CN108629134A - A kind of similitude intensifying method in field small in manifold - Google Patents
A kind of similitude intensifying method in field small in manifold Download PDFInfo
- Publication number
- CN108629134A CN108629134A CN201810446278.XA CN201810446278A CN108629134A CN 108629134 A CN108629134 A CN 108629134A CN 201810446278 A CN201810446278 A CN 201810446278A CN 108629134 A CN108629134 A CN 108629134A
- Authority
- CN
- China
- Prior art keywords
- image
- manifold
- hidden variable
- similitude
- random
- 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
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T9/00—Image coding
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Computer Hardware Design (AREA)
- Evolutionary Computation (AREA)
- Geometry (AREA)
- General Engineering & Computer Science (AREA)
- Multimedia (AREA)
- Compression Or Coding Systems Of Tv Signals (AREA)
Abstract
A kind of similitude intensifying method in field small in manifold, includes the following steps：1) image random perturbation：The image of input is disturbed according to certain strength of turbulence, completely new image of the output one after disturbance；2) hidden variable random perturbation：The encoded rear hidden variable generated is inputted, hidden variable is disturbed according to certain strength of turbulence, exports the hidden variable after disturbance；3) manifold constructs：It maps an image in a smooth manifold, under the premise of same strength of turbulence, two images that same image is generated through described two perturbation techniques are almost similar.The present invention provides a kind of similitude intensifying method in field small in manifold, realizes the novelty constraint of parametric method, on the one hand increases the similitude of the solution in popular upper small neighbourhood, on the other hand, increases more novel solutions for manifold, reduces the accounting of solution inferior.
Description
Technical field
The invention belongs to field of computer aided design, and in particular to the cross-sectional image of device modeling is in the base by coding
On plinth, the method for finding optimal moulding.
Background technology
With the development of computer technology, computer picture has been widely applied among each field at present.It is set in industry
Meter field is widely used computer to assist carrying out the design work of industrial equipment especially accurate device.
In the design of current device element, designer is mainly still designed using the priori of the mankind new
It makes.But the optimal moulding of many parts of appliance tends not to be immediately arrived at according to the professional knowledge of priori so that new modeling is set
Meter efficiency is greatly affected.Such as in the wind wheel moulding of design optimization wind-driven generator, under complicated use environment, it is very difficult to
Optimal wind wheel moulding is directly extrapolated according to existing knowwhy.
Then we expect the cross-sectional image by device element by way of coding on dimensionality reduction to the hidden variable of low-dimensional, into
And in the spatially searching optimal solution of low-dimensional.But it can not be directly by the way that device modeling hidden variable be mapped to a manifold sky
Between it is upper tested for optimizing, main reason is that the similarity hypothesis of the nonlinearity of image encoder and optimizing algorithm
To contradiction.Random optimizing algorithm often assumes that all solutions in manifold in a certain contiguous range are all similar mutually, however image is certainly
There is a large amount of minimum point in encoder, largely use nonlinearity high in network structure or even do not connect in the training process
Continuous computing unit.
Invention content
In view of the foregoing deficiencies of prior art, the present invention provides a kind of similitude reinforcing side in field small in manifold
Method realizes the novelty constraint of parametric method, on the one hand increases the similitude of the solution in popular upper small neighbourhood, on the other hand,
Increase more novel solutions for manifold, reduces the accounting of solution inferior.
The technical solution adopted by the present invention to solve the technical problems is：
A kind of similitude intensifying method in field small in manifold, the described method comprises the following steps：
1) image random perturbation：The image of input is disturbed according to certain strength of turbulence, one process of output is disturbed
Completely new image after dynamic；
2) hidden variable random perturbation：The encoded rear hidden variable generated is inputted, according to certain strength of turbulence to hidden variable
It is disturbed, exports the hidden variable after disturbance；
3) manifold constructs：It maps an image in a smooth manifold, under the premise of same strength of turbulence, same
Two images that image is generated through described two perturbation techniques are almost similar.
Further, in the step 3), before doing similitude reinforcing to image in manifold, can pre- place first be done to image
Reason.Sparse binarization operation is for example done image in the pretreatment, can improve the later stage do similitude reinforcing speed.
Further, in the step 2), normal state random distribution or uniformly random distribution are used according to a certain interference strength
Equal random digit generation methods do random perturbation on the different dimensions of image.
Further, in the step 1), encoding and decoding are carried out to image using confrontation self-encoding encoder, and generated after making coding
Hidden variable meet a certain predefined prior probability distribution.
Preferably, in the step 3), using the manifold constructing technology based on deep neural network, declined using gradient
Method updates the parameter of network, and decision device is enable to identify that input picture is from original image random perturbation or to come from hidden variable
Random perturbation.
Beneficial effects of the present invention are mainly manifested in：The optimal moulding of needs can be easily found in manifold.
Description of the drawings
Fig. 1 is the technology composition schematic diagram of similitude intensifying method in small field in manifold of the present invention.
Fig. 2 is structural schematic diagram of the similitude intensifying method in an example in small field in manifold of the present invention.
Fig. 3 is flow diagram of the similitude intensifying method in an example in small field in manifold of the present invention.
Original paper label declaration, similitude intensifying method, 21- image random perturbation technologies, 22- are hidden in small field in 2- manifolds
Variable random perturbation technology, 23- manifold constructing technologies.
Specific implementation mode
The invention will be further described below in conjunction with the accompanying drawings.
A kind of referring to Fig.1~Fig. 3, similitude intensifying method in field small in manifold, the described method comprises the following steps：
1) image random perturbation：The image of input is disturbed according to certain strength of turbulence, one process of output is disturbed
Completely new image after dynamic；
2) hidden variable random perturbation：The encoded rear hidden variable generated is inputted, according to certain strength of turbulence to hidden variable
It is disturbed, exports the hidden variable after disturbance；
3) manifold constructs：It maps an image in a smooth manifold, under the premise of same strength of turbulence, same
Two images that image is generated through described two perturbation techniques are almost similar.
Further, in the step 3), before doing similitude reinforcing to image in manifold, can pre- place first be done to image
Reason.Sparse binarization operation is for example done image in the pretreatment, can improve the later stage do similitude reinforcing speed.
Further, in the step 2), normal state random distribution or uniformly random distribution are used according to a certain interference strength
Equal random digit generation methods do random perturbation on the different dimensions of image.
Further, in the step 1), encoding and decoding are carried out to image using confrontation self-encoding encoder, and generated after making coding
Hidden variable meet a certain predefined prior probability distribution.
Preferably, in the step 3), using the manifold constructing technology based on deep neural network, declined using gradient
Method updates the parameter of network, and decision device is enable to identify that input picture is from original image random perturbation or to come from hidden variable
Random perturbation.
In image random perturbation step, when strength of turbulence is 0, output image is identical as input picture, exports hidden variable
It is identical as input hidden variable；
With the increase of described image random perturbation technology strength of turbulence, exporting the difference of image and input picture can also increase
Greatly, output hidden variable also will increase with the difference for inputting hidden variable.
Hidden variable disturbance treatment process includes with lower module：Image coding module, for image to be encoded to hidden variable；It is hidden
Variable disturbs module, for being disturbed by certain strength of turbulence to the hidden variable after coding；Image decoder module, for that will disturb
Hidden variable after dynamic is decoded as image.
In described image coding module, the image information of higher-dimension is encoded to the hidden variable data of low-dimensional；Described image solution
It is image by the hidden variable data convert of low-dimensional in code device.
In hidden variable perturbation steps, the disturbance hidden variable of output cannot significantly change the statistical probability for originally inputting hidden variable
Characteristic；The size that hidden variable may be implemented in a large amount of Repeated Disturbances is proportional to the traversal of each solution in the neighborhood of strength of turbulence.
The manifold Construction treatment process includes：Novelty decision device, for inputting the figure generated through image random perturbation
Picture, and the hidden variable generated through hidden variable random perturbation is decoded, the intensity that two images are disturbed is judged by novelty decision device
It is whether similar.The novelty decision device can be according to the parameter of output result and the difference update decision device of correct result.
Fig. 2 is a kind of construction module schematic diagram of embodiment of the present invention.
Described image random perturbation module generates the image after a disturbance according to a strength of turbulence.Disturb image with
The difference of original image increases with the increase of strength of turbulence.Image random perturbation device MB (x, s) is defined, as given original image x
Image o after a disturbance similar with x is provided with strength of turbulence s, perturbator MB (x, s), the difference of the two is with maximum probability direct ratio
In strength of turbulence s.
In the present embodiment, the realization means using discrete cosine transform (DCT) as image random perturbation device, use
Method is as follows：
1) two-dimensional dct transform is used to calculate the frequency domain matrix and jth image x of every image in image data basejFrequency
Domain matrix is Mj。
2) each element respective maximin in the database is calculated in each matrix, maximum value matrix MAX is constituted
With minimum value matrix MIN.
3) M is normalized to each matrixj=(Mj- MIN)/(MAX-MIN), wherein subtraction and division is matrix element
Grade operation.
4) with -1~1 be uniformly distributed generation one and MjThe identical random matrix R of size.
5) to MjCarry out random perturbation Ij=Mj+ R*s, wherein addition and multiplication operate for matrix element grade.
6) to IjIt carries out inverse normalization and inverse two-dimensional dct transform can be obtained the image o after random perturbationj。
7) with ojThe absolute value of middle all pixels value is threshold value to ojCarry out binaryzation.
8) use the decision device in image encoder to image ojJudged, if decision device thinks ojIt is not belonging to artwork
As the data of data set, then repeatedly step 4)~8), the otherwise image o after output disturbancej。
In some embodiments, step 4 can use Gaussian Profile or other random distribution modes to generate a random square
Battle array.
As shown in Fig. 2, hidden variable random perturbation module includes a self-encoding encoder and a hidden variable random perturbation device.Make
Image is encoded to the hidden variable data z of low-dimensional with the coded portion in self-encoding encoder, using hidden variable random perturbation device to hidden
Variable is disturbed, and the hidden variable after disturbance is finally decoded as a new picture by the decoded portion of self-encoding encoder.
It should be noted that the disturbance hidden variable of hidden variable random perturbation module output needs to obey the statistics of former hidden variable
Probability nature, and a large amount of Repeated Disturbances may be implemented the size of hidden variable and be proportional to each solution in the neighborhood of strength of turbulence
Traversal.
Under the premise of meeting the feature of the above hidden variable, in this present embodiment, the hidden variable random perturbation device is defined
For ML (z, s), specific implementation is as follows：
1) with -1~1 a random vector P identical with hidden variable z-dimension is generated for the uniformly random distribution of bound.
2) random perturbation v=z+P*s is carried out to hidden variable z, wherein addition and multiplication is the Element-Level operation of vector
3) random perturbation results of the output v as z
In some embodiments, the bound of the random number of generation can be defined as any other real number.
In some embodiments, random number, such as Gaussian Profile etc. can be generated according to other distribution of probability.
With reference to figure 2, the manifold constitutes module and generates network based on a confrontation to realize.By image random perturbation
For the image of module output as positive sample, the image that hidden variable disturbs module output is that negative sample inputs novelty arbiter, by
Novelty arbiter judges that picture is generated by which module.
In the present embodiment, the manifold building method is as follows：
1) disturbance strength of turbulence s is generated, and an image x is randomly selected from training library
2) image random perturbation device MB (x, s) is used to generate disturbance image o
3) it uses hidden variable random perturbation device ML (z, s) to generate disturbance hidden variable v, and disturbance image t is generated by encoder G
4) optimization formula (3), is allowed to obtain minimum value.Wherein D is novelty decision device, and LD is pair of novelty decision device
Anti- error function, LR (G) are predefined self-encoding encoder reconstructed error function.
