CN109711442A - Unsupervised layer-by-layer generation confrontation feature representation learning method - Google Patents
Unsupervised layer-by-layer generation confrontation feature representation learning method Download PDFInfo
- Publication number
- CN109711442A CN109711442A CN201811536668.2A CN201811536668A CN109711442A CN 109711442 A CN109711442 A CN 109711442A CN 201811536668 A CN201811536668 A CN 201811536668A CN 109711442 A CN109711442 A CN 109711442A
- Authority
- CN
- China
- Prior art keywords
- network
- image
- generation
- optimization
- layer
- 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
- 238000000034 method Methods 0.000 title claims abstract description 48
- 238000009826 distribution Methods 0.000 claims abstract description 12
- 238000005457 optimization Methods 0.000 claims description 52
- 230000006870 function Effects 0.000 claims description 16
- 230000004069 differentiation Effects 0.000 claims description 15
- 238000004380 ashing Methods 0.000 claims description 11
- 238000004519 manufacturing process Methods 0.000 claims description 10
- 238000009825 accumulation Methods 0.000 claims description 5
- 238000013139 quantization Methods 0.000 claims description 3
- 230000035945 sensitivity Effects 0.000 claims description 3
- 238000001228 spectrum Methods 0.000 claims description 2
- 230000004913 activation Effects 0.000 description 14
- 238000013528 artificial neural network Methods 0.000 description 8
- 230000008569 process Effects 0.000 description 6
- 238000005070 sampling Methods 0.000 description 6
- 239000011248 coating agent Substances 0.000 description 5
- 238000000576 coating method Methods 0.000 description 5
- 239000000284 extract Substances 0.000 description 4
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 description 3
- 239000000203 mixture Substances 0.000 description 3
- 230000002776 aggregation Effects 0.000 description 2
- 238000004220 aggregation Methods 0.000 description 2
- 230000008859 change Effects 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 230000000087 stabilizing effect Effects 0.000 description 2
- HUTDUHSNJYTCAR-UHFFFAOYSA-N ancymidol Chemical compound C1=CC(OC)=CC=C1C(O)(C=1C=NC=NC=1)C1CC1 HUTDUHSNJYTCAR-UHFFFAOYSA-N 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 239000008199 coating composition Substances 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 238000003745 diagnosis Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 230000008034 disappearance Effects 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 230000003121 nonmonotonic effect Effects 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 238000010183 spectrum analysis Methods 0.000 description 1
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses an unsupervised layer-by-layer generation countermeasure feature representation learning method, which is characterized in that a stacking network is formed by 2 or more generation type countermeasure networks, wherein the first generation type countermeasure network takes multidimensional random noise as input, and the rest generation type countermeasure networks take the random noise and the implicit feature of the previous branch as input; generating images with corresponding sizes through each generation network, optimizing each discrimination network according to the cross entropy of each discrimination network, and optimizing the whole generation network according to expected values of intermediate layer characteristics of the discrimination network and similar statistical characteristics between the generated images; and extracting an abstract semantic feature representation vector by using the optimized SGANs, and determining the hash feature representation of the image by combining a hash representation method. The invention can generate high-level abstract semantic features and better learn the distribution of real images.
Description
Technical field
The present invention relates to machine learning techniques, and in particular to a kind of unsupervised layer-by-layer generation confrontation character representation study side
Method.
Background technique
Massive image retrieval based on content is widely used in e-commerce, medical diagnosis and trade mark and intellectual property neck
Domain usually carries out Similarity matching by the characteristics of image extracted, measures content similar image.The mode of manual extraction feature is by face
The character representation as image such as color, texture, shape, profile, improves the accuracy of image retrieval to a certain extent.But hand
The mode of work can only extract the color of image low level, Texture eigenvalue, cannot extract the high-level abstract semantics of image
Feature.Character representation study can be automatically extracted to classification, retrieval or prediction from image to task useful feature, such as deep
Degree neural network can extract image from the raw information of Pixel-level to abstract semantic concept information, and automatic study is abundant to image
Semantic feature.But exists in the image data of magnanimity largely without the data of label, have the deep learning method of supervision can only
Using the semantic feature for the image study image for having label on a small quantity, it is easy to produce over-fitting, generalization ability is poor.GAN is a kind of nothing
The deep learning model of supervision can utilize the distribution for differentiating the Game Learning true picture of network and generation network, but single
A network that generates is difficult to generate image multiplicity and with enough detailed information.
