detection of noisy triangulated CSGbased objects using deep learning | |
| CN113658338A (en) | Point cloud tree monomer segmentation method and device, electronic equipment and storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |