CN110245249A - A kind of three-dimensional CAD model intelligent search method based on the double-deck depth residual error network - Google Patents
A kind of three-dimensional CAD model intelligent search method based on the double-deck depth residual error network Download PDFInfo
- Publication number
- CN110245249A CN110245249A CN201910401751.7A CN201910401751A CN110245249A CN 110245249 A CN110245249 A CN 110245249A CN 201910401751 A CN201910401751 A CN 201910401751A CN 110245249 A CN110245249 A CN 110245249A
- Authority
- CN
- China
- Prior art keywords
- view
- network
- dimensional cad
- cad model
- residual error
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/53—Querying
- G06F16/532—Query formulation, e.g. graphical querying
Abstract
The invention discloses a kind of three-dimensional CAD model intelligent search methods based on the double-deck depth residual error network, using depth residual error network as basic framework, construct the double-deck depth residual error network including screen and sorting network.Training dataset 1 and training dataset 2 are made by obtaining enterprise's history three-dimensional CAD model entity view/wire frame view line number Data preprocess of going forward side by side, data set 1 is further used for the parameter training of screen, data set 2 is used for the parameter training of sorting network, trained network is saved as into .h5 file.The trained double-deck depth residual error network is called, user can realize the precise search to similar three-dimensional CAD model by the entity view of three-dimensional CAD model, wire frame view, engineering drawing, engineering sketch etc..Three-dimensional CAD model intelligent search method proposed by the present invention has the characteristics that retrieval input is flexible, retrieval time is short, retrieval precision is high, provides effective support for the intelligent retrieval and efficient reuse of enterprise's magnanimity three-dimensional CAD model.
Description
Technical field
The invention belongs to advanced manufacturing technology intelligent information technology fields, and in particular to one kind is based on the double-deck depth residual error
The three-dimensional CAD model intelligent search method of network.
Background technique
According to statistics, in enterprise's New Product Development, about 40% components can use ready-made design, about 40%
Components can be completed by modifying existing design, and only about 20% components need completely new design.Therefore, about
The successful case that reuse enterprise is existing in 80% product component R&D process, is demonstrated by engineering practice can be greatly
The q&r of product is improved, and reduces by about 60% research and development of products time, to substantially reduce grinding for enterprise's new product
Cost is sent out, its market competitiveness is improved.It with the booming of cad technique and is widely applied, enterprise has accumulated the three-dimensional of magnanimity
The knowledge of CAD model, these models and its carrying provides case resource abundant for the research and development of new product.In new-product development
The existing three-dimensional CAD model of enterprise and its embedded technical papers, frock clamp and product production plan etc. are reused in the process
The quality and efficiency of research and development, the R&D cycle for shortening product of product not only can be improved, moreover it is possible to promote enterprise personnel in knowledge
It practises, it is inspired to be innovated again, and then promote the innovation ability and core competitiveness of enterprise.However, enterprise is effective due to lacking
Three-dimensional CAD model gopher, cause a large amount of three-dimensional CAD models to be sunk into sleep in enterprise database, thus can not be in new product
It plays a role in R&D process.
In view of the above-mentioned problems, academia expands numerous studies with reuse to three-dimensional CAD model retrieval, and achieve certain
Beneficial effect.The research of current related three-dimensional CAD model retrieval both at home and abroad is focused primarily on using semantic descriptions, shape description
Symbol, geometry descriptor etc. characterize the model；And further by calculating three-dimensional CAD model in input descriptor and database
The similitude or otherness of corresponding descriptor achieve the purpose that three-dimensional CAD model is retrieved.Although the above method can be compared
More accurate search result, but due to the building of descriptor complexity, cause the enforcement difficulty of the above method big；Secondly, for enterprise
For industry employee, descriptor is complicated and indigestion, practicability are poor；It is calculated in database finally, such method needs to be traversed for
The similarity of all descriptors and input descriptor, causes retrieval time longer, and retrieval time can be in enterprise database
Three-dimensional CAD model increase and increase.It can be seen that enterprise there is still a need for it is a kind of input it is flexible and convenient, practical, accuracy is high
Three-dimensional CAD model search method.
Substantially, three-dimensional CAD model view (such as entity view, wire frame view, engineering drawing, engineering sketch) is as enterprise
The important medium of industry employee expression and exchange research and development of products thinking is the ideal chose of three-dimensional CAD model retrieval input.Meanwhile
Three-dimensional CAD model view contains model geometry abundant and shape information, can be used to characterize the model and by itself and
Other three-dimensional CAD models distinguish.On the other hand, as deep learning is in field of image recognition (such as recognition of face, handwritten word
Identification etc.) extensive use, especially depth residual error network achieves in tasks such as ImageNet image classification, detection and positioning
Accuracy rate more higher than manual identified, indicates that deep learning achieves breakthrough in field of image recognition.
Summary of the invention
In view of the above-mentioned deficiencies in the prior art, the technical problem to be solved by the present invention is that providing a kind of based on bilayer
The three-dimensional CAD model intelligent search method of depth residual error network provides branch for the retrieval and reuse of enterprise's magnanimity three-dimensional CAD model
Support.
