CN114119930A - Three-dimensional model correction method and device based on deep learning and storage medium - Google Patents
Three-dimensional model correction method and device based on deep learning and storage medium Download PDFInfo
- Publication number
- CN114119930A CN114119930A CN202210099204.XA CN202210099204A CN114119930A CN 114119930 A CN114119930 A CN 114119930A CN 202210099204 A CN202210099204 A CN 202210099204A CN 114119930 A CN114119930 A CN 114119930A
- Authority
- CN
- China
- Prior art keywords
- dimensional model
- data
- model
- file
- training
- 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 40
- 238000013135 deep learning Methods 0.000 title claims abstract description 25
- 238000012937 correction Methods 0.000 title claims abstract description 16
- 238000003062 neural network model Methods 0.000 claims abstract description 26
- 238000012549 training Methods 0.000 claims description 77
- 239000013598 vector Substances 0.000 claims description 75
- 230000006870 function Effects 0.000 claims description 66
- 230000004913 activation Effects 0.000 claims description 21
- 238000012360 testing method Methods 0.000 claims description 19
- 230000004048 modification Effects 0.000 claims description 16
- 238000002715 modification method Methods 0.000 claims description 11
- 238000004590 computer program Methods 0.000 claims description 9
- 238000010276 construction Methods 0.000 claims description 7
- 238000004458 analytical method Methods 0.000 claims description 6
- 238000011478 gradient descent method Methods 0.000 claims description 6
- 238000012986 modification Methods 0.000 claims description 4
- 238000007781 pre-processing Methods 0.000 claims description 4
- 238000012795 verification Methods 0.000 claims description 4
- 238000013528 artificial neural network Methods 0.000 description 25
- 210000002569 neuron Anatomy 0.000 description 12
- 230000008569 process Effects 0.000 description 8
- 238000005070 sampling Methods 0.000 description 4
- 238000007476 Maximum Likelihood Methods 0.000 description 3
- 238000004364 calculation method Methods 0.000 description 3
- 239000000203 mixture Substances 0.000 description 3
- 239000003607 modifier Substances 0.000 description 3
- FIRDBEQIJQERSE-QPPQHZFASA-N 2',2'-Difluorodeoxyuridine Chemical compound FC1(F)[C@H](O)[C@@H](CO)O[C@H]1N1C(=O)NC(=O)C=C1 FIRDBEQIJQERSE-QPPQHZFASA-N 0.000 description 2
- 230000008901 benefit Effects 0.000 description 2
- 238000013461 design Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 230000008676 import Effects 0.000 description 2
- 239000011159 matrix material Substances 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 230000008439 repair process Effects 0.000 description 2
- 239000007787 solid Substances 0.000 description 2
- 241001502050 Acis Species 0.000 description 1
- 241000405961 Scomberomorus regalis Species 0.000 description 1
- 238000000627 alternating current impedance spectroscopy Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000009795 derivation Methods 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 238000012886 linear function Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 238000013138 pruning Methods 0.000 description 1
- 238000012887 quadratic function Methods 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T17/00—Three dimensional [3D] modelling, e.g. data description of 3D objects
- G06T17/30—Polynomial surface description
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/047—Probabilistic or stochastic networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2210/00—Indexing scheme for image generation or computer graphics
- G06T2210/32—Image data format
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Computing Systems (AREA)
- Artificial Intelligence (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- General Engineering & Computer Science (AREA)
- Biomedical Technology (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Probability & Statistics with Applications (AREA)
- Algebra (AREA)
- Mathematical Analysis (AREA)
- Mathematical Optimization (AREA)
- Pure & Applied Mathematics (AREA)
- Computer Graphics (AREA)
- Geometry (AREA)
- Processing Or Creating Images (AREA)
Abstract
The invention discloses a three-dimensional model correction method based on deep learning, which comprises the steps of constructing a data set by acquiring three-dimensional models of different types of files and topological geometric data of the three-dimensional model of each type of file; then, a deep neural network model is established, and the deep neural network model is trained according to the data set to obtain a three-dimensional model corrector; and then acquiring the three-dimensional model of the file to be corrected, analyzing the three-dimensional model to obtain corresponding topological geometric data, and matching the topological geometric data of the three-dimensional model of the file to be corrected with the three-dimensional model corrector to realize the correction of the topological geometric data of the three-dimensional model of the file to be corrected. The method can be applied to software for loading the three-dimensional model of the file, and the topological geometric data of the three-dimensional model of the imported file is corrected so that the three-dimensional model of the file can be loaded correctly. The invention also discloses a three-dimensional model correction device based on deep learning and a storage medium.
Description
Technical Field
The invention relates to the field of CAD software import, in particular to a three-dimensional model correction method and device based on deep learning and a storage medium.
Background
For the current mainstream CAD software, it has its own file type, such as Z3 type of file of ZW 3D. In the interaction process of each piece of software, the condition that a file model generated by other pieces of software is required to be imported into CAD software is avoided, so that the CAD software is required to analyze the imported file model to obtain topological data and geometric data, and then the file model is reproduced according to the topological data and the geometric data obtained through analysis. However, some file types of file models may not be analyzed in the CAD software, or some topological geometric data in the file models may have errors, so that the analyzed data of the CAD software cannot be used, and further the file models cannot be loaded and displayed, and errors also exist in subsequent modeling, and therefore, the repair of the topological geometric data of the models imported into the CAD software becomes a problem that needs to be solved at present.
Disclosure of Invention
In order to overcome the defects of the prior art, one of the purposes of the present invention is to provide a three-dimensional model correction method based on deep learning, which can solve the problem that the three-dimensional model of the file imported into the software in the prior art cannot be loaded in the software due to the topology geometric data error.