Summary of the invention
The purpose of the present invention is to provide a kind of unsupervised layer-by-layer generations to fight character representation learning method.
The technical solution for realizing the aim of the invention is as follows: a kind of unsupervised layer-by-layer generation confrontation character representation study side
Method includes the following steps:
Step 1, the building SGANs network architecture: pass through 2 or more production confrontation network (GAN) composition accumulation nets
Network (SGANs), wherein first production confrontation network is input with the random noise of multidimensional, remaining production fight network with
The hidden feature of random noise and previous branch is input；
Step 2 carries out the network optimization: the image of corresponding size is generated by each generation network, according to each differentiation network
Cross entropy, optimize each differentiation network, according to differentiate network middle layer feature desired value and generate image between it is similar
Statistical nature optimizes entire generation network；
Step 3 carries out character representation: using the SGANs of optimization, abstract semantics character representation vector is extracted, in conjunction with hash
Representation method determines the hash character representation of image.
As a preferred implementation manner, in step 2, each differentiation network is optimized respectively, optimization object function
Are as follows:
In formula,It indicates each optimization aim for differentiating network, passes through maximizationCarry out the differentiation network optimization, xiIt indicates
True picture,Indicate true picture distribution,It indicates to generate image,Indicate the model profile of generation image, E table
Show that expectation, D expression are the probability of true picture.
As a preferred implementation manner, in step 2, desired value matching optimization and structure one are carried out to each generation network
After the optimization of cause property, optimize whole generation network, method particularly includes:
Desired value matching optimization, optimization object function are carried out to each generation network are as follows:
In formula,Indicate each desired value matching optimization target for generating network, fiIt (x) is i-th of differentiation network middle layer
Activation primitive, E indicate expectation function, xiIndicate true picture,Indicate true picture distribution,It indicates to generate figure
Picture,Indicate the model profile of generation image；
Structural integrity optimization, optimization object function are carried out to each generation network are as follows:
In formula,Indicate each structural integrity optimization aim for generating network, λ1、λ2For weighting coefficient, μ=∑kxk/ N table
Show the mean value for generating the pixel of image, ∑=∑k(xk-μ)(xk-μ)T/ N indicates to generate the variance of the pixel of image, xk=
(R,G,B)TIndicate that the pixel of generation image, N are the pixel number for generating image；
The whole network that generates is optimized, optimization object function are as follows:
In formula, LGFor the optimization aim for entirely generating network, pass through maximization LGThe entire optimization for generating network is carried out,Indicate the optimization aim of i-th of generation network, n indicates the number of production confrontation network.
As a kind of more preferable embodiment, λ1、λ2Respectively 1 and 4.
As a preferred implementation manner, in step 3, the hash representation method of use include local sensitivity ashing technique,
Compose ashing technique, iterative quantization ashing technique and core ashing technique.
Compared with prior art, the present invention its remarkable advantage are as follows: 1) feature of the invention by complicated image hierarchy structure
It indicates that learning tasks are divided into multiple subtasks, gradually learns image low level color, Texture eigenvalue to high-level abstract language
Adopted feature；2) present invention generates the similar statistics feature of image according to different generation networks, and structure one is added in the network optimization
Cause property optimization aim, the training of stabilizing network preferably learn the distribution of true picture.
Detailed description of the invention
Fig. 1 is the schematic network structure of single GAN.
Fig. 2 is the network structure of the unsupervised SGANs of the present invention.
Fig. 3 is residual error network of the present invention.
Fig. 4 is present invention confrontation character representation flow chart.
Specific embodiment
The present invention program is further illustrated in the following with reference to the drawings and specific embodiments.