LD (D)=E [log (1-D (o, x, s))]+E [log (D (t, x, s))] (1)
LD (G)=E [log (1-D (t, x, s))] (2)
minLR(G)+minLD(D)+minLD(G) (3)
5) step 1) is repeated to 4), until formula (3) is optimal and successfully builds manifold.
Illustrate that embodiments of the present invention, those skilled in the art can be by this specification below by way of specific specific example
Disclosed content understands other advantages and effect of the present invention easily.The present invention can also pass through in addition different specific realities
Example mode be embodied or practiced, the various details in this specification can also be based on different viewpoints with application, without departing from
Various modifications or alterations are carried out under the spirit of the present invention.It should be noted that in the absence of conflict, following embodiment and implementation
Feature in example can be combined with each other.
It should be noted that the diagram provided in following embodiment only illustrates the basic structure of the present invention in a schematic way
Think, component count ring will when only display is with related component in the present invention rather than according to actual implementation in illustrating then, reality
The quantity and ratio of each component can be a kind of random change when implementation.
Claims (5)
1. a kind of similitude intensifying method in field small in manifold, which is characterized in that the described method comprises the following steps：
1) image random perturbation：The image of input is disturbed according to certain strength of turbulence, output one is after disturbance
Completely new image；
2) hidden variable random perturbation：The encoded rear hidden variable generated is inputted, hidden variable is carried out according to certain strength of turbulence
Disturbance exports the hidden variable after disturbance；
3) manifold constructs：It maps an image in a smooth manifold, under the premise of same strength of turbulence, same image
Two images generated through described two perturbation techniques are almost similar.
2. a kind of similitude intensifying method in field small in manifold as described in claim 1, which is characterized in that the step
3) in, before doing similitude reinforcing to image in manifold, first image can be pre-processed.
3. a kind of similitude intensifying method in field small in manifold as claimed in claim 1 or 2, which is characterized in that described
In step 2), according to a certain interference strength using the random digit generation methods such as normal state random distribution or uniformly random distribution in image
Different dimensions on do random perturbation.
4. a kind of similitude intensifying method in field small in manifold as claimed in claim 1 or 2, which is characterized in that described
In step 1), encoding and decoding are carried out to image using confrontation self-encoding encoder, and it is a certain predetermined so that the hidden variable generated after coding is met
The prior probability distribution of justice.
5. a kind of similitude intensifying method in field small in manifold as claimed in claim 1 or 2, which is characterized in that described
In step 3), using the manifold constructing technology based on deep neural network, the parameter of network is updated using the method that gradient declines,
Decision device is set to identify that input picture is from original image random perturbation or to come from hidden variable random perturbation.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810446278.XA CN108629134B (en) | 2018-05-11 | 2018-05-11 | Similarity strengthening method for small fields in manifold |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810446278.XA CN108629134B (en) | 2018-05-11 | 2018-05-11 | Similarity strengthening method for small fields in manifold |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108629134A true CN108629134A (en) | 2018-10-09 |
| CN108629134B CN108629134B (en) | 2022-05-03 |
Family
ID=63692740
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810446278.XA Active CN108629134B (en) | 2018-05-11 | 2018-05-11 | Similarity strengthening method for small fields in manifold |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108629134B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109949278A (en) * | 2019-03-06 | 2019-06-28 | 西安电子科技大学 | Hyperspectral abnormity detection method based on confrontation autoencoder network |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102930283A (en) * | 2012-08-10 | 2013-02-13 | 合肥工业大学 | Self-adaptive robust constraint maximum variation mapping (CMVM) feature dimension reduction method for image retrieval of plant laminae |
