CN108629134B - Similarity strengthening method for small fields in manifold - Google Patents
Similarity strengthening method for small fields in manifold Download PDFInfo
- Publication number
- CN108629134B CN108629134B CN201810446278.XA CN201810446278A CN108629134B CN 108629134 B CN108629134 B CN 108629134B CN 201810446278 A CN201810446278 A CN 201810446278A CN 108629134 B CN108629134 B CN 108629134B
- Authority
- CN
- China
- Prior art keywords
- image
- disturbance
- random
- manifold
- hidden variable
- 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.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 26
- 238000005728 strengthening Methods 0.000 title abstract description 4
- 238000005516 engineering process Methods 0.000 claims abstract description 12
- 238000013507 mapping Methods 0.000 claims abstract description 5
- 239000011159 matrix material Substances 0.000 claims description 17
- 238000009826 distribution Methods 0.000 claims description 13
- 238000010276 construction Methods 0.000 claims description 7
- 238000007781 pre-processing Methods 0.000 claims description 5
- 230000002708 enhancing effect Effects 0.000 claims description 4
- 238000013528 artificial neural network Methods 0.000 claims description 3
- 238000011478 gradient descent method Methods 0.000 claims description 3
- 238000010606 normalization Methods 0.000 claims description 3
- 238000013459 approach Methods 0.000 claims description 2
- 238000009827 uniform distribution Methods 0.000 claims description 2
- 238000013461 design Methods 0.000 description 5
- 238000010586 diagram Methods 0.000 description 3
- 230000008569 process Effects 0.000 description 3
- 238000005457 optimization Methods 0.000 description 2
- 238000012549 training Methods 0.000 description 2
- 239000013598 vector Substances 0.000 description 2
- 241000282414 Homo sapiens Species 0.000 description 1
- 230000004075 alteration Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000011960 computer-aided design Methods 0.000 description 1
- 230000008094 contradictory effect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012803 optimization experiment Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
Images
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 similarity strengthening method for small domains on manifold comprises the following steps: 1) image random disturbance: disturbing an input image according to a certain disturbance intensity, and outputting a completely new disturbed image; 2) hidden variable random disturbance: inputting the encoded hidden variable, disturbing the hidden variable according to a certain disturbance intensity, and outputting the disturbed hidden variable; 3) the manifold structure: and mapping the image to a smooth manifold, wherein two images generated by the same image through the two disturbance technologies are nearly similar on the premise of the same disturbance intensity. The invention provides a similarity strengthening method in a small field of manifold, which realizes the novelty constraint of a parameterization method, on one hand, increases the similarity of solutions in a popular small neighborhood, on the other hand, adds more novel solutions for the manifold and reduces the ratio of poor solutions.
Description
Technical Field
The invention belongs to the field of computer aided design, and particularly relates to a method for searching an optimal model of a cross-sectional image of an equipment model on the basis of coding.
Background
With the development of computer technology, computer images are widely used in various fields. In the field of industrial design, computers have been widely used to assist in the design work of industrial equipment, particularly precision devices.
In the design of current equipment devices, designers mainly use a priori knowledge of human beings to design new models. However, the optimal models of many equipment components cannot be directly obtained according to prior professional knowledge, so that the design efficiency of a new model is greatly influenced. For example, when the wind wheel model of the wind driven generator is designed and optimized, under a complex use environment, the optimal wind wheel model is difficult to directly calculate according to the existing theoretical knowledge.
Therefore, the sectional image of the equipment device is reduced to a low-dimensional hidden variable in a coding mode, and an optimal solution is further searched in a low-dimensional space. But it cannot be directly used for the optimization experiment by mapping the hidden variables of the device model onto a manifold space, and the main reason is that the high non-linearity of the image encoder is contradictory to the similarity assumption of the optimization algorithm. The random optimization algorithm usually assumes that all solutions in a certain neighborhood range on the manifold are similar to each other, however, a large number of minimum value points exist in the training process of the image self-encoder, and a large number of calculation units with extremely high nonlinearity or even discontinuity are adopted in the network structure.
Disclosure of Invention
In view of the above drawbacks of the prior art, the present invention provides a similarity enhancement method for small domains on a manifold, which implements the novelty constraint of a parameterization method, on one hand, increases the similarity of solutions in popular small neighborhoods, and on the other hand, adds more novel solutions to the manifold, and reduces the ratio of poor solutions.