The present invention devises accumulation network SGANs, and the character representation learning tasks of complicated image hierarchy structure are divided into
Multiple subtasks gradually learn image low level color, Texture eigenvalue to high-level abstract semantics feature, and by its feature
Applied to image retrieval.The characteristics of similar statistical nature should be had by generating image according to different generation networks, in the network optimization
Structural integrity optimization aim is added, the training of stabilizing network preferably learns the distribution of true picture.Net is described in detail below
Structure, optimization aim and the details of realization of network.
One, the SGANs network architecture
GAN is used to capture the distribution of real image data.The single structure for generating network is as shown in Figure 1.Each GAN is by one
A generation network G and a differentiation network D composition.Generating network G is a deconvolution neural network, and input is a multidimensional
Random noise vector z ∈ N (0,1) generates the image x=G (z) with true image same size by up-sampling.Differentiate net
Network D is a convolutional neural networks, for differentiating that image is true picture or generates image, is equivalent to a two classification
Device:
P (x=real)=D (x)
SGANs is made of multiple GAN.Fig. 2 gives the structure of unsupervised SGANs used herein.It is shown in Fig. 2
The accumulation network that is made of 3 GAN.True image can be re-adjusted to different sizes, and different generation networks is raw
At the image of corresponding size.Wherein neural network FiImplicit feature h is obtained by up-samplingi, GiWith coming by hiGenerate image.
Differentiate network DiDifferentiate that image corresponding to i-th of GAN is the probability of true picture.Network 0 is generated with the random noise z of multidimensional
∈ (0,1) passes through neural network F as input0Up-sampling obtains hidden feature:
h0=F0(z)
Hidden feature h0Pass through neural network G0Generate image:
For generating network i, hidden feature:
hi=Fi(hi-1,z)
Pass through hidden feature hiGenerate image:
Two, the optimization aim of network
2.1 differentiate the optimization aim of network
In GAN, network is differentiated for differentiating the image that image is still generated from true picture, maximum possible is sentenced
Not Chu image source.In SGANs, true picture is sized as x0,x1,...,xn, n is of GAN in whole network
Number.Differentiate that the optimization object function of network is similar with the objective function of two classification problems, for i-th of generation network, differentiates
The cross entropy optimization aim of network are as follows:
In training, every image xiFrom true image distributionEvery generation imageFrom mould
Type distributionEach differentiation network, which only focuses on, is currently located the corresponding image of network, and each differentiation network respectively optimizes.
2.2 generate the optimization aim of network
Being originally generated in formula confrontation network and generating network G for cheating differentiation network to generate image discriminating is really to scheme
Picture maximizes the probability for generating that image is true picture.In order to avoid the problem of gradient disappears in the training process, using most
Smallization 1-D (xsyn) come replace maximize D (xsyn), it is originally generated the optimization aim of formula confrontation network are as follows:
In order to stably generate the training process of formula confrontation network, several optimization methods for generating network have been suggested.It is special
Sign matching (Feature Matching) method generates network over training on differentiating network by preventing, and stably generates formula net
The training process of network.Characteristic matching requires to generate the statistical nature that image matches true picture as far as possible, differentiates that network is used to
It is specified to need matched statistical nature.In the implementation, training generates the desired value that net mate differentiates network middle layer feature.It is false
If f (x) is the activation primitive for differentiating network middle layer, then the optimization aim of network is generated are as follows:
In SGANs, network i is fought for production, differentiates that the activation primitive of network middle layer is fi, then optimize mesh
It is designated as:
Different GAN in SGANs will generate different size of image, and some similar systems should be possessed between these images
Count feature.The mean value and variance of the pixel in each channel between image are generated by minimizing, and guarantee to generate what network generated
Have structured consistency between image, the optimization aim of structural integrity is added, promotes the quality for generating image.
Assuming that xk=(R, G, B)TIt indicates to generate a pixel in image, generates the mean value of the pixel of image:
μ=∑kxk/N
Generate the variance of the pixel of image are as follows:
∑=∑k(xk-μ)(xk-μ)T/N
Wherein N is the number of pixel in image.
The structural integrity optimization aim of i-th (i > 0) a GAN are as follows:
λ herein1And λ2Value be respectively 1 and 4.Generate network 0 optimization aim beSo for i-th
A GAN, optimization aim become:
Multiple scale true pictures can be obtained by the same continuous picture signal with different sample rates.SGANs's
It generates network and multiple and different but relevant image distribution is approached by joint training, so entirely accumulating the excellent of the generation network of network
Change target are as follows:
Wherein, n is the number of GAN in entire accumulation network.
2.3 network training processes
Algorithm 1 is the training process of SGANs.The image for generating network and generating corresponding size each first, it is then successively excellent
Change each differentiation network, finally optimizes entire generation network using the optimization aim of obtained all generation networks.
Algorithm 1SGANs training process
Three, the realization of SGANs
Herein by taking 3 GAN form unsupervised SGANs as an example, it is described in detail below in each GAN and generates network and differentiation net
The specific implementation details of network, the structure of the number of plies and every layer network including network.Differentiate that neural network shares 3, respectively
D_Net32, D_Net64 and D_Net128, the size of corresponding input picture are respectively 32x32,64x64 and 128x128.With mould
As introducing in type, differentiate that network is mainly used for differentiating the true and false of image.Due to the input of first production confrontation network
It is different from the generation input of network later, realizing that generating network Time Division is Init_G and Next_G.Init_G is used for first
A generation network obtains hidden feature h with 100 dimensions random noise z ∈ N (0,1) as input0.Subsequent generation network all by
Next_G is realized, using the hidden feature of previous branch and random noise z as input, respectively obtains implicit feature h1And h2。
Image_Net is used to generate image by hidden feature.Image_Net(h0) generate 32x32 size image, Image_Net (h1)
Generate the image of 64x64 size, Image_Net (h2) generate 128x128 size image.
3.1 generate network implementations
Table 1 is upblock module, refers mainly to up-sample input.Wherein, UpSample will input as (N, C, H, W)
Tensor up-sampling is (N, C, H*scale_factor, W*scale_factor).N refers to the size of batch size, and C is defeated
The port number entered, H are the picture altitudes of input, and W is the width of input picture.
Conv refers to convolutional layer.In_channels is input feature vector figure (featrue map) quantity, out_
Channels refers to the quantity of the characteristic pattern of output.The third parameter of Conv is the size of convolution kernel (kernel), the 4th
A parameter refers to the size of step-length (stride), the last one parameter is to fill the size of (padding).
BN (Batch Normalization) solves the problems, such as that the gradient in backpropagation disappears and gradient is exploded.It is public
Formula are as follows:
Wherein E [x] and Var [x] is respectively the expectation and variance of corresponding dimension.γ and β is the parameter that can learn.
GLU (Gated Linear Units) is an activation primitive, and nonlinear factor is added for network, also can be one
GLU unlike determining the problem of alleviating gradient disappearance in degree and Relu is a continuous function, nonmonotonic.Assuming that Y=
[A,B]∈R2d, then
Wherein A, B ∈ Rd.The size of the tensor shape of the output of GLU ([AB]) is the half of Y.
Table 1 up-samples module upblock
Fig. 3 is a residual error network, it is assumed that H (x)-x → F (x) in deep neural network, if it can be assumed that multiple non-thread
Property layer combination can be similar to a complicated function, then equally assume that the residual error of hidden layer is similar to some complicated letter
Number, hidden layer can be indicated are as follows: H (x)=F (x)+x.Residual error network, which solves the neural network number of plies, excessively causes gradient to disappear
The problem of, while also improving the performance of network.Table 2 is residual block used herein.
2 residual block resblock of table
| Operation | Specific implementation |
| conv | Conv(channel_num,channel_num*2,3,1,1) |
| BN | Batch_norm2d(channel_num*2) |
| activation | GLU() |
| conv | Conv(channel_num,channel_num,3,1,1) |
| BN | Batch_norm2d(channel_num) |
| add | add(x) |
Table 3 is the structure of Init_G network.Init_G network is using random noise vector z as input.It is linear by one
Layer, BN layers, active coating and 3 up-sampling parts form.It is that network addition is non-linear that the output of linear layer is activated with GLU.
Table 3Init_G network structure
| Operation | Specific implementation |
| linear | Linear(in_dim,img_size*32) |
| BN | BatchNorm1d(img_size*32) |
| activation | GLU() |
| upblock | upblock(img_size,img_size) |
| upblock | upblock(img_size,img_size/2) |
| upblock | upblock(img_size/2,img_size/4) |
Table 4 is the network structure of Next_G.Next_G is by a convolutional layer, a residual block and a up-sampling block group
At.Random vector z is replicated first, then with hidden feature hi-1Connection obtains tensor t conduct in the dimension of axis=1
The input of Next_G.Netxt_G network is primarily used to by hidden feature hi-1Obtain hidden feature hi, standard is done to generate image
It is standby.
Table 4Next_G network structure
| Operation | Specific implementation |
| conv | Conv(t,img_size,3,1,1) |
| BN | Batch_norm2d(img_size*2) |
| activation | GLU() |
| resblock | resblock(img_size) |
| upblock | upblock(img_size,img_size/2) |
Table 5 is Image_Net network, which is made of a convolutional layer.Tanh activation primitive is added after convolutional layer.
Image_Net is with hidden feature hiAs input, the image of corresponding size is generated.
Table 5Image_Net network structure
| Operation | Specific implementation |
| conv | Conv(in_channels,out_channels,3,1,1) |
| activation | tanh() |
3.2 differentiate the realization of network
Table 6 is encode_image module.The module is using image data as input, by multiple convolutional layers, activation and BN
Composition.
Table 6encode_image module
Table 7 is the network structure of D_NET32.D_NET32 is by an encode_image module, linear layer and active coating group
At.
Table 7D_NET32 network structure
| Operation | Specific implementation |
| encode_image | encode_image(img_size) |
| linear | Linear(img_sizeximage_size,1) |
| activation | Sigmoid (): judge whether image is true picture |
Table 8 is block_leakyRelu module, is made of a convolutional layer, BN and active coating.
Table 8block_leakyRelu module
| Operation | Specific implementation |
| conv | Conv(in_channels,out_channels,3,1,1) |
| BN | BatchNorm2d(out_channels) |
| activation | LeakyRelu(0.2) |
Table 9 is down_block module, similar with block_leakyRelu to be made of convolutional layer, BN and active coating.
Down_block makes the bulk of tensor reduce half.
Table 9down_block module
| Operation | Specific implementation |
| conv | Conv(in_channels,out_channels,4,2,1) |
| BN | BatchNorm2d(out_channels) |
| activation | LeakyRelu(0.2) |
Table 10 is the network structure of D_NET64.D_NET64 is by an encode_imge module and multiple BN, active coating group
At.
Table 10D_NET64 network structure
Table 11 is the network structure of D_NET128.D_NET128 is by an encode_imge module, 2 own_block moulds
Block, 2 block_leakyRelu modules, a linear layer and active coating composition.
Table 11D_NET128 network structure
| Operation | Specific implementation |
| encode_image | encode_image(img_size) |
| down_block | down_block(img_sizex8,img_sizex16) |
| down_block | down_block(img_sizex16,img_sizex32) |
| block_leakyRelu | block_leakyRelu(img_sizex32,img_sizex16) |
| block_leakyRelu | block_leakyRelu(img_sizex16,img_sizex8) |
| linear | Linear(img_sizeximage_size,hash_dim) |
| activation | Sigmoid (): judge whether image is true picture |
Four, hierarchical structure character representation and image retrieval
The high dimensional feature expression of image is difficult to set up effective index and promotes effectiveness of retrieval.In face of mass image data,
The distance between image measurement will also devote a tremendous amount of time, and the retrieval of accurate neighbour's image is unpractical.By close
It realizes that similarity retrieval is able to ascend the time efficiency of retrieval like arest neighbors method, meets the needs of most of applications.
