CN110210119A - A kind of high efficiency phase developing method based on deep layer convolutional neural networks - Google Patents
A kind of high efficiency phase developing method based on deep layer convolutional neural networks Download PDFInfo
- Publication number
- CN110210119A CN110210119A CN201910467171.8A CN201910467171A CN110210119A CN 110210119 A CN110210119 A CN 110210119A CN 201910467171 A CN201910467171 A CN 201910467171A CN 110210119 A CN110210119 A CN 110210119A
- Authority
- CN
- China
- Prior art keywords
- phase
- neural networks
- convolutional neural
- deep layer
- layer convolutional
- 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.)
- Pending
Links
- 238000013527 convolutional neural network Methods 0.000 title claims abstract description 61
- 238000000034 method Methods 0.000 title claims abstract description 46
- 238000012549 training Methods 0.000 claims abstract description 34
- 230000004913 activation Effects 0.000 claims abstract description 10
- 230000008569 process Effects 0.000 claims abstract description 10
- 238000010606 normalization Methods 0.000 claims abstract description 7
- 238000005070 sampling Methods 0.000 claims abstract description 4
- 238000004422 calculation algorithm Methods 0.000 claims description 10
- 238000012360 testing method Methods 0.000 claims description 4
- 238000012545 processing Methods 0.000 claims description 2
- 238000012706 support-vector machine Methods 0.000 claims description 2
- 238000005259 measurement Methods 0.000 abstract description 12
- 230000006870 function Effects 0.000 description 17
- 238000013135 deep learning Methods 0.000 description 6
- 238000000605 extraction Methods 0.000 description 4
- 238000004458 analytical method Methods 0.000 description 3
- 230000000694 effects Effects 0.000 description 3
- 241000208340 Araliaceae Species 0.000 description 2
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 description 2
- 235000003140 Panax quinquefolius Nutrition 0.000 description 2
- 238000013473 artificial intelligence Methods 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 235000008434 ginseng Nutrition 0.000 description 2
- 238000000691 measurement method Methods 0.000 description 2
- 230000003287 optical effect Effects 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 238000011084 recovery Methods 0.000 description 2
- 230000001154 acute effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 230000000052 comparative effect Effects 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000004880 explosion Methods 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 230000001537 neural effect Effects 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 238000010587 phase diagram Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
-
- 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T17/00—Three dimensional [3D] modelling, e.g. data description of 3D objects
Abstract
The present invention provides a kind of high efficiency phase developing methods based on deep layer convolutional neural networks.Method includes the following steps: step 1: obtaining wrapped phase by software emulation and phase wraps up the data of number, and thus establish training sample database;Step 2: using convolutional layer, pond layer, batch normalization, activation primitive ReLu, up-sampling layer and Softmax classifier, building has the deep layer convolutional neural networks of residual error access;Step 3: the data set that step 1 is obtained pre-processes, and pretreated image obtains network model parameter as training data, training deep layer convolutional neural networks model;Step 4: inputting wrapped phase to be deployed, using the convolutional neural networks model in step 3, wrapped phase is unfolded and is visualized.The present invention solves the Construct question of sample database, under the premise of guaranteeing measurement efficiency, realizes high-precision phase unwrapping.
Description
Technical field
The present invention relates to a kind of high efficiency phase developing methods based on deep layer convolutional neural networks, belong to optics, calculate
Machine vision and field of artificial intelligence.
Background technique
In recent years, with the development of computer technology, computer measurement meter is become for the three dimension profile measurement of object
One important branch in calculation field is widely used in every field, such as recognition of face and reconstruction, satellite radar interference are surveyed
Amount etc..In many three dimension profile measurement methods, the optical 3-dimensional surface shape measurement method based on phase analysis class, due to having
Non-contact, the advantages that measuring speed is fast, measurement accuracy is high, extensive concern and research are obtained.And the light based on phase analysis
Three steps: the mapping of phase recovery, phase unwrapping and phase to 3 d shape depth can be substantially divided by learning measurement.