| CN106691378A (en) * | 2016-12-16 | 2017-05-24 | 深圳市唯特视科技有限公司 | Deep learning vision classifying method based on electroencephalogram data |
| CN107895373A (en) * | 2017-11-14 | 2018-04-10 | 西安建筑科技大学 | Image partition method based on regional area uniformity manifold constraint MRF models |
| CN107977629A (en) * | 2017-12-04 | 2018-05-01 | 电子科技大学 | A kind of facial image aging synthetic method of feature based separation confrontation network |
-
2018
- 2018-05-11 CN CN201810446278.XA patent/CN108629134B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102930283A (en) * | 2012-08-10 | 2013-02-13 | 合肥工业大学 | Self-adaptive robust constraint maximum variation mapping (CMVM) feature dimension reduction method for image retrieval of plant laminae |
| CN106691378A (en) * | 2016-12-16 | 2017-05-24 | 深圳市唯特视科技有限公司 | Deep learning vision classifying method based on electroencephalogram data |
| CN107895373A (en) * | 2017-11-14 | 2018-04-10 | 西安建筑科技大学 | Image partition method based on regional area uniformity manifold constraint MRF models |
| CN107977629A (en) * | 2017-12-04 | 2018-05-01 | 电子科技大学 | A kind of facial image aging synthetic method of feature based separation confrontation network |
Non-Patent Citations (2)
| Title |
|---|
| 何俊等: "一种新的多角度人脸表情识别方法", 《计算机应用研究》 * |
| 王秀美等: "一种基于局部保持的隐变量模型", 《王秀美等》 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109949278A (en) * | 2019-03-06 | 2019-06-28 | 西安电子科技大学 | Hyperspectral abnormity detection method based on confrontation autoencoder network |
| CN109949278B (en) * | 2019-03-06 | 2021-10-29 | 西安电子科技大学 | Hyperspectral anomaly detection method based on antagonistic self-coding network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108629134B (en) | 2022-05-03 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Wang et al. | Hf-neus: Improved surface reconstruction using high-frequency details | |
| Pont et al. | Wasserstein distances, geodesics and barycenters of merge trees | |
| Zhu et al. | Robust registration of partially overlapping point sets via genetic algorithm with growth operator | |
| Cui et al. | Lightweight attention module for deep learning on classification and segmentation of 3-D point clouds | |
| CN115983148B (en) | CFD simulation cloud image prediction method, system, electronic equipment and medium | |
| Wang et al. | Discrete wavelet transform-based simple range classification strategies for fractal image coding | |
| CN117454495B (en) | CAD vector model generation method and device based on building sketch outline sequence | |
| Wang et al. | Deep joint source-channel coding for multi-task network | |
| Zhu et al. | Fast superpixel segmentation by iterative edge refinement | |
| Zhu et al. | Semantic image segmentation with shared decomposition convolution and boundary reinforcement structure | |
| US20220284720A1 (en) | Method for grouping cells according to density and electronic device employing method | |
| CN108629134A (en) | A kind of similitude intensifying method in field small in manifold | |
| CN110335196A (en) | A kind of super-resolution image reconstruction method and system based on fractal decoding | |
| CN113822825A (en) | Optical building target three-dimensional reconstruction method based on 3D-R2N2 | |
| CN114219701A (en) | Dunhuang fresco artistic style conversion method, system, computer equipment and storage medium | |
| US20220138554A1 (en) | Systems and methods utilizing machine learning techniques for training neural networks to generate distributions | |
| Li et al. | Towards communication-efficient digital twin via AI-powered transmission and reconstruction | |
| Bai et al. | A ginzburg-landau-h-1 model and its sav algorithm for image inpainting | |
| CN104320659B (en) | Background modeling method, device and equipment | |
| CN116186899A (en) | Data-driven supercritical airfoil pressure distribution prediction method, system and medium | |
| CN115310209A (en) | VAE-based pneumatic shape migration optimization method and related device | |
| CN116228014A (en) | DC power distribution network benefit evaluation system and method for infrastructure network access | |
| Čomić et al. | Dimension-independent multi-resolution Morse complexes | |
| CN116546219A (en) | Point cloud geometric color joint compression method based on learning | |
| Balmelli et al. | Efficient algorithms for embedded rendering of terrain models |
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 |