CN109801375A - Porous material three-dimensional reconstruction method based on depth convolution confrontation neural network - Google Patents
Porous material three-dimensional reconstruction method based on depth convolution confrontation neural network Download PDFInfo
- Publication number
- CN109801375A CN109801375A CN201910068310.XA CN201910068310A CN109801375A CN 109801375 A CN109801375 A CN 109801375A CN 201910068310 A CN201910068310 A CN 201910068310A CN 109801375 A CN109801375 A CN 109801375A
- Authority
- CN
- China
- Prior art keywords
- image
- porous material
- network
- neural network
- generation
- 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
Landscapes
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a kind of porous material three-dimensional reconstruction methods based on depth convolution confrontation neural network, comprising the following steps: SEM scanning acquisition surface image S1, is carried out to the porous material of selection;S2, the image of acquisition is pre-processed, including image cutting, binaryzation, denoising;S3, the image pre-processed importing DCGAN network is subjected to confrontation generation training;S4, three-dimensionalreconstruction is carried out by generation network image generated.The present invention, which uses, is based on SEM sem image, carries out three-dimensionalreconstruction in conjunction with the method for depth convolutional neural networks confrontation neural network (DCGAN).Compared to conventional method, the model that the present invention generates is more nearly entity, may be advantageously employed in the research in every such as microfluidic direction, at the same also have the characteristics that cheap, method is easy, using quick.
Description
Technical field
The invention belongs to materials science fields, in particular to a kind of to fight neural network (DCGAN) based on depth convolution
Porous material three-dimensional reconstruction method.
Background technique
Porous gas static pressure bearing has high-precision, low vibration, pollution-free compared with traditional mechanical contact bearing
And the service life it is long the features such as, and while the rapid development of the industry such as aviation in recent years, semiconductor, bioengineering, composite material
Making the bearing of Ultra-precision Turning to loading material, more stringent requirements are proposed.And Porous Aerostatic Bearing be able to meet wanting for epoch
It asks, or even is known as 21 century novel bearing most with prospects.
Important component of the porous gas static pressure bearing as ultra-precision machine tool, Static and dynamic performance are always to study
The emphasis of concern.There are thousands of a slight restriction holes on porous material surface, makes it while guaranteeing high rigidity, high-mechanic
Stability also with higher, but the big quantity in hole and its irregular distribution also keep scholar in check to its theoretical model significantly
The research of construction, structurally and functionally rule.In the past in the research process of porous material performance, we are often by Porous
Material ideal is melted into a homogeneous material, ignores Permeability Distribution caused by material inhomogeneities itself unevenly and is embodied
Air film clearance pressure out is unevenly distributed.Therefore, in order to further probe into porous gas static pressure bearing static and dynamic performance
And the flow field movement in Ultra-precision Turning, we must work as the genuine property of porous material in view of simulation calculation
In, to replace previous idealized model, to obtain the model and data closer to actual conditions.
Have much for the research of porous material three-dimensionalreconstruction at present, can substantially be divided into three classes, one kind is based on quasi-
Porous material is fitted by the distribution of molding type with equally distributed ball, column.This reconstructing method implements simply, but
Be it is big with truth difference, be not available the model when studying microfluidic direction etc..And the second class method is then based on general
The random distribution of rate, using the bead or cylinder of random distribution come model of fit.Compared to first method, after the method reconstruct
The aspect of model is more abundant, it may have certain anisotropy, but still have larger difference with true hole distribution situation.Third
Class is then the method based on CT scan, is directly scanned using high-precision micro nano CT to material and establishes threedimensional model.Cause
For porous material self-characteristic, hole is at 1-10 μm, therefore this method is too high for equipment requirement, therefore price is too high,
Rebuild primary cost and requirement of experiment grave fault.
Summary of the invention
It is an object of the invention to overcome the deficiencies of the prior art and provide one kind to be based on SEM sem image, rolls up in conjunction with depth
Product neural network confrontation neural network (DCGAN) carries out three-dimensionalreconstruction, may be advantageously employed in items such as microfluidic direction
Research, have the characteristics that cheap, method is easy, using it is quick based on depth convolution fight neural network porous material
Expect three-dimensional reconstruction method.
The purpose of the present invention is achieved through the following technical solutions: based on the porous of depth convolution confrontation neural network
Material three-dimensional reconstruction method, comprising the following steps:
S1, SEM scanning acquisition surface image is carried out to the porous material of selection;
S2, the image of acquisition is pre-processed, including image cutting, binaryzation, denoising;
S3, the image pre-processed importing DCGAN network is subjected to confrontation generation training;
S4, three-dimensionalreconstruction is carried out by generation network image generated.
Further, the step S1 includes following sub-step:
S11, it is scanned and is taken pictures with porous material surface of the SEM to selection, obtain material surface image, and protected
Save as the tiff format with light and shade information;
S12, using accurate grinding mode, scanning area is ground, after every grinding certain depth using step S11
Method, which is scanned, takes pictures, and obtains multilayer material surface image, establishes tranining database.
Further, the step S2 includes following sub-step:
S21, the step S1 image obtained is split, coordinate system is established by origin of picture centre, in any quadrant
The middle topography for choosing 64 pixels is cut;
S22, binary conversion treatment is carried out to the topography cut, makes material entities part and aperture sections point in image
From the threshold value of binary conversion treatment is the brightness maxima of image and the average value of minimum value;
S23, Sigma filtering and noise reduction is carried out to the image after binary conversion treatment, saves hole side while removing noise
The irregular shape on boundary.
Further, the step S3 includes following sub-step:
S31, the training environment according to required for DCGAN are trained environment configurations, on anaconda and server
Python3.6, pip, hdf5 environment are configured, needs is trained using independent display card and configuration CUDA installation kit is installed;
S32, the main function of DCGAN is called in anaconda and carries out the setting of parameter, imagesize is before
Image preprocessing is set as 64, batchsize and 128, niter is selected to be set as 1000, and learning rate is set as 0.00001, generates
Feature number is 64 in network, identifies that feature number is 16 in network, initial noise vector magnitude is 512, is set
Main function is run after parameter is trained generation;
After S33, generation training terminate, generator can generate hdf5 transition file under specified generation catalogue, after
Processing routine is tiff file to generating image to carry out format conversion, obtains interpolation fitting image between layers.
Further, the step S4 includes following sub-step:
S41, the generation image between first layer and the second layer is ranked up from top to bottom according to depth information, it will be intermediate
The layer-by-layer draft of tomographic image forms the solid for having thickness;The draft distance is surface sweeping spacing and the ratio for generating amount of images
Value;
S42, the step S41 solid formed is stacked gradually, carries out the Boolean calculation of layer and interlayer again, makes body and body
Between formed an entirety;
S43, the operation that other all clearance layers are repeated with step S41~S42, it is final to realize porous material Three-dimensional Gravity
Structure.
The beneficial effects of the present invention are: the present invention uses the reconstructing method based on feature, it is based on SEM sem image, in conjunction with
Depth convolutional neural networks fight neural network (DCGAN), carry out three-dimensionalreconstruction.Compared to conventional method, mould that the present invention generates
Type is more nearly entity, may be advantageously employed in the research in every such as microfluidic direction, while also having cheap, method
It is easy, using it is quick the features such as.
Detailed description of the invention
Fig. 1 is the flow chart of porous material three-dimensional reconstruction method of the invention;
Fig. 2 is the scan image that the present embodiment obtains;
Fig. 3 is the middle layer that the training of the present embodiment generates;
Fig. 4 is the reconstruction model that this present embodiment obtains.
Specific embodiment
The present invention uses the reconstructing method based on feature, is based on SEM sem image, fights in conjunction with depth convolutional neural networks
Neural network (DCGAN) carries out three-dimensionalreconstruction.Compared to conventional method, the model that artificial intelligence generates is more nearly entity, can be very
The good research for every such as microfluidic direction, at the same also have the characteristics that cheap, method is easy, using quick.
DCGAN (Deep Convolutional Generative Adversarial Networks), i.e. depth convolution
Generate confrontation network.It is the series of optimum and improvement done on the basis of GAN in fact.The principle of GAN is carried out first
Illustrating --- GAN is proposed in 2014 Nian Qi papers by Ian Goodfellow, and model includes two neural networks, G
(Generator is generated) network and D (Discriminator differentiates) network.When input is picture, G network is substantially
The network of one generation picture, it receives a random noise signal z, generates picture, i.e. G (z) by this noise;D is then
A differentiation network, its effect be differentiate a picture whether " true ", judge input from authentic specimen image still
From G network.GAN final output is substantially what G network generated, has successfully cheated " the pseudo- image " of D network.With GAN into
During row training, the purpose for generating network is to try to generate true picture deception differentiation network, and differentiates the mesh of network
Be then to try to G generate picture and true picture distinguish.It interacts, mutually in this way, G and D constitute one
Confrontation, until a convergent closed loop procedure.
DCGAN uses the principle of GAN, has only used two CNN (Convolutional Neutral Network, convolution
Neural network) substitute the G network and D network in GAN, compared to being greatly improved the stability and generation newly joined for GAN
Outcome quality.DCGAN is modified in that GAN:
(1) cancel all pond layers, up-sampled in generator G network using warp lamination, in identifier D network
The convolution with stride (Strided convolutions) is added and replaces pond layer (Pooling);
(2) using batch normalization (Batch normalization) in generator G and identifier D;
(3) remove full articulamentum, network is made to become full convolutional network;
(4) use ReLU (Rectified linear unit) as activation primitive, the last layer in generator G network
Tangent function (Tanh) is cut using double;
(5) use Leaky ReLU (Leaky rectified linear unit) as activation letter in identifier D network
Number, the last layer use Softmax.
Technical solution of the present invention is further illustrated with reference to the accompanying drawing.
As shown in Figure 1, a kind of porous material three-dimensionalreconstruction side based on depth convolution confrontation neural network of the invention
Method the following steps are included:
S1, SEM scanning acquisition surface image is carried out to the porous material of selection;Specifically include following sub-step:
S11, it is scanned and is taken pictures with porous material surface of the SEM to selection, obtain material surface image, and protected
Save as the tiff format with light and shade information;
S12, using accurate grinding mode, scanning area is ground, (such as 1mm, grinding after every grinding certain depth
Depth can be by operator's self-defining) it is scanned and is taken pictures using the method for step S11, multilayer material surface image is obtained,
As shown in Fig. 2, then establishing tranining database.
S2, the image of acquisition is pre-processed, including image cutting, binaryzation, denoising;Specifically include following sub-step
It is rapid:
S21, the step S1 image obtained is split, coordinate system is established by origin of picture centre, in any quadrant
The middle topography for choosing 64 pixels is cut;
S22, binary conversion treatment is carried out to the topography cut, makes material entities part and aperture sections point in image
From the threshold value of binary conversion treatment is the brightness maxima of image and the average value of minimum value;
S23, Sigma filtering and noise reduction is carried out to the image after binary conversion treatment, saves hole side while removing noise
The irregular shape on boundary.
S3, the image pre-processed importing DCGAN network is subjected to confrontation generation training;Including following sub-step:
S31, the training environment according to required for DCGAN are trained environment configurations, on anaconda and server
Python3.6, pip, hdf5 environment are configured, needs is trained using independent display card and configuration CUDA installation kit is installed;
S32, the main function of DCGAN is called in anaconda and carries out the setting of parameter, imagesize is before
Image preprocessing is set as 64, batchsize and 128, niter is selected to be set as 1000, and learning rate is set as 0.00001, generates
Feature number is 64 in network, identifies that feature number is 16 in network, initial noise vector magnitude is 512, is set
Main function is run after parameter is trained generation;
After S33, generation training terminate, generator can generate hdf5 transition file under specified generation catalogue, after
Processing routine is tiff file to generating image to carry out format conversion, interpolation fitting image between layers is obtained, such as Fig. 3 institute
Show.
S4, three-dimensionalreconstruction is carried out by generation network image generated;Including following sub-step:
S41, the generation image between first layer and the second layer is ranked up from top to bottom according to depth information, it will be intermediate
The layer-by-layer draft of tomographic image forms the solid for having thickness;The draft distance is surface sweeping spacing and the ratio for generating amount of images
Value;
S42, the step S41 solid formed is stacked gradually, carries out the Boolean calculation of layer and interlayer again, makes body and body
Between formed an entirety;
S43, the operation that other all clearance layers are repeated with step S41~S42, it is final to realize porous material Three-dimensional Gravity
Structure.
Fig. 4 is the reconstruction model that the present embodiment obtains, wherein Fig. 4 (a) is entity and slice graph model after draft, Fig. 4
(b) entity to be obtained after X-Y plane draft;The face Fig. 4 (c) Y-Z draft thickness h;The finally obtained entity of Fig. 4 (d).
Those of ordinary skill in the art will understand that the embodiments described herein, which is to help reader, understands this hair
Bright principle, it should be understood that protection scope of the present invention is not limited to such specific embodiments and embodiments.This field
Those of ordinary skill disclosed the technical disclosures can make according to the present invention and various not depart from the other each of essence of the invention
The specific variations and combinations of kind, these variations and combinations are still within the scope of the present invention.
Claims (5)
1. the porous material three-dimensional reconstruction method based on depth convolution confrontation neural network, which is characterized in that including following step
It is rapid:
S1, SEM scanning acquisition surface image is carried out to the porous material of selection;
S2, the image of acquisition is pre-processed, including image cutting, binaryzation, denoising;
S3, the image pre-processed importing DCGAN network is subjected to confrontation generation training;
S4, three-dimensionalreconstruction is carried out by generation network image generated.
2. the porous material three-dimensional reconstruction method according to claim 1 based on depth convolution confrontation neural network,
It is characterized in that, the step S1 includes following sub-step:
S11, it is scanned and is taken pictures with porous material surface of the SEM to selection, obtain material surface image, and be saved as
Tiff format with light and shade information;
S12, using accurate grinding mode, scanning area is ground, it is every grinding certain depth after utilize step S11 method
It is scanned and takes pictures, obtain multilayer material surface image, establish tranining database.
3. the porous material three-dimensional reconstruction method according to claim 1 based on depth convolution confrontation neural network,
It is characterized in that, the step S2 includes following sub-step:
S21, the step S1 image obtained is split, establishes coordinate system by origin of picture centre, is selected in any quadrant
The topography of 64 pixels is taken to be cut;
S22, binary conversion treatment is carried out to the topography cut, separates material entities part in image with aperture sections,
The threshold value of binary conversion treatment is the brightness maxima of image and the average value of minimum value;
S23, Sigma filtering and noise reduction is carried out to the image after binary conversion treatment, saves pore boundary while removing noise
Irregular shape.
4. the porous material three-dimensional reconstruction method according to claim 1 based on depth convolution confrontation neural network,
It is characterized in that, the step S3 includes following sub-step:
S31, the training environment according to required for DCGAN are trained environment configurations, configure on anaconda and server
Python3.6, pip, hdf5 environment are trained needs using independent display card and install configuration CUDA installation kit;
S32, the main function of DCGAN is called in anaconda and carries out the setting of parameter, imagesize image before
Pretreatment is set as 64, batchsize and 128, niter is selected to be set as 1000, and learning rate is set as 0.00001, generates network
Middle feature number is 64, identifies that feature number is 16 in network, initial noise vector magnitude is 512, sets parameter
Operation main function is trained generation later;
After S33, generation training terminate, generator can generate hdf5 transition file under specified generation catalogue, with post-processing
Program is tiff file to generating image to carry out format conversion, obtains interpolation fitting image between layers.