The technical scheme adopted by the invention for solving the technical problems is as follows:
a method for enhancing similarity of small domains on manifold, the method comprising the steps of:
1) image random disturbance: disturbing an input image according to a certain disturbance intensity, and outputting a completely new disturbed image;
2) hidden variable random disturbance: inputting the encoded hidden variable, disturbing the hidden variable according to a certain disturbance intensity, and outputting the disturbed hidden variable;
3) the manifold structure: and mapping the image to a smooth manifold, wherein two images generated by the same image through the two disturbance technologies are nearly similar on the premise of the same disturbance intensity.
Further, in the step 3), before similarity enhancement is performed on the image in the manifold, preprocessing is performed on the image. The preprocessing, such as performing sparse binarization operation on the image, can improve the speed of performing similarity enhancement in the later period.
Further, in the step 2), random disturbance is performed on different dimensions of the image by using a random number generation method such as normal random distribution or uniform random distribution according to a certain interference intensity.
Furthermore, in the step 1), the image is coded and decoded by using a confrontation self-coder, and the hidden variables generated after coding are made to conform to a certain predefined prior probability distribution.
Preferably, in the step 3), a manifold construction technology based on a deep neural network is adopted, and a gradient descent method is used to update parameters of the network, so that the determiner can identify whether the input picture is from random disturbance of the original image or from random disturbance of a hidden variable.
The invention has the following beneficial effects: the required optimal modeling can be conveniently found on the manifold.
Drawings
FIG. 1 is a schematic diagram of the similarity enhancement method in the small domain of manifold according to the present invention.
FIG. 2 is a schematic structural diagram of a similarity enhancement method in a small area of manifold according to an embodiment of the present invention.
FIG. 3 is a flow chart illustrating an exemplary method for enhancing similarity in a small manifold area according to the present invention.
Original label description, 2-similarity enhancement method in small field on manifold, 21-image random disturbance technology, 22-hidden variable random disturbance technology and 23-manifold construction technology.
Detailed Description
The invention is further described below with reference to the accompanying drawings.
Referring to fig. 1 to 3, a method for enhancing similarity of small domains on manifold, the method comprising the steps of:
1) image random disturbance: disturbing an input image according to a certain disturbance intensity, and outputting a completely new disturbed image;
2) hidden variable random disturbance: inputting the encoded hidden variable, disturbing the hidden variable according to a certain disturbance intensity, and outputting the disturbed hidden variable;
3) the manifold structure: and mapping the image to a smooth manifold, wherein two images generated by the same image through the two disturbance technologies are nearly similar on the premise of the same disturbance intensity.
Further, in the step 3), before similarity enhancement is performed on the image in the manifold, preprocessing is performed on the image. The preprocessing, such as performing sparse binarization operation on the image, can improve the speed of performing similarity enhancement in the later period.
Further, in the step 2), random disturbance is performed on different dimensions of the image by using a random number generation method such as normal random distribution or uniform random distribution according to a certain interference intensity.
Furthermore, in the step 1), the image is coded and decoded by using a confrontation self-coder, and the hidden variables generated after coding are made to conform to a certain predefined prior probability distribution.
Preferably, in the step 3), a manifold construction technology based on a deep neural network is adopted, and a gradient descent method is used to update parameters of the network, so that the determiner can identify whether the input picture is from random disturbance of the original image or from random disturbance of a hidden variable.
In the image random disturbance step, when the disturbance intensity is 0, the output image is the same as the input image, and the output hidden variable is the same as the input hidden variable;
with the increase of the disturbance intensity of the image random disturbance technology, the difference between an output image and an input image is increased, and the difference between an output hidden variable and an input hidden variable is also increased.
The hidden variable disturbance processing process comprises the following modules: the image coding module is used for coding the image into a hidden variable; the hidden variable disturbance module is used for disturbing the coded hidden variable according to certain disturbance intensity; and the image decoding module is used for decoding the disturbed hidden variables into images.
In the image coding module, coding high-dimensional image information into low-dimensional hidden variable data; in the image decoder, low-dimensional hidden variable data is restored into an image.
In the hidden variable disturbance step, the output disturbance hidden variable cannot obviously change the statistical probability characteristic of the original input hidden variable; the large number of repeated perturbations may enable traversal of each solution in the neighborhood where the magnitude of the hidden variable is proportional to the perturbation strength.
The manifold construction process comprises: and the novelty decision device is used for inputting the images generated by the random disturbance of the images, decoding the hidden variables generated by the random disturbance of the hidden variables, and judging whether the disturbed intensities of the two images are similar by the novelty decision device. The novelty decider is able to update the parameters of the decider based on the difference between the output result and the correct result.
Fig. 2 is a schematic block diagram of an embodiment of the present invention.