In the three dimension profile measurement based on phase analysis, since during phase recovery, phase distribution is to pass through
Arctan function operation obtains, thus the phase value being calculated be truncated arctan function codomain (- π, π] in, claim
Phase is wrapped up.In order to enable phase completely embodies the surface condition of three-dimension object, need to carry out the phase wrapped up
Amendment, is allowed to become absolute phase, such process is referred to as phase unwrapping.
Due to the complexity of Phase unwrapping, the precision of phase unwrapping has been largely fixed entire measuring system
Measurement accuracy.Influencing the factor of phase unwrapping precision, to be common in interference noise, the acute variation of phase and phase discontinuous etc.,
Currently used phase developing method is broadly divided into time phase expansion and space phase expansion, the former needs several to measure image
Auxiliary, to realize dynamic, more demanding to the frame per second of hardware device rapid survey;The latter is based on single width phase diagram, but
Measurement accuracy and robustness are lower, and computationally intensive, are not able to satisfy the demand of practical rapid survey still.There is presently no take into account Shandong
The method of stick and measurement efficiency.
Deep learning develops intimately in recent years, and is proved to possess powerful ability in feature extraction.Artificial intelligence,
Many fields such as computer vision and optical measurement, the method based on deep learning in most cases represent current field
The interior attainable state-of-the-art level of institute.Deep learning is a frame, contains many important algorithms, such as convolutional Neural
Network, autocoder, limited Boltzmann machine, feedback cycle neural network etc..
As a classical solution in deep learning, the status in deep learning field is convolutional neural networks
It is self-evident.Theoretically, deeper convolutional neural networks possess more powerful feature extraction and classification capacity, and
In practice, degenerate problem is produced as depth increases due to deep layer convolutional neural networks, network is caused to be difficult to instruct well
Practice.For degenerate problem, the method for comparative maturity is solved using the structure with residual error access, such as residual error network in field
Certainly this problem.The powerful classification capacity of deep layer convolutional neural networks, and residual error access of having arranged in pairs or groups exactly is utilized in the present invention
Network training is helped, the purpose of phase unwrapping is realized.
Summary of the invention
To solve the above problems, the invention discloses a kind of high efficiency phases based on deep layer convolutional neural networks
Position method of deploying, solves the Construct question of sample database, under the premise of guaranteeing measurement efficiency, realizes high-precision phase
Position expansion.
Above-mentioned purpose is achieved through the following technical solutions:
A kind of high efficiency phase developing method based on deep layer convolutional neural networks, this method comprises the following steps:
Step 1: wrapped phase being obtained by software emulation and phase wraps up the data of number, and thus establishes training sample
Database;
Step 2: being classified using convolutional layer, pond layer, batch normalization, activation primitive ReLu, up-sampling layer and Softmax
Device, building have the deep layer convolutional neural networks of residual error access;
Step 3: the data set that step 1 is obtained pre-processes, and pretreated image is as training data, training deep layer volume
Product neural network model, obtains network model parameter;
Step 4: inputting wrapped phase to be deployed, using the convolutional neural networks model in step 3, wrapped phase is unfolded
And it is visualized.
The high efficiency phase developing method based on deep layer convolutional neural networks generates described in step 1 in emulation
When wrapped phase, the carrier frequency of different frequency is superimposed to wrapped phase or is not superimposed, i.e., carrier frequency is 0, while basis
Wrapped phase before superposition, addition phase package number is as label, according to carrier frequency difference, establishes without carrier frequency, low carrier frequency,
Containing being less than, 5 complete radio-cycles, middle carrier frequency, that is, piece image are interior to contain 5 to 10 complete carrier frequency i.e. in piece image
Containing the database more than 10 complete radio-cycles in period and high carrier frequency, i.e. piece image, format is carried out for data
Standardization processing, treated wrapped phase picture and package number label are 8 single channel images.
The high efficiency phase developing method based on deep layer convolutional neural networks generates described in step 1 in emulation
When wrapped phase, wrapped phasePhase Φ (x, y), phase package number k (x, y) and carrier phase main value f is unfolded
(x, y) meets:
Wherein, phase package number k (x, y) is integer;Wrapped phaseCodomain be [0,2 π);Carrier phase f
The codomain of (x, y) be [0,2 π).
The high efficiency phase developing method based on deep layer convolutional neural networks, in building deep layer described in step 2
During convolutional neural networks, in order to avoid degenerate problem occur in deep layer convolutional neural networks, rolled up in the deep layer of phase unwrapping
Residual error access is added in product neural network;Add or delete part residual error access, or using residual error network or in which residual error mould
Block and its modified version realized this step function originally.
The high efficiency phase developing method based on deep layer convolutional neural networks, in building deep layer described in step 2
During convolutional neural networks, classified using Softmax classifier in deep layer convolutional neural networks end;Or it uses
Support vector machine classifier, k close on classifier classifier and realize this step function.
The high efficiency phase developing method based on deep layer convolutional neural networks, in building deep layer described in step 2
Nonlinear activation function or use during convolutional neural networks, using ReLu function as deep layer convolutional neural networks
Sigmoid function or hyperbolic tangent function realize this step function.
The high efficiency phase developing method based on deep layer convolutional neural networks, after being pre-processed described in step 3
Phase data be divided into training sample, verifying sample and test sample three parts, with training sample training deep layer convolutional Neural net
Network updates network parameter using the training method and back-propagation algorithm for having supervision, uses graphics processor (GPU) training deep layer
Convolutional neural networks, meanwhile, using verifying sample observation grid training process and its performance during training, finally, will
Test sample input deep layer convolutional neural networks are tested.
The utility model has the advantages that
In contrast to traditional phase developing method, method provided by the invention can take into account high efficiency and high-precision, simultaneously
It obtains efficiency and precision is better than the solution of conventional method;
In contrast to other phase developing methods, such as Schwartzkopf based on deep learning et al. proposition based on preceding
The phase developing method of Multilayer perceptron network is presented, phase developing method of the invention has higher precision, and phase
Position expansion result is unrelated with path;
In contrast to the convolutional neural networks that Spoorthi et al. is used, network committed memory of the invention may be significantly smaller,
Precision is higher, and apparent more easily trained.
Detailed description of the invention
Fig. 1 is total algorithm schematic diagram of the invention, and training and utilization including deep layer convolutional neural networks are trained
The Principle of Process of network progress phase unwrapping.
Fig. 2 is the schematic network structure for the deep layer convolutional neural networks that the present invention uses.
Specific embodiment
Refering to fig. 1, for the present invention in order to solve the Phase unwrapping, the technical solution used is to provide a kind of base
In the method for deep layer convolutional neural networks, comprising the following steps:
Step 1: building data set.In a plane, high by the random two dimension of four positions, standard deviation, peak values
The superposition of this function, to simulate the absolute phase of the 3 d shape in objective reality;Carrier phase is superimposed in absolute phase
Main value, and the main value of superimposed result is taken to simulate the wrapped phase by measuring after three dimension profile measurement;Wrapped phase
During corresponding phase package number takes main value by absolute phase, the number of cycles intercepted is determined.It is following to be based on
The code of Matlab language is the important algorithm code in the embodiment for construct data set and data prediction.
The absolute phase of 3 d shape is first simulated,
Phase (x, y)=exp (- (((x-128)+ox1) .^2+ ((y-128)+oy1) .^2) ./sig1^2) * p1+exp
(-(((x-128)+ox2).^2+((y-128)+oy2).^2)./sig2^2)*p2+exp(-(((x-128)+ox3).^2+((y-
128)+oy3).^2)./sig3^2)*p3+exp(-(((x-128)+ox4).^2+((y-128)+oy4).^2)./sig4^2)*
p4;
Wherein, ox1, ox2, ox3, ox4, oy1, oy2, oy3, oy4 are the 8 Gaussian function positional shift being randomly generated ginsengs
Number;Sig1, sig2, sig3, sig4 are 4 Gaussian function standard deviation criterias being randomly generated;P1, p2, p3, p4 are 4 random
The Gaussian function peak parameters of generation, and this parameter is positive and negative random.
Then wrapped phase corresponding to above-mentioned 3 d shape absolute phase and phase package number are simulated,
N (x, y)=floor (phase (x, y));
Phasei (x, y)=phase (x, y)-n (x, y)+carry (x, y)-floor (phase (x, y)-n (x, y)+
carry(x,y));Wherein, n (x, y) is the phase package number simulated;Carry (x, y) is the master of the carrier phase of simulation
Value;Phasei (x, y) is the wrapped phase simulated.
Finally the phase data of acquisition is pre-processed, and phase data library is written.
Step 2: being classified using convolutional layer, pond layer, batch normalization, activation primitive ReLu, up-sampling layer and Softmax
Device, building have the deep layer convolutional neural networks of residual error access.
Convolutional layer: convolution algorithm is carried out using image of the different convolution kernels to input, to obtain a sheet by a sheet characteristic pattern
Operation layer, be module very traditional in convolutional neural networks.
Pond layer: compressing the data of input, removes unessential sample in characteristic pattern, to reduce trained ginseng
Number reduces network internal storage consumption, prevents the module of network over-fitting, be the conventional module in convolutional neural networks.Pond layer master
There are average pondization and two kinds of maximum value pondization, rule of thumb, the present invention selects maximum value pond.
Criticize normalization layer: the layer that the result of convolutional layer output is normalized, in order to avoid instructing in network
Occur gradient in experienced process to disappear and gradient explosion phenomenon.Criticizing the convolutional neural networks field of normalization operation in recent years is
One more orthodox operation.
Activation primitive ReLu: the purpose of activation primitive be introduced during network training it is non-linear, to avoid ladder
Degree disappears, and the activation primitive used in the present embodiment can also realize this using activation primitives such as such as sigmoid, tanh for ReLu
Function.
Up-sample layer: corresponding with pond layer, for the feature that releasing network extracts, the purpose used is so that image
Size it is identical as wrapped phase picture size to be deployed.
Softmax classifier: being the mainstream classifier for realizing classification in recent years, using other classifiers such as supporting vector
Machine (SVM) etc., can also realize this function.
Residual error access: since network characteristic of the invention is that level is deeper, the superpower of deep layer convolutional neural networks is utilized
Ability in feature extraction, and deep layer network bring side effect is to cause to be difficult to trained problem since network is degenerated.In order to solve
This problem, widespread practice is introducing residual error access (bibliography " the Deep Residual in network in field
Learning for Image Recognition ", author Kaiming He etc.).Residual error access only uses initial data
Scale operation, there is no the operations for carrying out complicated to avoid to remain the gradient of initial data to the full extent
The problem of network is degenerated.
Referring to Fig.2, figure is one embodiment model of deep layer convolutional neural networks provided by the invention.Novelty of the invention
Place is that taking the lead in combining the newest technology such as residual error access, batch normalization, Softmax classifier carries out phase unwrapping network
It builds, due to the reference of residual error access, network of the invention possesses very deep depth, shares 38 concatenated convolutional layers, guarantees
Powerful ability in feature extraction.Meanwhile the present invention utilizes various optimisation techniques, while guaranteeing the operational effect of network,
The request memory of network training 4GB is compressed to, it is sufficient to meet the configuration requirement of most mainstream hardwares.If using deep
The half precision learning framework in learning areas more forward position is spent, the request memory of network training can be down to 2GB.
Step 3: the data set that step 1 is obtained pre-processes, and pretreated image is as training data, using there is supervision
Back-propagation algorithm training deep layer convolutional neural networks model, obtain network model parameter.Detailed process is as follows.