5. the porous material three-dimensional reconstruction method according to claim 1 based on depth convolution confrontation neural network,
It is characterized in that, the step S4 includes following sub-step:
S41, the generation image between first layer and the second layer is ranked up from top to bottom according to depth information, by middle layer figure
As layer-by-layer draft, the solid for having thickness is formed;The draft distance is surface sweeping spacing and the ratio for generating amount of images;
S42, the step S41 solid formed is stacked gradually, carries out the Boolean calculation of layer and interlayer again, makes between body and body
Form an entirety;
S43, the operation that other all clearance layers are repeated with step S41~S42, it is final to realize porous material three-dimensionalreconstruction.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910068310.XA CN109801375B (en) | 2019-01-24 | 2019-01-24 | Three-dimensional reconstruction method of porous material based on deep convolution anti-neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910068310.XA CN109801375B (en) | 2019-01-24 | 2019-01-24 | Three-dimensional reconstruction method of porous material based on deep convolution anti-neural network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109801375A true CN109801375A (en) | 2019-05-24 |
CN109801375B CN109801375B (en) | 2021-02-09 |
Family
ID=66560276
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910068310.XA Active CN109801375B (en) | 2019-01-24 | 2019-01-24 | Three-dimensional reconstruction method of porous material based on deep convolution anti-neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109801375B (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111402397A (en) * | 2020-02-28 | 2020-07-10 | 清华大学 | TOF depth data optimization method and device based on unsupervised data |
CN112348947A (en) * | 2021-01-07 | 2021-02-09 | 南京理工大学智能计算成像研究院有限公司 | Three-dimensional reconstruction method for deep learning based on reference information assistance |
CN113781442A (en) * | 2021-09-13 | 2021-12-10 | 安徽农业大学 | Rice seedling plant three-dimensional structure model reconstruction method |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101706966A (en) * | 2009-11-06 | 2010-05-12 | 上海第二工业大学 | Method for three-dimensional reconstruction of porous medium on basis of two-dimensional images and multi-point statistical method |
CN107368852A (en) * | 2017-07-13 | 2017-11-21 | 西安电子科技大学 | A kind of Classification of Polarimetric SAR Image method based on non-down sampling contourlet DCGAN |
CN107563385A (en) * | 2017-09-02 | 2018-01-09 | 西安电子科技大学 | License plate character recognition method based on depth convolution production confrontation network |
CN107633486A (en) * | 2017-08-14 | 2018-01-26 | 成都大学 | Structure Magnetic Resonance Image Denoising based on three-dimensional full convolutional neural networks |
US20180059041A1 (en) * | 2016-08-08 | 2018-03-01 | B.G. Negev Technologies & Applications Ltd. At Ben-Gurion University | High sensitivity broad-target porous graphene oxide capacitive vapor sensor |
CN108088864A (en) * | 2017-12-15 | 2018-05-29 | 浙江隆劲电池科技有限公司 | A kind of material three-dimensional microstructure reconstructing method and system |
US20180293721A1 (en) * | 2017-04-07 | 2018-10-11 | Kla-Tencor Corporation | Contour based defect detection |
-
2019
- 2019-01-24 CN CN201910068310.XA patent/CN109801375B/en active Active
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101706966A (en) * | 2009-11-06 | 2010-05-12 | 上海第二工业大学 | Method for three-dimensional reconstruction of porous medium on basis of two-dimensional images and multi-point statistical method |
US20180059041A1 (en) * | 2016-08-08 | 2018-03-01 | B.G. Negev Technologies & Applications Ltd. At Ben-Gurion University | High sensitivity broad-target porous graphene oxide capacitive vapor sensor |
US20180293721A1 (en) * | 2017-04-07 | 2018-10-11 | Kla-Tencor Corporation | Contour based defect detection |
CN107368852A (en) * | 2017-07-13 | 2017-11-21 | 西安电子科技大学 | A kind of Classification of Polarimetric SAR Image method based on non-down sampling contourlet DCGAN |