And the image random disturbance module generates a disturbed image according to a disturbance intensity. The difference between the disturbance image and the original image increases with the intensity of the disturbance. An image randomizer MB (x, s) is defined, which gives a perturbed image o similar to x, given an original image x and a perturbation intensity s, the difference of which is roughly proportional to the perturbation intensity s.
In this embodiment, a Discrete Cosine Transform (DCT) is used as an implementation approach of the image randomizer, and the following method is adopted:
1) computing a frequency domain matrix for each image in an image database using a two-dimensional DCT transform, and a jth image xjHas a frequency domain matrix of Mj。
2) And calculating the maximum and minimum values of each element in each matrix in the database to form a maximum matrix MAX and a minimum matrix MIN.
3) Normalizing M for each matrixj＝(MjMIN)/(MAX-MIN), where subtraction and division are matrix element level operations.
4) Generating a sum M with a uniform distribution of-1 to 1jRandom matrices R of the same size.
5) To MjMaking random perturbations Ij＝Mj+ R s, where addition and multiplication are matrix element level operations.
6) To IjThe image o after random disturbance can be obtained by carrying out inverse normalization and inverse two-dimensional DCTj。
7) With ojThe absolute value of all pixel values in the image is a threshold value pair ojAnd carrying out binarization.
8) Using a decision device in an image encoder for an image ojMaking a decision if the decision maker considers ojIf the data does not belong to the original image data set, repeating the steps 4) to 8), otherwise, outputting the disturbed image oj。
In some embodiments, step 4 may generate a random matrix using a gaussian distribution or other random distribution.
As shown in fig. 2, the hidden variable random perturbation module comprises an autoencoder and a hidden variable random perturber. And finally, a decoding part of the self-encoder decodes the disturbed hidden variable into a new picture.
It should be noted that the disturbance hidden variable output by the hidden variable random disturbance module is required to obey the statistical probability characteristic of the original hidden variable, and the traversal of each solution in the neighborhood where the size of the hidden variable is proportional to the disturbance intensity can be realized by a large number of repeated disturbances.
On the premise of satisfying the above characteristic of the hidden variable, in this embodiment, the hidden variable random perturber is defined as ML (z, s), and the specific implementation method is as follows:
1) and uniformly and randomly distributing the random vector P with the size of the hidden variable z as the same as that of the hidden variable z by taking-1 as an upper limit and a lower limit.
2) Random perturbation v ═ z + P × s on hidden variable z, element-level operations where addition and multiplication are vectors
3) Outputting v as a result of the random perturbation of z
In some embodiments, the upper and lower limits of the generated random numbers may be defined as any other real number.
In some embodiments, the random numbers may be generated according to other probability distribution manners, such as a gaussian distribution or the like.
Referring to fig. 2, the manifold construction module is implemented based on a countermeasure generation network. And taking the image output by the image random disturbance module as a positive sample, taking the image output by the hidden variable disturbance module as a negative sample, inputting the negative sample into a novelty discriminator, and judging which module generates the image by the novelty discriminator.
In the present embodiment, the manifold construction method is as follows:
1) generating disturbance intensity s, and randomly selecting an image x from the training library
2) Generation of a perturbed image o using an image randomizer MB (x, s)
3) Generating a disturbance hidden variable v using a hidden variable randomizer ML (z, s), and generating a disturbance image t by an encoder G
4) The formula (3) is optimized to obtain the minimum value. Where D is the novelty decision, LD is the countermeasure error function of the novelty decision, and LR (G) is the predefined self-encoder reconstruction 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) Repeating steps 1) to 4) until equation (3) is optimal and the manifold is successfully constructed.
The embodiments of the present invention are described below with reference to specific embodiments, and other advantages and effects of the present invention will be easily understood by those skilled in the art from the disclosure of the present specification. The invention may be embodied or carried out in various other specific forms, and it is to be understood that various changes, modifications, and alterations may be made in the details of the description without departing from the spirit of the invention. It should be noted that the features in the following embodiments and examples may be combined with each other without conflict.
It should be noted that the drawings provided in the following embodiments are only for illustrating the basic idea of the present invention, and the drawings only show the components related to the present invention rather than the number of components in actual implementation, and the number and the ratio of the components in actual implementation can be changed freely.