The invention adopts the following technical scheme:
A kind of three-dimensional CAD model intelligent search method based on the double-deck depth residual error network is made using depth residual error network
For basic framework, building includes the double-deck depth residual error network of screen FilterNet and sorting network RankNet, filtering
The three-dimensional CAD model unrelated with input, sorting network are tentatively excluded when network FilterNet is retrieved for three-dimensional CAD model
RankNet is for being ranked up three-dimensional CAD model relevant to input by similarity and being exported to user；By obtaining enterprise
History three-dimensional CAD model entity view/wire frame view is gone forward side by side line number Data preprocess production training dataset 1 and training dataset 2,
Then data set 1 is used for the parameter training of screen FilterNet, data set 2 is used for the ginseng of sorting network RankNet
Number training, saves as .h5 file for trained network；The trained double-deck depth residual error network is called, three-dimensional CAD mould is passed through
Entity view, wire frame view, engineering drawing, the engineering sketch of type realize the precise search to similar three-dimensional CAD model.
Specifically, depth residual error network overall structure includes view preprocessing module, residual error module and post-processing module structure
At, wherein residual error module O (x) are as follows:
O (x)=max (0, F (x, { Wi})+Wsx)
Wherein, x is the output of a upper layer network, F (x, { Wi) it is that residual error learns, WiFor the weight parameter of residual error module,
WsX is identical mapping, WsFor square matrix, for matching x and F (x, { Wi) dimension.
Specifically, entity view is defined as three-dimensional CAD model with the view obtained after the colouring of sideline, wire frame view definition is
The view obtained after three-dimensional CAD model Hidden line elimination is characterized based on the three-dimensional CAD model of entity view and wire frame view are as follows:
Wherein U is training dataset model sum) indicate that some is given
Three-dimensional CAD model, affiliated model classification are cw,W is the model class of training data lump
Not,Indicate the entity view set of the model,Indicate the wire frame view set of the model, entity view and wire frame view
The number of figure is n, and pixel size is k × l.
Specifically, pretreatment includes scaling, greyscale transformation and normalized, to the entity view and wire frame of acquisition
It is as follows that view does scaling:
Wherein, [x0 y0It 1] is the point on entity view or wire frame view, [x1 y1After 1] being scaled for the scale of view
To new view on corresponding point, and the pixel size of new view is p × q；
It is as follows using weighting method progress greyscale transformation to the view after scaling:
V is the picture element matrix after view greyscale transformation, VR、VG、VBFor three Color Channels pair of red, green, blue before view transformation
The picture element matrix answered, the dimension of above-mentioned picture element matrix are p × q；
Data set normalization is as follows:
Wherein, m=2 × n × U is total number of views.
Specifically, it is training dataset 1, expression formula that pretreated view, which is pressed affiliated model category division, are as follows:
Wherein,Expression model classification is cwThe view set of the three-dimensional CAD model of (such as hydraulic pump), and institute in the set
The label for having view is cw。
Specifically, data set 1 to be imported into the parameter of screen FilterNet and training network, work as screen
When FilterNet reaches the training the number of iterations K of setting, training process terminates, and the screen FilterNet after training is protected
FilterNet.h5 file is saved as, is commented in screen FilterNet network parameter training process using softmax loss function
Error between valence input and output, expression formula are as follows:
Wherein, n is the quantity for most inputting view in small batches,For most in small batches input view i-th view label value,
y(i)Pass through the output valve after the network forward-propagating of depth residual error for the view；
During network parameter W training, improved depth residual error network optimizes loss function using stochastic gradient descent algorithm,
Stochastic gradient descent method expression formula are as follows:
Wi+1:=Wi+vi+1
Wherein, i is iteration index, and τ is momentum, and v is momentum variable, and μ is weight loss, and ε is learning rate.
Specifically, it is training dataset 2, expression formula that pretreated view, which is pressed affiliated model partition, are as follows:
Wherein,Indicate modelThe set of all views, and the label of the view in the set is
Specifically, sorting network RankNet network parameter is initialized using screen FilterNet network parameter, it will
Data set 2 imported into the parameter of sorting network RankNet and training network；Sorting network RankNet network parameter training process
The middle error evaluated using center-softmax loss function between input and output, expression formula are as follows:
Wherein, xiIndicate the depth of i-th most small quantities of (mini-batch) input view of the full articulamentum output of RankNet
Feature, and the view belongs to three-dimensional CAD modelView,Indicate three-dimensional CAD modelThe depth characteristic of all views
Center, λ indicate center loss softmax loss between weight；Sorting network RankNet network parameter W training period
Between, loss function is optimized using stochastic gradient descent algorithm.
Further, it is less than threshold gamma when the average trained accuracy rate of continuous 200 iteration of sorting network RankNet is promoted
When, training process terminates, and the sorting network RankNet after training is saved as RankNet.h5 file；Sorting network RankNet
The average trained accuracy rate of continuous 200 iteration promotes expression formula are as follows:
Wherein, AiThe training accuracy rate of RankNet after expression i-th iteration；Training terminates Rule of judgment are as follows:
Specifically, first pre-processing input view, then the trained double-deck depth residual error network passes through input
N view computation databases in three-dimensional CAD model similarity expression formula are as follows:
Wherein, FpAnd RsRespectively screen Filternet and sorting network RankNet is directed to the output of i-th view
Matrix, WpFor unit matrix, for matching FpAnd RsDimension,Representing matrix dot product,It indicates and input view
The highest M three-dimensional CAD model of similarity is pushed to user as search result.
Compared with prior art, the present invention at least has the advantages that
A kind of three-dimensional CAD model intelligent search method based on the double-deck depth residual error network of the present invention, using depth residual error
Network constructs the double-deck depth residual error network, including upper level filter network and lower layer's sorting network, upper layer mistake as basic framework
Tentatively exclude the three-dimensional CAD model unrelated with input when strainer network is retrieved for three-dimensional CAD model, lower layer's sorting network for into
One step pair three-dimensional CAD model relevant to input is ranked up by similarity and is exported to user.By above-mentioned system integrating, help
In the retrieval time for reducing three-dimensional CAD model, retrieval rate is improved；By obtaining enterprise's history three-dimensional CAD model entity view
Figure/wire frame view is gone forward side by side line number Data preprocess production training dataset 1 and training dataset 2, further uses training dataset 1
In the parameter training of upper level filter network, training dataset 2 is used for the parameter training of lower layer's sorting network, by trained net
Network saves as .h5 file, when three-dimensional CAD model is retrieved, calls the trained double-deck depth residual error network, user can be by defeated
Entity view, wire frame view, engineering drawing, the engineering sketch isometric drawing for entering three-dimensional CAD model are realized to similar three-dimensional CAD model
Precise search.The three-dimensional CAD model intelligent search method based on the double-deck depth residual error network proposed has retrieval input spirit
Living, the features such as retrieval time is short, retrieval precision is high, provides for the intelligent retrieval of enterprise's magnanimity three-dimensional CAD model with efficiently reusing
Effectively support.
Further, the view acquisition of three-dimensional CAD model is divided into two steps, first is that by after the sideline colouring of three-dimensional CAD model band
Form new model, by the model in n virtual camera of space reasonable Arrangement, it is automatic using Blender software tool
Obtain n entity views of the model；Second is that new model will be formed after three-dimensional CAD model Hidden line elimination, and use and reality
Stereogram obtains the n bracing cable frame view that identical mode obtains the model.The purpose of entity view and wire frame view is obtained simultaneously
It is to obtain model information as much as possible from multiple space angles of two kinds of three-dimensional CAD model different display states, to mention
It is flexible defeated when rising the double-deck depth residual error network to the understandability of three-dimensional CAD model, and then three-dimensional CAD model being supported to retrieve
Enter, and improves retrieval rate.
Further, entity view and wire frame view are carried out by scaling, greyscale transformation and normalized pre-
Processing, wherein entity view and wire frame view are converted into the picture of unified size by scaling, so that supported bilayer depth is residual
The parameter training of poor network；Greyscale transformation by three-dimensional rgb pixel matrix dimensionality reduction to one-dimensional gray-scale pixels matrix, to reduce bilayer
The parameter training time of residual error network；The convergence rate and training accuracy of depth residual error network can be improved at normalization.
When further, using the training upper layer network parameter of training dataset 1, assessed using softmax loss function defeated
Error between entering and exporting, and loss function is optimized by stochastic gradient descent algorithm, this method can allow upper layer network maximum
Change ground study to the ability for distinguishing different classes of three-dimensional CAD model.
Further, lower layer's sorting network parameter is initialized using upper level filter network parameter, further using training number
When according to the training lower layer's network parameter of collection 2, the error between input and output is assessed using center-softmax loss function, and
Optimize loss function by stochastic gradient descent algorithm, this method can allow lower layer's sorting network maximumlly to learn to each three-dimensional
CAD model profound level recognizes feature, for calculating the similarity between three-dimensional CAD model.
Further, when three-dimensional CAD model is retrieved, the trained double-deck depth residual error network is called, user can pass through
Entity view, wire frame view, engineering drawing, the engineering sketch isometric drawing for inputting three-dimensional CAD model are realized to similar three-dimensional CAD mould
The precise search of type, so that the three-dimensional CAD model that retrieval obtains directly be used or use enterprise's new product indirectly by modifying
In research and development, and then shorten the R&D cycle, reduce R & D Cost, improves the performance of enterprises.
In conclusion the present invention is understood using enterprise staff and known three-dimensional CAD model view (such as entity view, line
Frame view, engineering drawing, engineering sketch etc.) input as three-dimensional CAD model retrieval, pass through and calls the trained double-deck depth
Residual error network realizes the intelligent retrieval to three-dimensional CAD model, and this method is flexible with retrieval input, retrieval time is short, retrieval
The features such as accuracy rate is high can provide effective support for the intelligent retrieval of enterprise's magnanimity three-dimensional CAD model and efficient reuse.
Below by drawings and examples, technical scheme of the present invention will be described in further detail.
Detailed description of the invention
Fig. 1 is that the three-dimensional CAD model based on the double-deck depth residual error network retrieves block schematic illustration；
Fig. 2 is the double-deck depth residual error schematic network structure；
Fig. 3 is the partial 3-D CAD model schematic diagram that data set includes；
Fig. 4 is that three-dimensional CAD model entity view and wire frame view obtain schematic diagram.