It is that the Caffe deep learning frame of the invention used can by the data prediction in the database obtained in step 1
With the data of receiving, the more orthodox back-propagation algorithm of use has used the cross entropy of Softmax as the damage of the present embodiment
Function is lost, has cooperated momentum (Momentum) optimization algorithm, has updated the parameter of network.Meanwhile by repeatedly attempting, I is obtained
One group of effect relatively good network training hyper parameter.Learning rate takes 0.0008, and momentum takes 0.9, and batch size takes 4, by 50-
The training of 100 periods (epoch), may finally obtain relatively good training result.The network hyper parameter provided is carried out
Certain fine tuning can also realize the purpose of this step.
Step 4: refering to fig. 1, using trained deep layer convolutional neural networks model in step 3, in network inputs
Part inputs wrapped phase information to be deployed, wraps up number by the phase that wrapped phase can be obtained in network operations.In conjunction with
The present invention writes visual code in Python, may be implemented what wrapped phase and phase the package number in Fig. 1 were superimposed
Function, the final purpose for realizing phase unwrapping, obtains high-precision absolute phase.
Above content is only one embodiment of the present of invention, by adding or deleting part convolutional layer, adding or deleting portion
Divide residual error access, the hyper parameter of the part or all of network training of change or using other network training optimization algorithms, the present invention
Provided network structure can also there are many variants, in those skilled in the art in the premise for not making creative labor
Under, the network variant that network obtains is changed through the above way to be included within the scope of the present invention.
Claims (7)
1. a kind of high efficiency phase developing method based on deep layer convolutional neural networks, it is characterised in that: this method includes as follows
Step:
Step 1: wrapped phase being obtained by software emulation and phase wraps up the data of number, and thus establishes training sample data
Library;
Step 2: utilizing convolutional layer, pond layer, batch normalization, activation primitive ReLu, up-sampling layer and Softmax classifier, structure
Build the deep layer convolutional neural networks with residual error access;
Step 3: the data set that step 1 is obtained pre-processes, and pretreated image is as training data, training deep layer convolution mind
Through network model, network model parameter is obtained;
Step 4: inputting wrapped phase to be deployed, using the convolutional neural networks model in step 3, expansion wrapped phase simultaneously will
It is visualized.
2. the high efficiency phase developing method according to claim 1 based on deep layer convolutional neural networks, it is characterised in that:
Described in step 1 when emulation generates wrapped phase, the carrier frequency of different frequency is superimposed to wrapped phase or is not superimposed, i.e.,
Carrier frequency is 0, while according to the wrapped phase before superposition, adding phase package number as label, not according to carrier frequency
Together, it establishes without carrier frequency, low carrier frequency, i.e., containing less than in 5 complete radio-cycles, middle carrier frequency, that is, piece image in piece image
Containing 5 to 10 complete radio-cycles and high carrier frequency, i.e., contain the number more than 10 complete radio-cycles in piece image
According to library, format specification processing is carried out for data, treated wrapped phase picture and package number label are 8 lists
Channel image.
3. the high efficiency phase developing method according to claim 1 based on deep layer convolutional neural networks, it is characterised in that:
Described in step 1 when emulation generates wrapped phase, wrapped phasePhase Φ (x, y) is unfolded, phase wraps up number k
(x, y) and carrier phase main value f (x, y) meet:
Wherein, phase package number k (x, y) is integer;Wrapped phaseCodomain be [0,2 π);Carrier phase f (x, y)
Codomain be [0,2 π).
4. the high efficiency phase developing method according to claim 1 based on deep layer convolutional neural networks, it is characterised in that:
Described in step 2 during constructing deep layer convolutional neural networks, asked in order to avoid degenerating occur in deep layer convolutional neural networks
Topic adds residual error access in the deep layer convolutional neural networks of phase unwrapping;Part residual error access is added or deleted, or using residual
Poor network or in which residual error module and its modified version realized this step function originally.
5. the high efficiency phase developing method according to claim 1 based on deep layer convolutional neural networks, it is characterised in that:
Described in step 2 during constructing deep layer convolutional neural networks, using Softmax classifier in deep layer convolutional neural networks
Classify end;Or support vector machine classifier is used, k closes on classifier classifier and realizes this step function.
6. the high efficiency phase developing method according to claim 1 based on deep layer convolutional neural networks, it is characterised in that:
Described in step 2 during constructing deep layer convolutional neural networks, using ReLu function as deep layer convolutional neural networks
Nonlinear activation function realizes this step function using sigmoid function or hyperbolic tangent function.
7. the high efficiency phase developing method according to claim 1 based on deep layer convolutional neural networks, it is characterised in that:
Pretreated phase data is divided into training sample, verifying sample and test sample three parts described in step 3, with training sample
This training deep layer convolutional neural networks update network parameter using the training method and back-propagation algorithm for having supervision, use figure
Shape processor (GPU) trains deep layer convolutional neural networks, meanwhile, using verifying sample observation grid training during training
Process and its performance, finally, test sample input deep layer convolutional neural networks are tested.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910467171.8A CN110210119A (en) | 2019-05-30 | 2019-05-30 | A kind of high efficiency phase developing method based on deep layer convolutional neural networks |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910467171.8A CN110210119A (en) | 2019-05-30 | 2019-05-30 | A kind of high efficiency phase developing method based on deep layer convolutional neural networks |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110210119A true CN110210119A (en) | 2019-09-06 |
Family
ID=67789821
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910467171.8A Pending CN110210119A (en) | 2019-05-30 | 2019-05-30 | A kind of high efficiency phase developing method based on deep layer convolutional neural networks |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110210119A (en) |
Cited By (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110500957A (en) * | 2019-09-10 | 2019-11-26 | 中国科学院苏州纳米技术与纳米仿生研究所 | A kind of active three-D imaging method, device, equipment and storage medium |
CN110751268A (en) * | 2019-09-27 | 2020-02-04 | 北京理工大学 | Phase aliasing error removing method and device based on end-to-end convolutional neural network |
CN111189414A (en) * | 2020-01-09 | 2020-05-22 | 西安知象光电科技有限公司 | Real-time single-frame phase extraction method |
CN111351450A (en) * | 2020-03-20 | 2020-06-30 | 南京理工大学 | Single-frame stripe image three-dimensional measurement method based on deep learning |
CN111461224A (en) * | 2020-04-01 | 2020-07-28 | 西安交通大学 | Phase data unwrapping method based on residual self-coding neural network |
CN111561877A (en) * | 2020-04-24 | 2020-08-21 | 西安交通大学 | Variable resolution phase unwrapping method based on point diffraction interferometer |
CN111794741A (en) * | 2020-08-11 | 2020-10-20 | 中国石油天然气集团有限公司 | Method for realizing sliding directional drilling simulator |
CN111797678A (en) * | 2020-05-15 | 2020-10-20 | 华南师范大学 | Phase unwrapping method and device based on composite neural network |
CN111928794A (en) * | 2020-08-04 | 2020-11-13 | 北京理工大学 | Closed fringe compatible single interference diagram phase method and device based on deep learning |
CN112116616A (en) * | 2020-08-05 | 2020-12-22 | 西安交通大学 | Phase information extraction method based on convolutional neural network, storage medium and equipment |
CN112836422A (en) * | 2020-12-31 | 2021-05-25 | 电子科技大学 | Interference and convolution neural network mixed scheme measuring method |
CN113093379A (en) * | 2021-03-25 | 2021-07-09 | 上海交通大学 | Photon Itanium machine-oriented orthogonal space phase modulation method |
CN113314216A (en) * | 2021-06-01 | 2021-08-27 | 南方科技大学 | Functional brain network construction method and device, electronic equipment and readable storage medium |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103940371A (en) * | 2014-05-12 | 2014-07-23 | 电子科技大学 | High-precision three-dimensional shape measurement method for jump object |
CN109253708A (en) * | 2018-09-29 | 2019-01-22 | 南京理工大学 | A kind of fringe projection time phase method of deploying based on deep learning |
CN109307483A (en) * | 2018-11-20 | 2019-02-05 | 西南石油大学 | A kind of phase developing method based on structured-light system geometrical constraint |
-
2019
- 2019-05-30 CN CN201910467171.8A patent/CN110210119A/en active Pending
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103940371A (en) * | 2014-05-12 | 2014-07-23 | 电子科技大学 | High-precision three-dimensional shape measurement method for jump object |
CN109253708A (en) * | 2018-09-29 | 2019-01-22 | 南京理工大学 | A kind of fringe projection time phase method of deploying based on deep learning |
CN109307483A (en) * | 2018-11-20 | 2019-02-05 | 西南石油大学 | A kind of phase developing method based on structured-light system geometrical constraint |
Non-Patent Citations (2)
Title |
---|
G. E. SPOORTHI 等: "PhaseNet: A Deep Convolutional Neural Network for Two-Dimensional Phase Unwrapping", 《IEEE SIGNAL PROCESSING LETTERS》 * |
KAIQIANG WANG 等: "One-step robust deep learning phase unwrapping", 《OPTICS EXPRESS》 * |
Cited By (25)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110500957B (en) * | 2019-09-10 | 2021-09-14 | 中国科学院苏州纳米技术与纳米仿生研究所 | Active three-dimensional imaging method, device, equipment and storage medium |
CN110500957A (en) * | 2019-09-10 | 2019-11-26 | 中国科学院苏州纳米技术与纳米仿生研究所 | A kind of active three-D imaging method, device, equipment and storage medium |
CN110751268B (en) * | 2019-09-27 | 2022-07-26 | 北京理工大学 | Phase aliasing error removing method and device based on end-to-end convolutional neural network |
CN110751268A (en) * | 2019-09-27 | 2020-02-04 | 北京理工大学 | Phase aliasing error removing method and device based on end-to-end convolutional neural network |
CN111189414A (en) * | 2020-01-09 | 2020-05-22 | 西安知象光电科技有限公司 | Real-time single-frame phase extraction method |
CN111189414B (en) * | 2020-01-09 | 2021-09-03 | 西安知象光电科技有限公司 | Real-time single-frame phase extraction method |
CN111351450A (en) * | 2020-03-20 | 2020-06-30 | 南京理工大学 | Single-frame stripe image three-dimensional measurement method based on deep learning |
CN111351450B (en) * | 2020-03-20 | 2021-09-28 | 南京理工大学 | Single-frame stripe image three-dimensional measurement method based on deep learning |
CN111461224B (en) * | 2020-04-01 | 2022-08-16 | 西安交通大学 | Phase data unwrapping method based on residual self-coding neural network |
CN111461224A (en) * | 2020-04-01 | 2020-07-28 | 西安交通大学 | Phase data unwrapping method based on residual self-coding neural network |
CN111561877A (en) * | 2020-04-24 | 2020-08-21 | 西安交通大学 | Variable resolution phase unwrapping method based on point diffraction interferometer |
CN111561877B (en) * | 2020-04-24 | 2021-08-13 | 西安交通大学 | Variable resolution phase unwrapping method based on point diffraction interferometer |
CN111797678A (en) * | 2020-05-15 | 2020-10-20 | 华南师范大学 | Phase unwrapping method and device based on composite neural network |
CN111797678B (en) * | 2020-05-15 | 2023-07-07 | 华南师范大学 | Phase unwrapping method and device based on composite neural network |
CN111928794B (en) * | 2020-08-04 | 2022-03-11 | 北京理工大学 | Closed fringe compatible single interference diagram phase method and device based on deep learning |
CN111928794A (en) * | 2020-08-04 | 2020-11-13 | 北京理工大学 | Closed fringe compatible single interference diagram phase method and device based on deep learning |
CN112116616B (en) * | 2020-08-05 | 2022-06-07 | 西安交通大学 | Phase information extraction method based on convolutional neural network, storage medium and equipment |
CN112116616A (en) * | 2020-08-05 | 2020-12-22 | 西安交通大学 | Phase information extraction method based on convolutional neural network, storage medium and equipment |
CN111794741A (en) * | 2020-08-11 | 2020-10-20 | 中国石油天然气集团有限公司 | Method for realizing sliding directional drilling simulator |
CN111794741B (en) * | 2020-08-11 | 2023-08-18 | 中国石油天然气集团有限公司 | Method for realizing sliding directional drilling simulator |
CN112836422A (en) * | 2020-12-31 | 2021-05-25 | 电子科技大学 | Interference and convolution neural network mixed scheme measuring method |
CN112836422B (en) * | 2020-12-31 | 2022-03-18 | 电子科技大学 | Interference and convolution neural network mixed scheme measuring method |
CN113093379A (en) * | 2021-03-25 | 2021-07-09 | 上海交通大学 | Photon Itanium machine-oriented orthogonal space phase modulation method |
CN113093379B (en) * | 2021-03-25 | 2022-02-25 | 上海交通大学 | Photon Itanium machine-oriented orthogonal space phase modulation method |
CN113314216A (en) * | 2021-06-01 | 2021-08-27 | 南方科技大学 | Functional brain network construction method and device, electronic equipment and readable storage medium |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110210119A (en) | A kind of high efficiency phase developing method based on deep layer convolutional neural networks | |
US11574097B2 (en) | Deep learning based identification of difficult to test nodes | |
Howard et al. | Mobilenets: Efficient convolutional neural networks for mobile vision applications | |
US20190295228A1 (en) | Image in-painting for irregular holes using partial convolutions | |
US20190035113A1 (en) | Temporally stable data reconstruction with an external recurrent neural network | |
US20190294972A1 (en) | Representing a neural network utilizing paths within the network to improve a performance of the neural network | |
CN110176054A (en) | For training the generation of the composograph of neural network model | |
CN108416327A (en) | A kind of object detection method, device, computer equipment and readable storage medium storing program for executing | |
US11062471B1 (en) | Neural network system for stereo image matching | |
CN112950775A (en) | Three-dimensional face model reconstruction method and system based on self-supervision learning | |
US11961001B2 (en) | Parallel forward and backward propagation | |
CN111210498A (en) | Reducing the level of detail of a polygon mesh to reduce the complexity of rendered geometry | |
CN109936745A (en) | For improving the method and system of the decompression of original video data | |
Denninger et al. | 3d scene reconstruction from a single viewport | |
DE102021121109A1 (en) | RECOVERY OF THREE-DIMENSIONAL MODELS FROM TWO-DIMENSIONAL IMAGES | |
DE102019106996A1 (en) | PRESENTING A NEURONAL NETWORK USING PATHS INSIDE THE NETWORK TO IMPROVE THE PERFORMANCE OF THE NEURONAL NETWORK | |
DE102022107232A1 (en) | PACKED ERROR CORRECTION CODE (ECC) FOR COMPRESSED PRIVACY | |
CN116736624A (en) | Parallel mask rule checking for evolving mask shapes in an optical proximity correction stream | |
Hattori et al. | Learning self-prior for mesh denoising using dual graph convolutional networks | |
DE112019001978T5 (en) | IMPROVING THE REALISM OF SCENES WITH WATER SURFACES DURING RENDERING | |
CN103678888B (en) | The flowing of a kind of heart blood based on Euler's fluid simulation algorithm schematically shows method | |
DE102021114013A1 (en) | TECHNIQUES FOR EFFICIENT SCANNING OF AN IMAGE | |
Mlakar et al. | Subdivision‐specialized linear algebra kernels for static and dynamic mesh connectivity on the gpu | |
CN106910246A (en) | Speckle three-D imaging method and device that space-time is combined | |
CN112861977B (en) | Migration learning data processing method, system, medium, equipment, terminal and application |
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 | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190906 |