Ashing technique is widely used in large-scale information retrieval task as one of approximate KNN method
In.The image feature representation of higher-dimension is mapped to the Hamming space of low-dimensional by ashing technique, use is shorter to be made of 0 and 1
Hash character representation of the binary-coding as image, so that the hash character representation of image has good aggregation properties, retrieval
Time efficiency can all be significantly improved.Assuming that indicating image with the binary-coding that length is L, then it is equivalent to image
Feature space is divided into 2LSub-spaces.The image of identical hash character representation will drop into same space, Hamming distance
From the image for 1 then in adjacent subspace.Assuming that image has 1000 classes, if the hash feature of image can be good at table
Aggregation properties between diagram picture then only need the hash character representation that length is 10 bits can be by different classes of image
It accurately assigns in different subspaces.
After the hash character representation for obtaining image, there are two types of the retrievals that basic method realizes similar image, improve
The accuracy and time efficiency of retrieval.One is hash table is established, different from traditional ashing technique, hash table is wanted in retrieval
The collision rate for increasing the hash character representation of similar image as far as possible, so that similar image drops into same bucket.Separately
A kind of outer way is the direct query image that calculates at a distance from image in image data base.Since hash character representation uses the Chinese
Prescribed distance calculates the distance between image, and calculating speed quickly, and hashes character representation and usually uses shorter bit, can be with
It is loaded into memory to accelerate the speed of retrieval.In addition it can use the number that inverted index reduces the candidate image for needing to compare
Amount, further speeds up the speed of retrieval.
Such as Fig. 4, extracting the high-level abstract characteristics of image by SGANs indicates, then using hashing technique by extracting
The character representation of higher-dimension obtains the low degree of image and hashes character representation.The preceding layer of the linear layer used herein for differentiating network
Abstract semantic feature of the output as image.LSH, SH, ITQ and KSH are that high dimensional feature is commonly mapped to low-dimensional
Hamming space, obtain data hash indicate method.It is most general using gaussian random projection matrix that local sensitivity hashes LSH
Rate is by similar image hash into same bucket.Image data will be carried over into as far as possible in the local feature of luv space
Hamming space.Spectrum hash SH carries out spectrum analysis to high dimensional data, converts Laplce spy for problem by relaxed constraints condition
The dimensionality reduction problem for levying figure, to obtain the hash character representation of image data.It is a kind of based on PCA that iterative quantization, which hashes ITQ,
Image hash algorithm makes the variance of all directions after rotating keep balancing as far as possible by rotating principal component.Core hash KSH passes through core
Method study obtains the data of hash function processing linearly inseparable.Above-mentioned ashing technique is applied to image retrieval
In, but the character representation of previous image is all that manual mode extracts.The color for being characterized in low level, texture are extracted by hand
Etc. information, cannot reflect that the semantic information of image abstraction, the accuracy rate of retrieval are low.
It is dissipated herein by what the characteristics of image based on manual extraction and the image abstraction semantic feature extracted based on SGANs were indicated
Column method compares, and studies effect of the character representation of the image based on SGANs in image retrieval.
Claims (5)
1. unsupervised layer-by-layer generation fights character representation learning method, which comprises the steps of:
Step 1, the building SGANs network architecture: fighting group of networks into accumulation network by 2 or more productions, wherein the
One production confrontation network is input with the random noise of multidimensional, remaining production fights network with random noise and previous
The hidden feature of branch is input；
Step 2 carries out the network optimization: the image of corresponding size is generated by each generation network, according to each friendship for differentiating network
Entropy is pitched, each differentiation network is optimized, according to the desired value for differentiating network middle layer feature and generates the similar statistics between image
Feature optimizes entire generation network；
Step 3 carries out character representation: using the SGANs of optimization, extracting abstract semantics character representation vector, indicates in conjunction with hash
Method determines the hash character representation of image.
2. unsupervised layer-by-layer generation according to claim 1 fights character representation learning method, which is characterized in that step 2
In, each differentiation network is optimized respectively, optimization object function are as follows:
In formula,It indicates each optimization aim for differentiating network, passes through maximizationCarry out the differentiation network optimization, xiIndicate true figure
Picture,Indicate true picture distribution,It indicates to generate image,Indicating the model profile of generation image, E indicates expectation,
D expression is the probability of true picture.