Fig. 5 is FilterNet training process schematic diagram；
Fig. 6 is RankNet training process schematic diagram；
Fig. 7 is the three-dimensional CAD model intelligent retrieval case schematic diagram based on a variety of views；
Fig. 8 is that case schematic diagram is assessed in three-dimensional CAD model intelligent retrieval.
Specific embodiment
The present invention provides a kind of three-dimensional CAD model intelligent search methods based on the double-deck depth residual error network, will be three-dimensional
CAD model view is retrieved as three-dimensional CAD model and is inputted, so that depth residual error network is introduced, using depth residual error network conduct
Basic framework constructs the double-deck depth residual error network including screen and sorting network；It is three-dimensional by obtaining enterprise's history
CAD model entity view/wire frame view is gone forward side by side line number Data preprocess production training dataset 1 and training dataset 2, then will be counted
It is used for the parameter training of screen according to collection 1, data set 2 is used for the parameter training of sorting network, trained network is protected
Save as .h5 file；Call the trained double-deck depth residual error network, by the entity view of three-dimensional CAD model, wire frame view,
Engineering drawing, engineering sketch realize the precise search to similar three-dimensional CAD model.Three-dimensional CAD model intelligence proposed by the present invention
Search method has the characteristics that retrieval input is flexible, retrieval time is short, retrieval precision is high, is enterprise's magnanimity three-dimensional CAD model
Intelligent retrieval and efficiently reuse provide effectively support.
A kind of three-dimensional CAD model intelligent search method based on the double-deck depth residual error network of the present invention, comprising the following steps:
S1, using depth residual error network as basic framework, construct the double-deck deep learning network；
Referring to Fig. 1, the double-deck deep learning network includes that upper level filter network (hereinafter referred to as FilterNet) and lower layer arrange
Sequence network (hereinafter referred to as RnakNet).Wherein, it is tentatively excluded when FilterNet is retrieved for three-dimensional CAD model unrelated with input
Three-dimensional CAD model, RankNet is for being further ranked up three-dimensional CAD model relevant to input by similarity and defeated
Out to user.FilterNet and RankNet is all made of depth residual error network as basic framework.
Depth residual error network overall structure include view preprocessing module, residual error module (Residual Block,
ResBlock it) is constituted with post-processing module, wherein residual error modular expression formula are as follows:
O (x)=max (0, F (x, { Wi})+Wsx)
Wherein, x is the output of a upper layer network, and O (x) is the output of residual error module, F (x, { Wi) it is that residual error learns (WiIt is residual
The weight parameter of difference module), WsX is identical mapping (WsFor square matrix, for matching x and F (x, { Wi) dimension).
S2, the multiple entity views and wire frame view that three-dimensional CAD model is obtained by virtual camera, should for sufficiently characterizing
Model；
Entity view is defined as three-dimensional CAD model with the view obtained after the colouring of sideline, and wire frame view definition is three-dimensional CAD
The view obtained after model Hidden line elimination, the expression formula based on the three-dimensional CAD model of entity view and wire frame view characterization are as follows:
Wherein U is training dataset model sum) indicate that some is given
Three-dimensional CAD model, affiliated model classification areW is the model of training data lump
Classification),Indicate the entity view set of the model,Indicate the wire frame view set of the model, entity view and line
The number of frame view is n, and pixel size is k × l.
S3, entity view and wire frame view are pre-processed by scaling, greyscale transformation and normalization；
The double-deck depth residual error network parameter training needs to input the picture of unified size, thus regards first to the entity of acquisition
Figure and wire frame view do scaling, expression formula are as follows:
Wherein, [x0 y0It 1] is the point on entity view or wire frame view, [x1 y1After 1] being scaled for the scale of view
To new view on corresponding point, and the pixel size of new view is p × q.
The color of view retrieves without substantial role three-dimensional CAD model, thus uses weighting method to the view after scaling
Carry out greyscale transformation, expression formula are as follows:
V is the picture element matrix after view greyscale transformation, VR、VG、VBFor (R) red before view transformation, green (G), three, indigo plant (B)
The corresponding picture element matrix of Color Channel, the dimension of above-mentioned picture element matrix are p × q.
Data set normalized can improve the convergence rate of the double-deck depth residual error network and training accuracy, data set are returned
One expression formula changed are as follows:
Wherein, m=2 × n × U is total number of views.
S4, pretreated view for training dataset 1 and is trained into number by affiliated model classification and affiliated model partition
According to collection 2；
S401, by pretreated view by affiliated model category division be training dataset 1, expression formula are as follows:
Wherein,Expression model classification is cwThe view set of the three-dimensional CAD model of (such as hydraulic pump), and institute in the set
The label for having view is cw。
S402, by pretreated view by affiliated model partition be training dataset 2, expression formula are as follows:
Wherein,Indicate modelThe set of all views, and the label of the view in the set is
S5, the parameter that the data set 1 that step S4 is obtained is imported into FilterNet and training network；
Using the mistake between softmax loss function evaluation input and output in FilterNet network parameter training process
Difference, expression formula are as follows:
Wherein, n is the quantity that most small quantities of (mini-batch) inputs view,For i-th for most inputting view in small batches
The label value of view, y(i)Pass through the output valve after the network forward-propagating of depth residual error for the view.
During network parameter W training, improved depth residual error network optimizes loss function using stochastic gradient descent algorithm,
Stochastic gradient descent method expression formula are as follows:
Wi+1:=Wi+vi+1
Wherein, i is iteration index, and τ is momentum, and v is momentum variable, and μ is weight loss, and ε is learning rate.
S6, when FilterNet reaches the training the number of iterations K of setting, training process terminates, after training
FilterNet saves as FilterNet.h5 file；
S7, the parameter that the data set 2 that step S4 is obtained is imported into RankNet and training network；
Using between center-softmax loss function evaluation input and output in RankNet network parameter training process
Error, expression formula are as follows:
Wherein, xiIndicate the depth of i-th most small quantities of (mini-batch) input view of the full articulamentum output of RankNet
Feature, and the view belongs to three-dimensional CAD modelView,Indicate three-dimensional CAD modelThe depth characteristic of all views
Center, λ indicate center loss softmax loss between weight.
During RankNet network parameter W training, loss function is optimized using the stochastic gradient descent algorithm in S5.
S8, when the average trained accuracy rate of continuous 200 iteration of RankNet is promoted and is less than threshold gamma, training process knot
RankNet after training is saved as RankNet.h5 file by beam；
The average trained accuracy rate of continuous 200 iteration of RankNet promotes expression formula are as follows:
Wherein, AiThe training accuracy rate of RankNet after expression i-th iteration.
Training terminates Rule of judgment are as follows:
S9, call trained FilterNet and RankNet, user can by the entity view of three-dimensional CAD model,
Wire frame view, engineering drawing, engineering sketch isometric drawing realize the precise search to similar three-dimensional CAD model；
Input view is pre-processed according to the view preprocess method in step S3 first；Then FilterNet and
The similarity expression formula of three-dimensional CAD model in the N view computation databases that RankNet passes through input are as follows:
Wherein, FpAnd RsRespectively Filternet and RankNet is directed to the output matrix of i-th view, WpFor unit square
Battle array, for matching FpAnd RsDimension,Representing matrix dot product,It indicates and input the highest M of view similarity
Three-dimensional CAD model is pushed to user as search result.
In order to make the object, technical scheme and advantages of the embodiment of the invention clearer, below in conjunction with the embodiment of the present invention
In attached drawing, technical scheme in the embodiment of the invention is clearly and completely described, it is clear that described embodiment is
A part of the embodiment of the present invention, instead of all the embodiments.The present invention being described and shown in usually here in attached drawing is real
The component for applying example can be arranged and be designed by a variety of different configurations.Therefore, below to the present invention provided in the accompanying drawings
The detailed description of embodiment be not intended to limit the range of claimed invention, but be merely representative of of the invention selected
Embodiment.Based on the embodiments of the present invention, those of ordinary skill in the art are obtained without creative efforts
The every other embodiment obtained, shall fall within the protection scope of the present invention.
Referring to Fig. 2, FilterNet and RankNet are all made of the depth residual error network being formed by stacking by 5 residual error modules
(ResNet18) it is used as basic framework, wherein FilterNet the last layer is softmax layers, for calculating input view and number
It is most related according to any class model in library, and (i.e. the view of softmax layers of output belongs to the general of certain one kind by model Category Relevance
Rate) primary filtration fall those and input the lower three-dimensional CAD model of drawings relativity；RankNet the last layer is center-
It softmax layers, sorts for further calculating the similarity of input view and related three-dimensional CAD model, and by similarity height
10 most like three-dimensional CAD models are pushed to user.
Referring to Fig. 3, training dataset includes that 56 model classifications amount to 560 three-dimensional CAD models, each classification includes
10 three-dimensional CAD models.Referring to Fig. 4, according to common 26 view locations in SolidWorks three-dimensional engineering drawing tool
It arranges virtual camera, Blender software is used to obtain 26 pixel sizes of each three-dimensional CAD model automatically as 256 × 256
Entity view uses Blender software to obtain 26 pixel sizes of each three-dimensional CAD model automatically as 256 × 256 wire frame
View.Entity view and wire frame view are pre-processed to obtain pixel size by scaling, greyscale transformation and normalization
For 224 × 224 gray scale view.By gray scale view according to the affiliated model category division of model be training dataset 1, data set 1
In each classification include 2 × 26 × 10=520 gray scale views, and the label of each gray scale view be the category item name or
Number.It is training dataset 2 that gray scale view, which is pressed affiliated model partition, and each model includes 2 × 26=52 ashes in data set 2
View is spent, and the label of each gray scale view is the title or number of the model.
Training dataset 1 is imported into FilterNet and carries out network parameter training；Training the number of iterations is set as K=
200000 times；The hyper parameter of network is arranged are as follows: τ=0.9, μ=0.00001, and as the number of iterations i≤100000, ε=0.01,
As i > 100000, ε=0.0001, most small quantities of (mini-batch) inputs amount of views n=32；When training the number of iterations reaches
When to setting value K, trained FilterNet is saved as ' FilterNet.h5 ' file by training iteration ends；Please refer to figure
5, FilterNet training precision Et=0.9980, verify precision Ev=0.9594, loss function L < 0.16, achieves higher
Training precision and measuring accuracy.
The network parameter (the all-network parameter except softmax layers) of FilterNet is loaded to RankNet, and will be trained
Data set 2 imports RankNet and carries out network parameter training；When network parameter training, between center loss and softmax loss
Weight setting are as follows: λ=0.1, RankNet training termination condition setting are as follows: γ=0.0001；Network hyper parameter setting and
FilterNet is consistent；When the average trained accuracy rate of continuous 200 iteration of RankNet, which is promoted, is less than threshold gamma, training process
Terminate, trained RankNet is saved as into ' RankNet.h5 ' file；Referring to Fig. 6, the training precision E of RankNett=
0.9971, verify precision Ev=0.9416, loss function L < 0.28, achieves higher training precision and measuring accuracy.
Referring to Fig. 7, when needing to retrieve three-dimensional CAD model, call trained FilterNet and
RankNet, user can be regarded by entity view, wire frame view, engineering drawing, the engineering sketch etc. for inputting three-dimensional CAD model
Figure realizes the precise search to similar three-dimensional CAD model, and retrieval time stablizes at 0.85 second or so.For assessment proposition method
Validity is assessed proposed method using average sequence (abbreviation MRR) reciprocal, expression formula are as follows:
Wherein, M is the three-dimensional CAD model sum that RankNet is pushed to user；If i-th of search result and input regard
Scheme similar, then Ranki=1, it is otherwise 0.
Referring to Fig. 8, when retrieving view is 3, MRR can reach 0.923 using the MRR > 0.898 of entity view；
Using the MRR > 0.852 of wire frame view, when retrieving view is 3, MRR can reach 0.909.Above-mentioned experiment shows the present invention
Three-dimensional CAD model search method based on the double-deck depth convolutional neural networks is flexible with input, retrieval time is short, retrieval is accurate
The advantages that rate is high.
The above content is merely illustrative of the invention's technical idea, and this does not limit the scope of protection of the present invention, all to press
According to technical idea proposed by the present invention, any changes made on the basis of the technical scheme each falls within claims of the present invention
Protection scope within.
Claims (10)
1. a kind of three-dimensional CAD model intelligent search method based on the double-deck depth residual error network, which is characterized in that residual using depth
For poor network as basic framework, building includes the double-deck depth residual error net of screen FilterNet and sorting network RankNet
Network tentatively excludes the three-dimensional CAD model unrelated with input, sequence when screen FilterNet is retrieved for three-dimensional CAD model
Network RankNet is for being ranked up three-dimensional CAD model relevant to input by similarity and being exported to user；Pass through acquisition
Enterprise's history three-dimensional CAD model entity view/wire frame view is gone forward side by side line number Data preprocess production training dataset 1 and training data
Data set 1, is then used for the parameter training of screen FilterNet, data set 2 is used for sorting network RankNet by collection 2
Parameter training, trained network is saved as into .h5 file；The trained double-deck depth residual error network is called, three-dimensional is passed through
Entity view, wire frame view, engineering drawing, the engineering sketch of CAD model realize the precise search to similar three-dimensional CAD model.
2. the three-dimensional CAD model intelligent search method according to claim 1 based on the double-deck depth residual error network, feature
It is, depth residual error network overall structure includes that view preprocessing module, residual error module and post-processing module are constituted, wherein residual
Difference module O (x) are as follows:
O (x)=max (0, F (x, { Wi})+Wsx)
Wherein, x is the output of a upper layer network, F (x, { Wi) it is that residual error learns, WiFor the weight parameter of residual error module, WsX is
Identical mapping, WsFor square matrix, for matching x and F (x, { Wi) dimension.
3. the three-dimensional CAD model intelligent search method according to claim 1 based on the double-deck depth residual error network, feature
It is, entity view is defined as three-dimensional CAD model with the view obtained after the colouring of sideline, and wire frame view definition is three-dimensional CAD mould
The view obtained after type Hidden line elimination is characterized based on the three-dimensional CAD model of entity view and wire frame view are as follows:
WhereinU be training dataset model sum) indicate some given three
Victoria C AD model, affiliated model classification are cw,W is the model classification of training data lump,Indicate the entity view set of the model,Indicate the wire frame view set of the model, entity view and wire frame view
Number be n, pixel size is k × l.
4. the three-dimensional CAD model intelligent search method according to claim 1 based on the double-deck depth residual error network, feature
It is, pretreatment includes scaling, greyscale transformation and normalized, and entity view and wire frame view to acquisition do ratio
It scales as follows:
Wherein, [x0 y0It 1] is the point on entity view or wire frame view, [x1 y11] it is obtained after being scaled for the scale of view
Corresponding point on new view, and the pixel size of new view is p × q；
It is as follows using weighting method progress greyscale transformation to the view after scaling:
V is the picture element matrix after view greyscale transformation, VR、VG、VBIt is corresponding for three Color Channels of red, green, blue before view transformation
Picture element matrix, the dimension of above-mentioned picture element matrix are p × q；
Data set normalization is as follows:
Wherein, m=2 × n × U is total number of views.