The second purpose of the present invention is to provide a three-dimensional model modification apparatus based on deep learning, which can solve the problem that the three-dimensional model of the file imported into the software in the prior art cannot be loaded in the software due to the topology geometric data error.
The invention further aims to provide a storage medium, which can solve the problem that a three-dimensional model of a file imported into software in the prior art cannot be loaded in the software due to topological and geometric data errors.
One of the purposes of the invention is realized by adopting the following technical scheme:
the three-dimensional model correction method based on deep learning comprises the following steps:
a data acquisition step: acquiring three-dimensional models of different types of files and topological geometric data of the three-dimensional models of each type of files to construct a data set;
a model construction step: creating a deep neural network model, and training the deep neural network model according to the data set to obtain a three-dimensional model corrector;
and (3) correcting: the method comprises the steps of obtaining a three-dimensional model of a file to be corrected, analyzing the three-dimensional model of the file to be corrected to obtain corresponding topological geometric data, and matching the topological geometric data of the three-dimensional model of the file to be corrected with a three-dimensional model corrector to realize correction of the topological geometric data of the three-dimensional model of the file to be corrected.
Further, the model building step further comprises:
a data set dividing step: dividing the data set into a training set and a test set;
model training: training the built deep neural network model according to the training set by building the deep neural network model to obtain a three-dimensional model corrector;
a model verification step: verifying the trained three-dimensional model corrector according to the test set to judge whether the three-dimensional model corrector meets the preset requirement, and if so, obtaining the three-dimensional model corrector; if not, executing a model training step or a data set dividing step.
Further, when the constructed model is trained in the model training step, a nonlinear activation function is adopted as an activation function of the training model; wherein the nonlinear activation function is any one of the following functions: an igmolar type function, a logistic sigmoidal function, a hyperbolic tangent function, a corrective linear unit function; a random gradient descent method is adopted as a training method.
Further, the topological-geometric data comprises topological data and geometric data; the topology data comprises a topology ensemble, a manifold entity container, an entity, a shell, a face, a ring edge, an edge and a vertex; the geometric data is a Nurbs expression and comprises control vertexes, node vectors and times.
Further, the data acquiring step further comprises: analyzing the three-dimensional model of each type of file, and storing the topological geometric data obtained by analysis in a topological table constructing mode; the topology table is constructed according to the sequence of topology totality, manifold entity container, entity, shell, face, ring edge, edge and vertex, and the information of all ring edges and edges of each face is stored when the face contained in the shell is constructed.
Further, the data acquiring step further comprises: constructing a plurality of input vectors and corresponding output vectors according to topological geometric data of the three-dimensional model of each type of file; the input vector is constructed by taking a surface as a unit and comprises the fact that the surface belongs to an entity or a sheet body, an index of the entity or the sheet body, an index of the surface, the number of outer rings, the number of inner rings, the number of ring edges of each ring, the index of the ring edges, the index of the corresponding edge of the ring edges, the index of head and tail end points of the edge, a Nurbs control vertex of the surface, a node vector and the number of times; the corresponding output vector includes the control vertices, node vectors and the degree of all edges and ring edges of the face.
Further, the data acquiring step further comprises: and after preprocessing each constructed input vector and corresponding output vector, summarizing the input vector and the output vector corresponding to the three-dimensional model of each type of file to form the data set. The second purpose of the invention is realized by adopting the following technical scheme:
a three-dimensional model modification apparatus based on deep learning includes a memory, a processor, and a computer program stored on the memory and running on the processor, the computer program being a three-dimensional model modification program, the processor implementing the steps of the three-dimensional model modification method based on deep learning as one of the objects of the present invention when executing the three-dimensional model modification program.
The third purpose of the invention is realized by adopting the following technical scheme:
a storage medium which is a computer-readable storage medium having stored thereon a computer program which is a three-dimensional model modification program, the three-dimensional model modification program being executed by a processor for performing the steps of a deep learning-based three-dimensional model modification method employed as one of the objects of the present invention.
Compared with the prior art, the invention has the beneficial effects that:
according to the method, the data sets are constructed for the three-dimensional models of the files of different types and the corresponding topological geometric data, and the construction of the three-dimensional model corrector is realized by combining the deep neural network according to the three-dimensional models of the files of different types and the corresponding topological geometric data, so that the method is applied to software to correct the topological geometric data of the three-dimensional models of the imported files, and the problem that the three-dimensional models of the files cannot be loaded in the software due to errors of the topological geometric data is solved.
Drawings
FIG. 1 is a flowchart of a three-dimensional model modification method based on deep learning according to the present invention;
fig. 2 is a flowchart of step S2 in fig. 1.
Detailed Description
The present invention will be further described with reference to the accompanying drawings and the detailed description, and it should be noted that any combination of the embodiments or technical features described below can be used to form a new embodiment without conflict.
Example one
The invention provides a preferred embodiment, a three-dimensional model correction method based on deep learning, as shown in fig. 1, comprising the following steps:
step S1, obtaining three-dimensional models of a plurality of different types of files and topological geometrical data of the three-dimensional models of each type of files and constructing a data set.
The method comprises the steps of collecting three-dimensional models of different types of files created by previous users, and then obtaining corresponding topological geometric data according to the three-dimensional model of each file. The three-dimensional model of the file related by the invention not only relates to the three-dimensional model created by the user, but also can be downloaded from the internet to the existing three-dimensional models of different types of files.
Preferably, the type of the three-dimensional model of the file refers to a format of the three-dimensional model. Generally, the file models generated by different software have different formats, such as the file format generated by currently popular general-purpose software and the file format generated by third-party software. The general software may include any one of the following: IGES, STEP, VDA, Parasolid (X _ T, X _ B), ACIS (SAT), and the like. The third party software may include any of: NX, SolidWorks, Inventor, CatiaV4\ V5\ V6, Cero \ ProE, SolidEdge, etc.
Preferably, the topological geometrical data of the three-dimensional model of the file comprise topological data and geometrical data. Wherein the topology data includes topology universes, manifold entity containers, entities, shells, faces, rings, ring edges, and vertices. More specifically, the topology population is the highest level of topological entities, which represents a space of connected entities, including manifold objects and non-manifold objects. The object may be an entity, a skin, a surface model, or the like. Each topology entity belongs to a topology population, and the topology population contains a container of manifold entities.
The manifold entity container is a manifold topology totality or a manifold part of a non-manifold topology totality and comprises one or more entities. Such as: if two entities have a common edge, they belong to two separate manifold entity containers.
The solid body is a space limited by the connecting edge and comprises one or more shells. For example, if two housings have a common face, two physical representations are required.
A shell is a collection of faces that make up a solid. Wherein, the shell can be external or internal; may be open or closed. An entity may be an outer closed shell comprising an outer open shell or may be an outer closed shell comprising zero, one or more inner closed shells.
The surface is a curved surface formed by one or more rings and is a unit forming the shell. The surface is a topological element, the corresponding geometric element is a curved surface, and the curved surface stores the composition information of the non-uniform rational B-spline surface. Wherein, the expression of the Nurbs of the curved surface is as follows:
The ring edge represents a two-dimensional ring edge on a plane, and is a basic unit constituting a ring. The ring edge is a topological element and the corresponding geometric element is a curve.
An edge, representing a three-dimensional edge located on a face, is defined by two vertices. Edges are topological elements and the corresponding geometric elements are curves. Wherein the Nurbs of the curve is expressed as follows:
A vertex, defining an edge, and representing a 3D point. An edge has two vertices, which may be said to define an edge, the vertices being topological elements and the corresponding geometric elements being points, each point storing coordinate information for a 3D point.
The geometric data is a Nurbs (Non-Uniform Rational B-spline) expression, and comprises a Nurbs curved surface, specifically comprises control vertexes, node vectors and times.
After the file model is imported into the CAD software, the CAD software analyzes the three-dimensional model of the file to obtain topological geometric data, and simultaneously constructs a corresponding topological table. The topology table is used for storing the topological geometric data obtained by analysis. Because the topological geometric data can be wrong due to failure of the original file modeling process or other factors, the three-dimensional model of the file cannot be loaded by the CAD software, the three-dimensional models of files of different types are collected to carry out model training so as to obtain a corrector of the topological geometric data of the three-dimensional model of each type of file, and then the topological geometric data with the mistake is corrected so that the three-dimensional model of the file can be correctly loaded by the CAD software. Therefore, the invention collects the three-dimensional models of different types of files and the topological geometrical data of the three-dimensional model of each type of file to form a data set for subsequent model training.
And step S2, creating a deep neural network model, and training the built deep neural network model according to the data set to obtain the three-dimensional model corrector.
The three-dimensional model corrector is realized in a model training mode so as to be applied to software for correcting topological geometric data of the three-dimensional model.
And step S3, acquiring topological geometric data of the three-dimensional model of the file to be corrected and importing the topological geometric data into a three-dimensional model corrector to correct the topological geometric data of the three-dimensional model of the file to be corrected.
The three-dimensional model of the file to be corrected is analyzed to obtain corresponding topological geometric data, so that the wrong topological geometric data can be corrected, and correct loading of software can be realized. The method is applied to various 3D model loading software, such as CAD software. When the CAD software imports the three-dimensional model of the corresponding file, the three-dimensional model corrector can be applied to correct the topological geometric data, so that the 3D model can be loaded correctly.
Preferably, as shown in fig. 2, step S2 further includes:
and step S21, dividing the data set into a training set and a testing set.
Wherein, the training set is used for model training. And the test set is used for testing the trained model. The data set constructed in the foregoing is divided according to a certain method, for example, by randomly selecting a part of the data as a training set and another part as a testing set.
And S22, constructing a deep neural network model, and training the constructed deep neural network model according to the training set to obtain the three-dimensional model corrector.
The built deep neural network model is trained according to the data in the training set to obtain three-dimensional model modifiers of each type, so that topological geometric data of the three-dimensional model of the file imported into the CAD software can be modified conveniently, the file can be loaded correctly by the CAD software, and the problem that file loading cannot be realized due to errors of the topological geometric data in the three-dimensional model of the file in the prior art is solved.
Preferably, when the trained three-dimensional model corrector may not meet the preset requirement, the method of the invention also verifies the trained three-dimensional model corrector through a test set to verify whether the trained three-dimensional model corrector meets the requirement or not, so as to improve the accuracy of subsequent correction.
Step S23, verifying the three-dimensional model corrector through the test set to judge whether the three-dimensional model corrector meets the preset requirement, if so, executing step S3; if not, returning to the step S22 to continue training the constructed deep neural network model according to the training set, or returning to the step S21 to re-divide the training set and the test set and then train the model.
Preferably, the present invention also constructs the data set by vector construction of topological geometry data of the three-dimensional model for each type of file. In particular, step S1 further includes constructing a number of input vectors and corresponding output vectors for topological geometry data of the three-dimensional model of the constructed file.
Wherein each input vector comprises the following data: whether the surface belongs to the entity or the sheet body, the index of the surface, the number of outer rings, the number of inner rings, the number of ring edges of each ring, the index of the ring edges, the index of the corresponding edges of the ring edges, the index of head and tail end points of the edges, a Nurbs control vertex of the surface, a node vector and the number of times.
And meanwhile, constructing a corresponding output vector according to each input vector, wherein the output vector is the control vertex, the node vector and the times of all edges and ring edges of the plane. And summarizing the plurality of constructed input vectors and the corresponding output vectors to form a data set.
Preferably, the input vector and the output vector in the present invention are both topology data and geometry data taken from the same surface, and the output vector is constrained by expressing the topology and geometry data of the surface through the input vector.
Preferably, in step S1, the topology geometric data obtained by the parsing is also stored in the form of a topology table, that is, the topology table is constructed to store the topology geometric data obtained by the parsing. The topology table is constructed according to a top-down sequence, specifically, the topology table is constructed according to a sequence of a topology total, a manifold entity container, an entity, a shell, a face, a ring edge, an edge and a vertex, and information of all ring edges and edges of each face is stored when the face contained in the shell is constructed.
For example, for a part in a three-dimensional model of a document, a topological population is first stored, and a part corresponds to a topological population. The topology as a whole has no geometrical information for storing the manifold entity container.
Then judging whether the model under the topology population is a manifold entity container, if so, storing a manifold entity container under the topology population; if not, the connection part of the model is divided until each divided model is a manifold entity container; meanwhile, one manifold entity container is created for each model after segmentation, and then each manifold entity container is stored under the topology population.
Storing topological geometry data of an entity for each individual entity under a manifold entity container. The entity comprises an outer shell and an inner shell, and topological geometrical data of one shell is stored for each shell. The housing includes all the faces constituting it, and data of each face is stored under the housing. There may be one ring or multiple rings for each data of the subsurface storage ring. And simultaneously storing the geometric data of the curved surface of the surface, wherein the curved surface stores the composition information of the non-uniform rational B-spline surface. And storing data of a ring edge and an edge under the ring, wherein the ring edge and the edge are a two-dimensional edge and a three-dimensional edge respectively, the ring edge and the edge respectively store the geometrical data of curves corresponding to the ring edge and the edge, and the curves store the composition information of the non-uniform rational B-spline curves.
That is, the topology geometric data obtained by analyzing the three-dimensional model of each imported file is stored by the method for constructing the topology table.
Preferably, in order to ensure the accuracy of model training, when the deep neural network model is constructed, the operations of structural design of the deep neural network involved in model training, allocation of a corresponding number of neurons to each hidden layer, initialization of uncertain parameters, determination of an activation function, and the like are also required.
Preferably, for the design of hyper-parameters of a deep neural network, the deep neural network generally consists of five layers: an input layer, three hidden layers and an output layer. The three hidden layers are respectively used for constructing a topological structure of a simulation model and projecting a three-dimensional edge to obtain the geometric information mapping of a ring edge and an edge.
The invention is based on a fully connected neural network, accepts 1500-dimensional vectors as input, and then extracts and synthesizes more abstract conceptual features by three hidden layers. Thus, if the deep neural network contains more trainable free parameters, the training error rate can become small. Such neural networks have a high expressive power and can be fine-tuned to a specific training set. In this case, however, the error rate of the neural network on the test sample may be unacceptably high, and "overfitting" is likely to occur.
If the deep neural network does not have enough trainable free parameters, "under-fitting" occurs, meaning that for training data the error rate of the fit to the training data is too high and the results of the test data are not good. Therefore, for the number of trainable free parameters, a compromise value needs to be selected to obtain a lower test error rate.
In the maximum likelihood set, the parameter space has an optimal bit of trainable free parameter number, which can bring the best generalization performance to the modifier. This optimum corresponds to the best balance between under-fitting and over-fitting. Based on domain knowledge, experience and experimentation, we found that the ratio of the total training data and the number of free parameters provided to the network was 100: 1, the performance of the corrector is optimal. According to the theory, the invention adopts 1500-28-400-56-180 full neural network configuration (wherein, the number is the neuron number of each layer of neural network model), and the generalization performance of the corrector is optimized as much as possible.
Preferably, the training set used in the present invention contains 5800 model faces in the data, each face being represented by a 1500-dimensional vector, and the total training data volume is 5800 x 1500= 8700000. Therefore, the number of the free parameter sets of the constructed deep neural network modifier is specifically as follows:
1500 × 28+28+28 × 400+400+400 × 56+56 +180= 86344.
Wherein the ratio of the total training data volume and the trainable parameters 8700000/86344= 100.76.
Meanwhile, the invention also tests the neural networks with other structural configurations in the experiment, and the result shows that the generalization performance of the neural networks is not as good as that of the neural networks with the parameters.
Furthermore, another key issue that needs to be carefully considered for the linear elements of the neurons of the hidden layer is weight initialization. If the weight is initialized to be an overlarge value, the deep neural network falls into a poor local minimum value; if the weight is initialized to be too small, the gradient change of the previous layer in the network is very small, so that the deep neural network with a plurality of hidden layers cannot be trained.
In the invention, the initial weight of a certain hidden layer is subjected to hook sampling from a symmetrical interval depending on the specific activated number. Wherein, the logic sigmoid function has a weight sampling interval of
；
Is the number of neurons contained in the i-1 st hidden layer,
is the number of neurons contained in the ith hidden layer.
For the hyperbolic tangent function, the weight sampling interval is
。
This way of weight initialization can ensure that, in the early stage of training, the activation function runs in its functional interval, and the information (function signal and error signal) in the deep neural network can be easily transmitted in the forward direction (activation signal flow from input to output) and in the backward direction (gradient signal flow from output to input).
Preferably, the activation function for a deep neural network should generally satisfy the following properties: (1) non-linear, otherwise the deep neural network will not provide any computational power above the two-layer network.
(2) Saturation, i.e. there are maximum and minimum output values, which may define the upper and lower boundaries of the weights and activation functions, thus making the number of training times also limited.
(3) Continuity and smoothness, i.e., the activation function and its derivative function, are defined over their entire range of arguments. The existence of the derivative function of the activation function is crucial to the derivation of the back propagation learning rule, so that a threshold function and a coincidence function cannot be selected; piecewise linear functions add computational complexity and do not provide additional benefits.
(4) Monotonicity. We expect the derivative function to have the same sign throughout the argument range, and if not monotonic and have multiple local maxima, will introduce additional and undesirable extrema on the error surface.
(5) The linear characteristic is provided when the net activation value is small, which enables the system to realize a linear model when the error is low.
Preferably, the present invention generally selects the nonlinear activation function as a sigmoid-type function, a logistic sigmoid function, a hyperbolic tangent function, or a corrective linear unit function.
Preferably, the present invention selects a non-negative log-likelihood function as the error function. The learning process is equivalent to a model based on a sample space (training set D) in a parameter space of the model
Performing maximum likelihood estimation. The likelihood function L is as follows:
wherein
All trainable free parameters of the classifier are represented, D represents the training set, W is the weight matrix, b is the bias vector, and { output } is the output direction of the ith plane in the training setQuantity, { input } is the input vector for the ith plane in the training set.
Preferably, the error function used to train the deep neural network is as follows:
Preferably, "weight attenuation" is a heuristic rule that controls the complexity of the deep neural network to avoid overfitting. This rule forces some neuron weights in the deep neural network to take values that are approximately zero, while allowing other weights to retain their relatively large values. Therefore, the neuron weights in the network are roughly divided into two categories: neuron weight which has great influence on network performance; neuron weights that have little or no impact on network performance are also referred to as excess weights. Without complexity regularization and network pruning, these neuron weights are likely to take entirely arbitrary values, or to yield slightly reduced training errors to force the network to over-fit the training data, resulting in poor generalization performance.
Weight attenuation penalizes those unnecessary weights by adding a regularization term to the error function.
In the present invention, the error function is redefined as follows:
The error function takes a maximum likelihood method as a method for evaluating errors, has a theoretical basis of probability and mathematical statistics, can better reflect error conditions, has better mathematical properties, is continuous and differentiable, and can be used for training a correction weight of deep neural network back propagation.
Preferably, in the model training, the calculation method of the hidden layer in the propagation process of the invention is as follows:
. Wherein, H is the output result of a certain hidden layer, W is the weight matrix of the two-layer neuron transformation, b is the bias vector of the two-layer neuron, and v is the input vector of the previous layer.
The invention adopts a random gradient descent method as a training method, and the specific formula is as follows:
Compared with other optimization methods, the random gradient descent optimization method has the advantages that the convergence speed of the training process of the three-dimensional model corrector of the file is higher, the fluctuation of the training process is smaller, and the repair accuracy of the three-dimensional model of the file is higher.
Preferably, in the model training process, generally speaking, as long as the learning rate is small enough to guarantee convergence, its value merely determines the deep neural network arrival error function
The velocity of the minimum value of (c), and does not determine the final weight value. In practice, however, the learning rate may actually affect the final network performance, since deep neural networks are rarely trained sufficiently to actually minimize the error.
One basic method of setting the learning rate of the present invention is to assume an error function
Can reasonably be approximated as a quadratic function, thus giving:
。
the optimal learning rate can be directly found from the following formula:
。
the optimal learning rate calculated by the above formula is adopted as the learning rate of the random gradient descent method in the invention. The vibration can be avoided, the divergence can be avoided, and the training speed is not too slow.
Preferably, the present invention also preprocesses the input vectors and the output vectors prior to performing model training. Wherein, the input vector is constructed by taking a surface as a unit; and the output vector is the geometric data of the edge and the ring edge, namely the Nurbs curve data. In order to ensure the consistency of data dimensions and better ensure that a trained three-dimensional model corrector has stronger generalization performance, perspective projection is carried out on a Nurbs curve of three-dimensional geometric data comprising a Nurbs curved surface and a Nurbs side. After perspective projection, all data are unified into two-dimensional data, the supervision effect that the two-dimensional geometric data are ignored due to the fact that the training process excessively tends to the three-dimensional geometric data due to the fact that the difference of mathematical properties of input data is large is avoided, and the effects of input vectors and output vectors in supervision learning are exerted as far as possible.
The geometric data of the annular edge is obtained by projecting the three-dimensional curve data of the edge on a plane. Because the geometric data of the annular edge is a two-dimensional Nurbs curve, the annular edge can be obtained by projection calculation of the Nurbs curve of the three-dimensional edge on a curved surface, and the method specifically comprises the following steps:
(1) and uniformly sampling 1500 points on the three-dimensional Nurbs curve, and recording the coordinates of the three-dimensional points.
(2) And projecting each three-dimensional point to the Nurbs curved surface, and calculating to obtain the coordinates of the two-dimensional points in the curved surface parameter space. For the projected point (u, v) on the surface, the following condition must be satisfied:
Suppose that:
for the course of projection, we know the initial point of projection (u)0, v0) Then, only dU and dV need be found to obtain the proxels (u, v).
Therefore, will
Substituting into formula (1), then expanding according to Taylor's formula of the binary function to obtain formula (2):
for equation (2), if the other terms than the first two terms are set to 0, equation (2) can be converted into equation (3):
will be provided with
Then equation (3) can yield equation (4) and equation (5):
substituting equation (4) and equation (5) into equation (3) yields:
order to
Then, equation (6) is obtained:
since zRHS, dFdU, dFdV are all vectors, for the convenience of calculation, if both sides of equation (6) are multiplied by dFdU and dFdV, respectively, then equation (7) and equation (8) can be obtained:
solving by combining the formula (7) and the formula (8) as an equation set to obtain dU and dV, and further carrying out
The projection point (u, v) is obtained.
(3) And constructing a two-dimensional Nurbs curve, namely the geometric data of the two-dimensional ring edge, by using all the calculated two-dimensional uv points through an interpolation method, so that an output vector { output } can be generated.
Example two
Based on the first embodiment, the present invention further provides a deep learning based three-dimensional model modification apparatus, including a memory, a processor, and a computer program stored in the memory and running on the processor, where the computer program is a three-dimensional model modification program, and the processor implements the following steps when executing the three-dimensional model modification program:
a data acquisition step: acquiring three-dimensional models of different types of files and topological geometric data of the three-dimensional models of each type of files to construct a data set;
a model construction step: creating a deep neural network model, and training the deep neural network model according to the data set to obtain a three-dimensional model corrector;
and (3) correcting: the method comprises the steps of obtaining a three-dimensional model of a file to be corrected, analyzing the three-dimensional model of the file to be corrected to obtain corresponding topological geometric data, and matching the topological geometric data of the three-dimensional model of the file to be corrected with a three-dimensional model corrector to realize correction of the topological geometric data of the three-dimensional model of the file to be corrected.
Further, the model building step further comprises:
a data set dividing step: dividing the data set into a training set and a test set;
model training: training the built deep neural network model according to the training set by building the deep neural network model to obtain a three-dimensional model corrector;
a model verification step: verifying the trained three-dimensional model corrector according to the test set to judge whether the three-dimensional model corrector meets the preset requirement, and if so, obtaining the three-dimensional model corrector; if not, executing a model training step or a data set dividing step.
Further, when the constructed model is trained in the model training step, a nonlinear activation function is adopted as an activation function of the training model; wherein the nonlinear activation function is any one of the following functions: an igmolar type function, a logistic sigmoidal function, a hyperbolic tangent function, a corrective linear unit function; a random gradient descent method is adopted as a training method.
Further, the topological-geometric data comprises topological data and geometric data; the topology data comprises a topology ensemble, a manifold entity container, an entity, a shell, a face, a ring edge, an edge and a vertex; the geometric data is a Nurbs expression and comprises control vertexes, node vectors and times.
Further, the data acquiring step further comprises: analyzing the three-dimensional model of each type of file, and storing the topological geometric data obtained by analysis in a topological table constructing mode; the topology table is constructed according to the sequence of topology totality, manifold entity container, entity, shell, face, ring edge, edge and vertex, and the information of all ring edges and edges of each face is stored when the face contained in the shell is constructed.
Further, the data acquiring step further comprises: constructing a plurality of input vectors and corresponding output vectors according to topological geometric data of the three-dimensional model of each type of file; the input vector is constructed by taking a surface as a unit and comprises the fact that the surface belongs to an entity or a sheet body, an index of the entity or the sheet body, an index of the surface, the number of outer rings, the number of inner rings, the number of ring edges of each ring, the index of the ring edges, the index of the corresponding edge of the ring edges, the index of head and tail end points of the edge, a Nurbs control vertex of the surface, a node vector and the number of times; the corresponding output vector includes the control vertices, node vectors and the degree of all edges and ring edges of the face.
Further, the data acquiring step further comprises: and after preprocessing each constructed input vector and corresponding output vector, summarizing the input vector and the output vector corresponding to the three-dimensional model of each type of file to form the data set.
EXAMPLE III
A storage medium which is a computer-readable storage medium having stored thereon a computer program which is a three-dimensional model modification program, the three-dimensional model modification program being executed by a processor to perform the steps of:
a data acquisition step: acquiring three-dimensional models of different types of files and topological geometric data of the three-dimensional models of each type of files to construct a data set;
a model construction step: creating a deep neural network model, and training the deep neural network model according to the data set to obtain a three-dimensional model corrector;
and (3) correcting: the method comprises the steps of obtaining a three-dimensional model of a file to be corrected, analyzing the three-dimensional model of the file to be corrected to obtain corresponding topological geometric data, and matching the topological geometric data of the three-dimensional model of the file to be corrected with a three-dimensional model corrector to realize correction of the topological geometric data of the three-dimensional model of the file to be corrected.
Further, the model building step further comprises:
a data set dividing step: dividing the data set into a training set and a test set;
model training: training the built deep neural network model according to the training set by building the deep neural network model to obtain a three-dimensional model corrector;
a model verification step: verifying the trained three-dimensional model corrector according to the test set to judge whether the three-dimensional model corrector meets the preset requirement, and if so, obtaining the three-dimensional model corrector; if not, executing a model training step or a data set dividing step.
Further, when the constructed model is trained in the model training step, a nonlinear activation function is adopted as an activation function of the training model; wherein the nonlinear activation function is any one of the following functions: an igmolar type function, a logistic sigmoidal function, a hyperbolic tangent function, a corrective linear unit function; a random gradient descent method is adopted as a training method.
Further, the topological-geometric data comprises topological data and geometric data; the topology data comprises a topology ensemble, a manifold entity container, an entity, a shell, a face, a ring edge, an edge and a vertex; the geometric data is a Nurbs expression and comprises control vertexes, node vectors and times.
Further, the data acquiring step further comprises: analyzing the three-dimensional model of each type of file, and storing the topological geometric data obtained by analysis in a topological table constructing mode; the topology table is constructed according to the sequence of topology totality, manifold entity container, entity, shell, face, ring edge, edge and vertex, and the information of all ring edges and edges of each face is stored when the face contained in the shell is constructed.
Further, the data acquiring step further comprises: constructing a plurality of input vectors and corresponding output vectors according to topological geometric data of the three-dimensional model of each type of file; the input vector is constructed by taking a surface as a unit and comprises the fact that the surface belongs to an entity or a sheet body, an index of the entity or the sheet body, an index of the surface, the number of outer rings, the number of inner rings, the number of ring edges of each ring, the index of the ring edges, the index of the corresponding edge of the ring edges, the index of head and tail end points of the edge, a Nurbs control vertex of the surface, a node vector and the number of times; the corresponding output vector includes the control vertices, node vectors and the degree of all edges and ring edges of the face.
Further, the data acquiring step further comprises: and after preprocessing each constructed input vector and corresponding output vector, summarizing the input vector and the output vector corresponding to the three-dimensional model of each type of file to form the data set.
The above embodiments are only preferred embodiments of the present invention, and the protection scope of the present invention is not limited thereby, and any insubstantial changes and substitutions made by those skilled in the art based on the present invention are within the protection scope of the present invention.
Claims (9)
1. The three-dimensional model correction method based on deep learning is characterized by comprising the following steps of:
a data acquisition step: acquiring three-dimensional models of different types of files and topological geometric data of the three-dimensional models of each type of files to construct a data set;
a model construction step: creating a deep neural network model, and training the deep neural network model according to the data set to obtain a three-dimensional model corrector;
and (3) correcting: the method comprises the steps of obtaining a three-dimensional model of a file to be corrected, analyzing the three-dimensional model of the file to be corrected to obtain corresponding topological geometric data, and matching the topological geometric data of the three-dimensional model of the file to be corrected with a three-dimensional model corrector to realize correction of the topological geometric data of the three-dimensional model of the file to be corrected.
2. The deep learning-based three-dimensional model modification method according to claim 1, wherein the model construction step further comprises:
a data set dividing step: dividing the data set into a training set and a test set;
model training: training the built deep neural network model according to the training set by building the deep neural network model to obtain a three-dimensional model corrector;
a model verification step: verifying the trained three-dimensional model corrector according to the test set to judge whether the three-dimensional model corrector meets the preset requirement, and if so, obtaining the three-dimensional model corrector; if not, executing a model training step or a data set dividing step.
3. The deep learning-based three-dimensional model modification method according to claim 2, wherein in the model training step, when the constructed model is trained, a nonlinear activation function is used as an activation function of the training model; wherein the nonlinear activation function is any one of the following functions: an igmolar type function, a logistic sigmoidal function, a hyperbolic tangent function, a corrective linear unit function; a random gradient descent method is adopted as a training method.
4. The deep learning-based three-dimensional model modification method according to claim 1, wherein the topological geometric data includes topological data and geometric data; the topology data comprises a topology ensemble, a manifold entity container, an entity, a shell, a face, a ring edge, an edge and a vertex; the geometric data is a Nurbs expression and comprises control vertexes, node vectors and times.
5. The deep learning-based three-dimensional model modification method according to claim 4, wherein the data acquisition step further comprises: analyzing the three-dimensional model of each type of file, and storing the topological geometric data obtained by analysis in a topological table constructing mode; the topology table is constructed according to the sequence of topology totality, manifold entity container, entity, shell, face, ring edge, edge and vertex, and the information of all ring edges and edges of each face is stored when the face contained in the shell is constructed.
6. The deep learning-based three-dimensional model modification method according to claim 5, wherein the data acquisition step further comprises: constructing a plurality of input vectors and corresponding output vectors according to topological geometric data of the three-dimensional model of each type of file; the input vector is constructed by taking a surface as a unit and comprises the fact that the surface belongs to an entity or a sheet body, an index of the entity or the sheet body, an index of the surface, the number of outer rings, the number of inner rings, the number of ring edges of each ring, the index of the ring edges, the index of the corresponding edge of the ring edges, the index of head and tail end points of the edge, a Nurbs control vertex of the surface, a node vector and the number of times; the corresponding output vector includes the control vertices, node vectors and the degree of all edges and ring edges of the face.
7. The deep learning-based three-dimensional model modification method according to claim 6, wherein the data acquisition step further comprises: and after preprocessing each constructed input vector and corresponding output vector, summarizing the input vector and the output vector corresponding to the three-dimensional model of each type of file to form the data set.
8. Deep learning based three-dimensional model modification apparatus comprising a memory, a processor and a computer program stored on the memory and running on the processor, the computer program being a three-dimensional model modification program, characterized in that the processor implements the steps of the deep learning based three-dimensional model modification method according to any one of claims 1 to 7 when executing the three-dimensional model modification program.