3. unsupervised layer-by-layer generation according to claim 1 fights character representation learning method, which is characterized in that step 2
In, after carrying out desired value matching optimization and structural integrity optimization to each generation network, optimize whole generation network, specific side
Method are as follows:
Desired value matching optimization, optimization object function are carried out to each generation network are as follows:
In formula,Indicate each desired value matching optimization target for generating network, fiIt (x) is swashing for i-th of differentiation network middle layer
Function living, E indicate expectation function, xiIndicate true picture,Indicate true picture distribution,It indicates to generate image,
Indicate the model profile of generation image；
Structural integrity optimization, optimization object function are carried out to each generation network are as follows:
In formula,Indicate each structural integrity optimization aim for generating network, λ1、λ2For weighting coefficient, μ=∑kxk/ N indicates life
At the mean value of the pixel of image, ∑=∑k(xk-μ)(xk-μ)T/ N indicates to generate the variance of the pixel of image, xk=(R, G,
B)TIndicate that the pixel of generation image, N are the pixel number for generating image；
The whole network that generates is optimized, optimization object function are as follows:
In formula, LGFor the optimization aim for entirely generating network, pass through maximization LGThe entire optimization for generating network is carried out,Indicate the optimization aim of i-th of generation network, n indicates the number of production confrontation network.
4. unsupervised layer-by-layer generation according to claim 3 fights character representation learning method, which is characterized in that λ1、λ2Point
It Wei 1 and 4.
5. unsupervised layer-by-layer generation according to claim 1 fights character representation learning method, which is characterized in that step 3
In, the hash representation method of use includes local sensitivity ashing technique, spectrum ashing technique, iterative quantization ashing technique and core hash
Method.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811536668.2A CN109711442B (en) | 2018-12-15 | 2018-12-15 | Unsupervised layer-by-layer generation confrontation feature representation learning method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811536668.2A CN109711442B (en) | 2018-12-15 | 2018-12-15 | Unsupervised layer-by-layer generation confrontation feature representation learning method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109711442A true CN109711442A (en) | 2019-05-03 |
| CN109711442B CN109711442B (en) | 2021-04-16 |
Family
ID=66256561
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811536668.2A Active CN109711442B (en) | 2018-12-15 | 2018-12-15 | Unsupervised layer-by-layer generation confrontation feature representation learning method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109711442B (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110287357A (en) * | 2019-05-31 | 2019-09-27 | 浙江工业大学 | A kind of iamge description generation method generating confrontation network based on condition |
| CN110414593A (en) * | 2019-07-24 | 2019-11-05 | 北京市商汤科技开发有限公司 | Image processing method and device, processor, electronic equipment and storage medium |
| CN112967379A (en) * | 2021-03-03 | 2021-06-15 | 西北工业大学深圳研究院 | Three-dimensional medical image reconstruction method for generating confrontation network based on perception consistency |
| CN111311702B (en) * | 2020-01-15 | 2023-04-28 | 浙江传媒学院 | Image generation and identification module and method based on BlockGAN |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105808723A (en) * | 2016-03-07 | 2016-07-27 | 南京邮电大学 | Image retrieval method based on image semantics and visual hashing |
| CN107220600A (en) * | 2017-05-17 | 2017-09-29 | 清华大学深圳研究生院 | A kind of Picture Generation Method and generation confrontation network based on deep learning |
| US20180201164A1 (en) * | 2017-01-16 | 2018-07-19 | Textron Inc. | Auxiliary seating system for light weight utility vehicle |
| CN108665005A (en) * | 2018-05-16 | 2018-10-16 | 南京信息工程大学 | A method of it is improved based on CNN image recognition performances using DCGAN |
| CN108694721A (en) * | 2017-04-04 | 2018-10-23 | 通用电气公司 | Light stream determines system |
-
2018
- 2018-12-15 CN CN201811536668.2A patent/CN109711442B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105808723A (en) * | 2016-03-07 | 2016-07-27 | 南京邮电大学 | Image retrieval method based on image semantics and visual hashing |
| US20180201164A1 (en) * | 2017-01-16 | 2018-07-19 | Textron Inc. | Auxiliary seating system for light weight utility vehicle |
| CN108694721A (en) * | 2017-04-04 | 2018-10-23 | 通用电气公司 | Light stream determines system |
| CN107220600A (en) * | 2017-05-17 | 2017-09-29 | 清华大学深圳研究生院 | A kind of Picture Generation Method and generation confrontation network based on deep learning |
| CN108665005A (en) * | 2018-05-16 | 2018-10-16 | 南京信息工程大学 | A method of it is improved based on CNN image recognition performances using DCGAN |
Non-Patent Citations (1)
| Title |
|---|
| ZHAOFAN QIU等: "《Deep Semantic Hashing with Generative Adversarial Networks》", 《ARXIV》 * |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110287357A (en) * | 2019-05-31 | 2019-09-27 | 浙江工业大学 | A kind of iamge description generation method generating confrontation network based on condition |
| CN110287357B (en) * | 2019-05-31 | 2021-05-18 | 浙江工业大学 | Image description generation method for generating countermeasure network based on condition |
| CN110414593A (en) * | 2019-07-24 | 2019-11-05 | 北京市商汤科技开发有限公司 | Image processing method and device, processor, electronic equipment and storage medium |
| CN111311702B (en) * | 2020-01-15 | 2023-04-28 | 浙江传媒学院 | Image generation and identification module and method based on BlockGAN |
| CN112967379A (en) * | 2021-03-03 | 2021-06-15 | 西北工业大学深圳研究院 | Three-dimensional medical image reconstruction method for generating confrontation network based on perception consistency |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109711442B (en) | 2021-04-16 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109670528B (en) | Data expansion method facing pedestrian re-identification task and based on paired sample random occlusion strategy | |
| CN110348399B (en) | Hyperspectral intelligent classification method based on prototype learning mechanism and multidimensional residual error network | |
| CN111382300B (en) | Multi-view three-dimensional model retrieval method and system based on pairing depth feature learning | |
| CN113469236B (en) | Self-tag learning deep clustering image recognition system and method | |
| CN109711442A (en) | Unsupervised layer-by-layer generation confrontation feature representation learning method | |
| CN108108751B (en) | Scene recognition method based on convolution multi-feature and deep random forest | |
| WO2019179403A1 (en) | Fraud transaction detection method based on sequence width depth learning | |
| Sinha et al. | Class-wise difficulty-balanced loss for solving class-imbalance | |
| CN110321805B (en) | Dynamic expression recognition method based on time sequence relation reasoning | |
| Mia et al. | Computer vision based local fruit recognition | |
| CN106778768A (en) | Image scene classification method based on multi-feature fusion | |
| CN107358172B (en) | Human face feature point initialization method based on human face orientation classification | |
| CN110211127A (en) | Image partition method based on bicoherence network | |
| CN114821022A (en) | Credible target detection method integrating subjective logic and uncertainty distribution modeling | |
| CN112818774A (en) | Living body detection method and device | |
| CN114092819B (en) | Image classification method and device | |
| Cai et al. | Rgb-d scene classification via multi-modal feature learning | |
| CN109948662B (en) | Face image depth clustering method based on K-means and MMD | |
| CN107766792A (en) | A kind of remote sensing images ship seakeeping method | |
| CN113850182B (en) | DAMR _ DNet-based action recognition method | |
| Song et al. | Srrm: Semantic region relation model for indoor scene recognition | |
| CN117853810A (en) | Hyperspectral image classification method for generating countermeasure network based on semi-supervised residual error | |
| CN117765258A (en) | Large-scale point cloud semantic segmentation method based on density self-adaption and attention mechanism | |
| CN117422876A (en) | Image instance segmentation method based on enhanced disposable aggregation network | |
| CN112364193A (en) | Image retrieval-oriented method for fusing multilayer characteristic deep neural network model |
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 |