5. the three-dimensional CAD model intelligent search method according to claim 1 based on the double-deck depth residual error network, feature
It is, it is training dataset 1, expression formula that pretreated view, which is pressed affiliated model category division, are as follows:
Wherein,Expression model classification is cwThe view set of the three-dimensional CAD model of (such as hydraulic pump), and all views in the set
The label of figure is cw。
6. the three-dimensional CAD model intelligent search method according to claim 1 or 5 based on the double-deck depth residual error network,
It is characterized in that, data set 1 is imported into the parameter of screen FilterNet and training network, as screen FilterNet
When reaching the training the number of iterations K of setting, training process terminates, and the screen FilterNet after training is saved as
FilterNet.h5 file is evaluated using softmax loss function in screen FilterNet network parameter training process defeated
Error between entering and exporting, expression formula are as follows:
Wherein, n is the quantity for most inputting view in small batches,For the label value of i-th view of most small quantities of input view, y(i)For
The view passes through the output valve after the network forward-propagating of depth residual error；
During network parameter W training, improved depth residual error network optimizes loss function using stochastic gradient descent algorithm, at random
Gradient descent method expression formula are as follows:
Wi+1:=Wi+vi+1
Wherein, i is iteration index, and τ is momentum, and v is momentum variable, and μ is weight loss, and ε is learning rate.
7. the three-dimensional CAD model intelligent search method according to claim 1 based on the double-deck depth residual error network, feature
It is, it is training dataset 2, expression formula that pretreated view, which is pressed affiliated model partition, are as follows:
Wherein,Indicate modelThe set of all views, and the label of the view in the set is
8. the three-dimensional CAD model intelligent search method according to claim 1 or claim 7 based on the double-deck depth residual error network,
It is characterized in that, sorting network RankNet network parameter is initialized using screen FilterNet network parameter, by data set 2
It imported into the parameter of sorting network RankNet and training network；It is used in sorting network RankNet network parameter training process
Error between center-softmax loss function evaluation input and output, expression formula are as follows:
Wherein, xiIndicate the depth characteristic of i-th most small quantities of (mini-batch) input view of the full articulamentum output of RankNet,
And the view belongs to three-dimensional CAD modelView,Indicate three-dimensional CAD modelThe center of the depth characteristic of all views,
λ indicates the weight between center loss and softmax loss；During sorting network RankNet network parameter W training, use
Stochastic gradient descent algorithm optimizes loss function.
9. the three-dimensional CAD model intelligent search method according to claim 8 based on the double-deck depth residual error network, feature
It is, when the average trained accuracy rate of continuous 200 iteration of sorting network RankNet, which is promoted, is less than threshold gamma, training process
Terminate, the sorting network RankNet after training is saved as into RankNet.h5 file；Sorting network RankNet continuous 200 times repeatedly
The average trained accuracy rate in generation promotes expression formula are as follows:
Wherein, AiThe training accuracy rate of RankNet after expression i-th iteration；Training terminates Rule of judgment are as follows:
10. the three-dimensional CAD model intelligent search method according to claim 1 based on the double-deck depth residual error network, special
Sign is, first pre-processes input view, the N views that then the trained double-deck depth residual error network passes through input
Calculate the similarity expression formula of three-dimensional CAD model in database are as follows:
Wherein, FpAnd RsRespectively screen Filternet and sorting network RankNet is directed to the output square of i-th view
Battle array, WpFor unit matrix, for matching FpAnd RsDimension,Representing matrix dot product,It indicates and input view phase
User is pushed to like highest M three-dimensional CAD model is spent as search result.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910401751.7A CN110245249B (en) | 2019-05-15 | 2019-05-15 | Three-dimensional CAD model intelligent retrieval method based on double-layer depth residual error network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910401751.7A CN110245249B (en) | 2019-05-15 | 2019-05-15 | Three-dimensional CAD model intelligent retrieval method based on double-layer depth residual error network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110245249A true CN110245249A (en) | 2019-09-17 |
| CN110245249B CN110245249B (en) | 2021-08-13 |
Family
ID=67884019
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910401751.7A Active CN110245249B (en) | 2019-05-15 | 2019-05-15 | Three-dimensional CAD model intelligent retrieval method based on double-layer depth residual error network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110245249B (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111914112A (en) * | 2020-07-07 | 2020-11-10 | 西安交通大学 | Part CAD model reusing method based on point cloud classification network |
| CN113609555A (en) * | 2021-07-16 | 2021-11-05 | 黄河勘测规划设计研究院有限公司 | Hydraulic metal structure design method based on big data technology |
| CN114119930A (en) * | 2022-01-27 | 2022-03-01 | 广州中望龙腾软件股份有限公司 | Three-dimensional model correction method and device based on deep learning and storage medium |
| CN115238591A (en) * | 2022-08-12 | 2022-10-25 | 杭州国辰智企科技有限公司 | Dynamic parameter checking and driving CAD automatic modeling engine system |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107122396A (en) * | 2017-03-13 | 2017-09-01 | 西北大学 | Three-dimensional model searching algorithm based on depth convolutional neural networks |
| CN108427729A (en) * | 2018-02-23 | 2018-08-21 | 浙江工业大学 | A kind of extensive picture retrieval method based on depth residual error network and Hash coding |
| US20190094875A1 (en) * | 2017-09-28 | 2019-03-28 | Nec Laboratories America, Inc. | Generating occlusion-aware bird eye view representations of complex road scenes |
-
2019
- 2019-05-15 CN CN201910401751.7A patent/CN110245249B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107122396A (en) * | 2017-03-13 | 2017-09-01 | 西北大学 | Three-dimensional model searching algorithm based on depth convolutional neural networks |
| US20190094875A1 (en) * | 2017-09-28 | 2019-03-28 | Nec Laboratories America, Inc. | Generating occlusion-aware bird eye view representations of complex road scenes |
| CN108427729A (en) * | 2018-02-23 | 2018-08-21 | 浙江工业大学 | A kind of extensive picture retrieval method based on depth residual error network and Hash coding |
Non-Patent Citations (2)
| Title |
|---|
| CHAO ZHANG等: "A view-based 3D CAD model reuse framework enabling product lifecycle", 《ELSEVIER》 * |
| 李荫民: "基于残差网络的三维模型检索算法", 《计算机科学》 * |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111914112A (en) * | 2020-07-07 | 2020-11-10 | 西安交通大学 | Part CAD model reusing method based on point cloud classification network |
| CN111914112B (en) * | 2020-07-07 | 2023-05-23 | 西安交通大学 | Part CAD model reuse method based on point cloud classification network |
| CN113609555A (en) * | 2021-07-16 | 2021-11-05 | 黄河勘测规划设计研究院有限公司 | Hydraulic metal structure design method based on big data technology |
| CN113609555B (en) * | 2021-07-16 | 2023-10-20 | 黄河勘测规划设计研究院有限公司 | Hydraulic metal structure design method based on big data technology |
| CN114119930A (en) * | 2022-01-27 | 2022-03-01 | 广州中望龙腾软件股份有限公司 | Three-dimensional model correction method and device based on deep learning and storage medium |
| CN114119930B (en) * | 2022-01-27 | 2022-10-28 | 广州中望龙腾软件股份有限公司 | Three-dimensional model correction method and device based on deep learning and storage medium |
| CN115238591A (en) * | 2022-08-12 | 2022-10-25 | 杭州国辰智企科技有限公司 | Dynamic parameter checking and driving CAD automatic modeling engine system |
| CN115238591B (en) * | 2022-08-12 | 2022-12-27 | 杭州国辰智企科技有限公司 | Dynamic parameter checking and driving CAD automatic modeling engine system |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110245249B (en) | 2021-08-13 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110245249A (en) | A kind of three-dimensional CAD model intelligent search method based on the double-deck depth residual error network | |
| CN106156003B (en) | A kind of question sentence understanding method in question answering system | |
| CN109376692A (en) | Migration convolution neural network method towards facial expression recognition | |
| CN108921822A (en) | Image object method of counting based on convolutional neural networks | |
| CN108985317A (en) | A kind of image classification method based on separable convolution sum attention mechanism | |
| CN108009594B (en) | A kind of image-recognizing method based on change grouping convolution | |
| CN106778682A (en) | A kind of training method and its equipment of convolutional neural networks model | |
| CN106991386A (en) | A kind of gesture identification method based on depth residual error network | |
| CN103761295B (en) | Automatic picture classification based customized feature extraction method for art pictures | |
| CN109063112A (en) | A kind of fast image retrieval method based on multi-task learning deep semantic Hash, model and model building method | |
| CN111652273B (en) | Deep learning-based RGB-D image classification method | |
| CN109102498A (en) | A kind of method of cluster type nucleus segmentation in cervical smear image | |
| CN107563407A (en) | A kind of character representation learning system of the multi-modal big data in network-oriented space | |
| CN109871892A (en) | A kind of robot vision cognitive system based on small sample metric learning | |
| US11841839B1 (en) | Preprocessing and imputing method for structural data | |
| CN109325513A (en) | A kind of image classification network training method based on magnanimity list class single image | |
| CN107992611A (en) | The high dimensional data search method and system of hash method are distributed based on Cauchy | |
| CN110674326A (en) | Neural network structure retrieval method based on polynomial distribution learning | |
| CN107506362A (en) | Image classification based on customer group optimization imitates brain storage method | |
| CN113034506A (en) | Remote sensing image semantic segmentation method and device, computer equipment and storage medium | |
| CN104036242A (en) | Object recognition method based on convolutional restricted Boltzmann machine combining Centering Trick | |
| Ma et al. | Research on fish image classification based on transfer learning and convolutional neural network model | |
| CN109840550A (en) | A kind of mobile subscriber's application preferences recognition methods based on deep neural network | |
| CN110110120A (en) | A kind of image search method and device based on deep learning | |
| CN113128560A (en) | Attention module enhancement-based CNN regular script style classification method |
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 |