CN107633486A (en) * | 2017-08-14 | 2018-01-26 | 成都大学 | Structure Magnetic Resonance Image Denoising based on three-dimensional full convolutional neural networks |
CN107563385A (en) * | 2017-09-02 | 2018-01-09 | 西安电子科技大学 | License plate character recognition method based on depth convolution production confrontation network |
CN108088864A (en) * | 2017-12-15 | 2018-05-29 | 浙江隆劲电池科技有限公司 | A kind of material three-dimensional microstructure reconstructing method and system |
Non-Patent Citations (2)
Title |
---|
何东杰: "基于单幅图像的多孔材料几何模型三维重构研究", 《中国优秀硕士学位论文全文数据库信息科技辑》 * |
张文东: "基于CNN的有遮挡三维人脸重建算法的研究", 《中国优秀硕士学位论文全文数据库信息科技辑》 * |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111402397A (en) * | 2020-02-28 | 2020-07-10 | 清华大学 | TOF depth data optimization method and device based on unsupervised data |
CN112348947A (en) * | 2021-01-07 | 2021-02-09 | 南京理工大学智能计算成像研究院有限公司 | Three-dimensional reconstruction method for deep learning based on reference information assistance |
CN112348947B (en) * | 2021-01-07 | 2021-04-09 | 南京理工大学智能计算成像研究院有限公司 | Three-dimensional reconstruction method for deep learning based on reference information assistance |
CN113781442A (en) * | 2021-09-13 | 2021-12-10 | 安徽农业大学 | Rice seedling plant three-dimensional structure model reconstruction method |
CN113781442B (en) * | 2021-09-13 | 2024-05-31 | 安徽农业大学 | Three-dimensional structure model reconstruction method for rice seedling plants |
Also Published As
Publication number | Publication date |
---|---|
CN109801375B (en) | 2021-02-09 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109801375A (en) | Porous material three-dimensional reconstruction method based on depth convolution confrontation neural network | |
CN109800516A (en) | A kind of porous material flow field model building method based on DCGAN | |
CN111968053B (en) | Image restoration method based on gate-controlled convolution generation countermeasure network | |
CN107194872B (en) | Remote sensed image super-resolution reconstruction method based on perception of content deep learning network | |
CN108961272B (en) | Method for generating skin disease image based on deep convolution countermeasure generation network | |
CN107709699B (en) | Generating three-dimensional micromodel of porous rock sample | |
CN110827408B (en) | Real-time three-dimensional reconstruction method based on depth sensor | |
CN107679543A (en) | Sparse autocoder and extreme learning machine stereo image quality evaluation method | |
CN111028335B (en) | Point cloud data block surface patch reconstruction method based on deep learning | |
CN108053478A (en) | A kind of particulate reinforced composite finite element modeling method based on pixel theory | |
CN109614874A (en) | A kind of Human bodys' response method and system based on attention perception and tree-like skeleton point structure | |
CN113281239B (en) | Multi-scale coal rock pore network generation method and device | |
CN108986218A (en) | A kind of building point off density cloud fast reconstructing method based on PMVS | |
CN108153706A (en) | A kind of ultra-large calculating grid reconstruction method | |
CN112348831A (en) | Shale SEM image segmentation method based on machine learning | |
CN116051382A (en) | Data enhancement method based on deep reinforcement learning generation type antagonistic neural network and super-resolution reconstruction | |
CN102722906B (en) | Feature-based top-down image modeling method | |
CN109583450A (en) | Salient region detecting method based on feedforward neural network fusion vision attention priori | |
CN111724331A (en) | Porous medium image reconstruction method based on generation network | |
CN104050639B (en) | Multi-view dense point cloud data fusion method based on two-sided filter | |
DE102020123979A1 (en) | Defect detection for semiconductor structures on a wafer | |
Hoppe et al. | Hydrophobicity of myxomycete spores: An undescribed aspect of spore ornamentation | |
CN108898594A (en) | A kind of detection method of homogeneous panel defect | |
CN101739676A (en) | Method for manufacturing face effigy with ultra-low resolution | |
CN112734933B (en) | Method for reducing three-dimensional structure of non-woven material through central axis of fiber |
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 |