9. A storage medium which is a computer-readable storage medium having stored thereon a computer program which is a three-dimensional model modification program, characterized in that: the three-dimensional model modification program is executed by a processor to perform the steps of the deep learning based three-dimensional model modification method according to any one of claims 1 to 7.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210099204.XA CN114119930B (en) | 2022-01-27 | 2022-01-27 | Three-dimensional model correction method and device based on deep learning and storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210099204.XA CN114119930B (en) | 2022-01-27 | 2022-01-27 | Three-dimensional model correction method and device based on deep learning and storage medium |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114119930A true CN114119930A (en) | 2022-03-01 |
| CN114119930B CN114119930B (en) | 2022-10-28 |
Family
ID=80361455
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210099204.XA Active CN114119930B (en) | 2022-01-27 | 2022-01-27 | Three-dimensional model correction method and device based on deep learning and storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114119930B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116502548A (en) * | 2023-06-29 | 2023-07-28 | 湖北工业大学 | Three-dimensional toy design method based on deep learning |
Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6804416B1 (en) * | 2001-03-16 | 2004-10-12 | Cognex Corporation | Method and system for aligning geometric object models with images |
| CN1614593A (en) * | 2004-10-14 | 2005-05-11 | 上海交通大学 | Similarity estimating method for three-dimensional CAD model based on shape |
| CN108520553A (en) * | 2018-03-30 | 2018-09-11 | 苏州喆堃智能科技有限公司 | The drawing method of Faro point cloud datas based on AutoCAD |
| CN110245249A (en) * | 2019-05-15 | 2019-09-17 | 西安交通大学 | A kind of three-dimensional CAD model intelligent search method based on the double-deck depth residual error network |
| CN110378047A (en) * | 2019-07-24 | 2019-10-25 | 哈尔滨工业大学 | A kind of Longspan Bridge topology ambiguity three-dimensional rebuilding method based on computer vision |
| EP3757903A1 (en) * | 2019-06-26 | 2020-12-30 | Robert Bosch GmbH | Optimizing geometry of physical structure using neural network |
| CN112699915A (en) * | 2020-12-07 | 2021-04-23 | 杭州电子科技大学 | Method for identifying CAD model assembly interface based on improved drawing attention network |
| CN113160411A (en) * | 2021-04-23 | 2021-07-23 | 杭州电子科技大学 | Indoor three-dimensional reconstruction method based on RGB-D sensor |
| WO2021150215A1 (en) * | 2020-01-22 | 2021-07-29 | Hewlett-Packard Development Company, L.P. | Geometrical compensation models |
| US20210390219A1 (en) * | 2019-04-30 | 2021-12-16 | Hewlett-Packard Development Company, L.P. | Geometrical compensation models |
| WO2021257094A1 (en) * | 2020-06-19 | 2021-12-23 | Hewlett-Packard Development Company, L.P. | Point cloud alignment |
-
2022
- 2022-01-27 CN CN202210099204.XA patent/CN114119930B/en active Active
Patent Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6804416B1 (en) * | 2001-03-16 | 2004-10-12 | Cognex Corporation | Method and system for aligning geometric object models with images |
| CN1614593A (en) * | 2004-10-14 | 2005-05-11 | 上海交通大学 | Similarity estimating method for three-dimensional CAD model based on shape |
| CN108520553A (en) * | 2018-03-30 | 2018-09-11 | 苏州喆堃智能科技有限公司 | The drawing method of Faro point cloud datas based on AutoCAD |
| US20210390219A1 (en) * | 2019-04-30 | 2021-12-16 | Hewlett-Packard Development Company, L.P. | Geometrical compensation models |
| CN110245249A (en) * | 2019-05-15 | 2019-09-17 | 西安交通大学 | A kind of three-dimensional CAD model intelligent search method based on the double-deck depth residual error network |
| EP3757903A1 (en) * | 2019-06-26 | 2020-12-30 | Robert Bosch GmbH | Optimizing geometry of physical structure using neural network |
| CN110378047A (en) * | 2019-07-24 | 2019-10-25 | 哈尔滨工业大学 | A kind of Longspan Bridge topology ambiguity three-dimensional rebuilding method based on computer vision |
| WO2021150215A1 (en) * | 2020-01-22 | 2021-07-29 | Hewlett-Packard Development Company, L.P. | Geometrical compensation models |
| WO2021257094A1 (en) * | 2020-06-19 | 2021-12-23 | Hewlett-Packard Development Company, L.P. | Point cloud alignment |
| CN112699915A (en) * | 2020-12-07 | 2021-04-23 | 杭州电子科技大学 | Method for identifying CAD model assembly interface based on improved drawing attention network |
| CN113160411A (en) * | 2021-04-23 | 2021-07-23 | 杭州电子科技大学 | Indoor three-dimensional reconstruction method based on RGB-D sensor |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116502548A (en) * | 2023-06-29 | 2023-07-28 | 湖北工业大学 | Three-dimensional toy design method based on deep learning |
| CN116502548B (en) * | 2023-06-29 | 2023-09-15 | 湖北工业大学 | Three-dimensional toy design method based on deep learning |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114119930B (en) | 2022-10-28 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111652321B (en) | Marine ship detection method based on improved YOLOV3 algorithm | |
| WO2021254499A1 (en) | Editing model generation method and apparatus, face image editing method and apparatus, device, and medium | |
| Sheng et al. | Improving video temporal consistency via broad learning system | |
| CN109977223B (en) | Method for classifying papers by using capsule mechanism-fused graph convolution network | |
| CN107679562B (en) | Analysis processing method and device for three-dimensional model | |
| CN111352965B (en) | Training method of sequence mining model, and processing method and equipment of sequence data | |
| Wei et al. | Feature preserving mesh simplification using feature sensitive metric | |
| CN114119930B (en) | Three-dimensional model correction method and device based on deep learning and storage medium | |
| WO2024016739A1 (en) | Method for training neural network model, electronic device, cloud, cluster, and medium | |
| CN111460193A (en) | Three-dimensional model classification method based on multi-mode information fusion | |
| CN115204171A (en) | Document-level event extraction method and system based on hypergraph neural network | |
| JP7488375B2 (en) | Method, device and computer-readable storage medium for generating neural networks | |
| Hu et al. | IMMAT: Mesh reconstruction from single view images by medial axis transform prediction | |
| Gao | Game-theoretic approaches for generative modeling | |
| WO2021046681A1 (en) | Complex scenario-oriented multi-source target tracking method | |
| CN110717402A (en) | Pedestrian re-identification method based on hierarchical optimization metric learning | |
| Wu et al. | Active 3-D shape cosegmentation with graph convolutional networks | |
| CN114880527A (en) | Multi-modal knowledge graph representation method based on multi-prediction task | |
| US20210303923A1 (en) | Cleaning dataset for neural network training | |
| CN115131605A (en) | Structure perception graph comparison learning method based on self-adaptive sub-graph | |
| Zhang et al. | MeshLink: a surface structured mesh generation framework to facilitate automated data linkage | |
| CN113344202A (en) | Novel deep multi-core learning network model training method, system and medium | |
| CN114861863A (en) | Heterogeneous graph representation learning method based on meta-path multi-level graph attention network | |
| CN113792459A (en) | Mesh fairing optimization method and device for finite element triangular mesh | |
| KR102494222B1 (en) | Method for creating 3d eyebrows automatically |
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 |