Claims (1)
1. A method for enhancing similarity of small domains on manifold, comprising the steps of:
1) image random disturbance: disturbing an input image according to a certain disturbance intensity, and outputting a completely new disturbed image;
2) hidden variable random disturbance: inputting the encoded hidden variable, disturbing the hidden variable according to a certain disturbance intensity, and outputting the disturbed hidden variable;
3) the manifold structure: mapping the images to a smooth manifold, wherein two images generated by the same image through the two disturbance technologies are nearly similar on the premise of the same disturbance intensity;
in the step 3), before similarity enhancement is performed on the images in the manifold, preprocessing is performed on the images; in the step 2), random disturbance is performed on different dimensions of the image by using a normal random distribution or uniform random distribution random number generation method according to certain interference intensity; in the step 1), a confrontation self-encoder is used for encoding and decoding the image, and a hidden variable generated after encoding is made to accord with a certain predefined prior probability distribution; in the step 3), a manifold construction technology based on a deep neural network is adopted, and parameters of the network are updated by using a gradient descent method, so that the judger can identify whether the input picture is from random disturbance of the original image or from random disturbance of a hidden variable;
the method adopts Discrete Cosine Transform (DCT) as an implementation approach of an image random perturber, and comprises the following steps:
1.1) computing the frequency domain matrix for each image in the image database using a two-dimensional DCT transform, image x, jjHas a frequency domain matrix of Mj；
1.2) calculating the respective maximum and minimum values of each element in each matrix in a database to form a maximum matrix MAX and a minimum matrix MIN;
1.3) normalization M of each matrixj＝(Mj-MIN)/(MAX-MIN), where subtraction and division are matrix element level operations;
1.4) generating a sum M with a uniform distribution of-1 to 1jRandom matrices R of the same size;
1.5) to MjMaking random perturbations Ij＝Mj+ R × s, s is the perturbation intensity, where addition and multiplication are matrix element-level operations;
1.6) pairs of IjThe image o after random disturbance can be obtained by carrying out inverse normalization and inverse two-dimensional DCTj；
1.7) with ojThe absolute value of all pixel values in the image is a threshold value pair ojCarrying out binarization;
1.8) image o using a decision taker in the image encoderjMaking a decision if the decision maker considers ojIf the data does not belong to the original image data set, repeating the steps 1.4) to 1.8), otherwise, outputting the disturbed image oj。
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 CN108629134A (en) | 2018-10-09 |
| CN108629134B true 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) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| 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 |
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 |
|---|
| 一种基于局部保持的隐变量模型;王秀美等;《王秀美等》;20100630;第23卷(第3期);第369-375页 * |
| 一种新的多角度人脸表情识别方法;何俊等;《计算机应用研究》;20180131;第35卷(第1期);第282-286页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108629134A (en) | 2018-10-09 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111079532B (en) | Video content description method based on text self-encoder | |
| Kuzmin et al. | Fp8 quantization: The power of the exponent | |
| CN110309321B (en) | Knowledge representation learning method based on graph representation learning | |
| CN114662788B (en) | Seawater quality three-dimensional time-space sequence multi-parameter accurate prediction method and system | |
| Guo et al. | A fully-pipelined expectation-maximization engine for Gaussian mixture models | |
| Xu et al. | Singular vector sparse reconstruction for image compression | |
| Kekre et al. | Vector quantized codebook optimization using k-means | |
| CN108629134B (en) | Similarity strengthening method for small fields in manifold | |
| CN111401514A (en) | Semi-supervised symbol network embedding method and system based on improved graph convolutional network | |
| Huai et al. | Latency-constrained DNN architecture learning for edge systems using zerorized batch normalization | |
| Hou | Permuted sparse representation for 3D point clouds | |
| Ruberto et al. | Image feature learning with genetic programming | |
| CN116912296A (en) | Point cloud registration method based on position-enhanced attention mechanism | |
| CN113835964B (en) | Cloud data center server energy consumption prediction method based on small sample learning | |
| CN115310618A (en) | Quantum noise cancellation method and apparatus in quantum operation, electronic device, and medium | |
| Ren et al. | Hardware implementation of KLMS algorithm using FPGA | |
| CN114065913A (en) | Model quantization method and device and terminal equipment | |
| CN110349228B (en) | Triangular mesh compression method for data-driven least square prediction | |
| Han et al. | Online burr video denoising by learning sparsifying transform | |
| Chen et al. | Improved modeling and generalization capabilities of graph neural networks with legendre polynomials | |
| CN114222997A (en) | Method and apparatus for post-training quantization of neural networks | |
| CN109089125A (en) | A kind of DCT coefficient distribution characteristics model optimization method in Video coding | |
| Wang et al. | OctPCGC-Net: Learning Octree-Structured Context Entropy Model for Point Cloud Geometry Compression | |
| CN116704588B (en) | Face image replacing method, device, equipment and storage medium | |
| CN111522975B (en) | Equivalent continuously-changed binary discrete optimization non-linear Hash image retrieval method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |