CN108171320A - A kind of image area switching network and conversion method based on production confrontation network - Google Patents
A kind of image area switching network and conversion method based on production confrontation network Download PDFInfo
- Publication number
- CN108171320A CN108171320A CN201711273921.5A CN201711273921A CN108171320A CN 108171320 A CN108171320 A CN 108171320A CN 201711273921 A CN201711273921 A CN 201711273921A CN 108171320 A CN108171320 A CN 108171320A
- Authority
- CN
- China
- Prior art keywords
- network
- image
- input
- true
- false
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
Abstract
The invention discloses a kind of image area switching networks and conversion method based on production confrontation network, are mainly included the following steps that including U-shaped generation network, true and false discrimination natwork and pairing discrimination natwork, image area transfer process:1), the U-shaped generation network of training, establish the network model of U-shaped generation network;2)Image to be converted is inputted after normalized through step 1)The network model of foundation completes the image area conversion of image to be converted;The present invention can realize the image area convert task of regional area in image, and image local domain conversion quality is high, and network judgement is strong, and the stability of image conversion is strong, substantially increases the authenticity of the image of generation.
Description
Technical field
The present invention relates to image area switch technology fields more particularly to a kind of image area based on production confrontation network to turn
Switching network and conversion method.
Background technology
Image area conversion is an important research direction in computer vision, is had broad application prospects.At present, it is right
The appearance of anti-formula generation network (Generative Adversarial Networks, GAN) achieves in image generation field
The achievement to attract people's attention, this also provides new solution for image area conversion.Network is fought using production, image is defeated
Enter and generate the image of network generation aiming field, the training of network is completed based on the game between generation network and discrimination natwork.
Production confrontation network is unsupervised learning method when initially proposing, by generating the game between network and discrimination natwork,
Gradually learn the data distribution into training set, allow generation network according to learning so as to fulfill one random value of input
Data distribution goes to be randomly generated data, the generation of earliest application, that is, image.Later, Conditional GAN are by giving
The input of GAN increases artificial condition, its data generated is made no longer to be randomly generated, but is given birth to according to the condition difference of input
Into different data, the proposition of Conditional GAN so that the conversion of image area can use production fight network into
Row.
Later, the Conditional GAN appearance from original GAN is improved, the Conditional GAN studied at present lead to
Cross increases artificial condition to the input of GAN, can be to specifically inputting the specific image data of generation, and is no longer to input immediately
Data generation image data immediately.The it is proposed of Conditional GAN allows the conversion of image area to use production
It fights network to carry out, i.e., input is original area image under its frame, and target area image can be exported by training study.
The image area conversion GAN realized under the frame has:(1) pix2pix GAN, the method for using supervision, network base
In one generate network and one confrontation discrimination natwork, solution be image whole domain convert task;(2) Cycle GAN make
With unsupervised method, by two generation networks, two confrontation networks, recycled using cycle-consistency loss
Training generation network and fight discrimination natwork realize image area conversion.Wherein, unsupervised method is although no longer need one
The training data of one pairing, but conversion effect is worse than the pix2pix networks of supervision, and what is be directed to is still image whole domain
Conversion, in the conversion of existing image area, for the domain convert task of regional area in image, there is no special GAN.
Invention content
The object of the present invention is to provide a kind of image area switching network and conversion method based on production confrontation network, energy
Enough realize the image area convert task of regional area in image, and image local domain conversion quality is high, network judgement is strong, figure
The stability of picture conversion is strong, substantially increases the authenticity of the image of generation.
The technical solution adopted by the present invention is:
It is a kind of based on production confrontation network image area switching network, including it is U-shaped generation network, true and false discrimination natwork and
Discrimination natwork is matched, U-shaped generation network includes coding network and decoding network, the input terminal access input picture of coding network
Input, the input terminal of the output terminal connection decoding network of coding network, the output terminal output network generation image of decoding network
Output;If it is target area image target with the true target area image that input picture Input is matched one by one;Network generates
Image Output inputs the negative sample input terminal of true and false discrimination natwork, aiming field figure as the training negative sample of true and false discrimination natwork
As positive sample input terminals of the target as the true and false discrimination natwork of training positive sample input of true and false discrimination natwork, true and false discriminating net
The value of network output feeds back to the true and false loss input terminal of decoding network as true and false penalty values;Network generates image Output and right
Answer negative sample input terminals of the input picture Input as training negative sample input pairing discrimination natwork, target area image target
With positive sample input terminals of the corresponding input picture Input as training positive sample input pairing discrimination natwork, discrimination natwork is matched
Input terminal is lost in the pairing that the value of output feeds back to decoding network as pairing penalty values;Network is generated into image Output and mesh
Structural similarity value feeds back to the compensation loss input terminal of decoding network as compensation penalty values between mark area image target.
The coding network includes eight layers of convolutional network, and the convolution kernel size of every layer of convolutional network is 3*3, step-length 2*2,
Every layer of convolutional network includes convolutional layer, Normalization layers of Batch and Leak ReLU active coatings, Leak ReLU activation
The alpha parameter of layer is 0.2;The decoding network includes eight layers of deconvolution network, and the deconvolution core size of every layer of deconvolution network is
3*3, step-length 2*2, every layer of deconvolution network include warp lamination, Normalization layers of Batch and active coating, first to
The active coating of layer 7 deconvolution network uses ReLU active coatings, and the active coating of the 8th layer of deconvolution network is activated using tanh
Layer.
The true and false discrimination natwork includes the true and false discriminating convolutional network that multilayer is transmitted successively, every layer of true and false discriminating convolution net
Network includes convolutional layer, Normalization layers of Batch and active coating, last layer of true and false active coating for differentiating convolutional network
Using Sigmoid activation primitives, other layer of true and false active coating for differentiating convolutional network uses ReLU functions.
The pairing discrimination natwork includes the Concat layers transmitted successively and multilayer pairing differentiates convolutional network, every layer of pairing
Differentiate that convolutional network includes convolutional layer, Normalization layers of Batch and active coating, the pairing of last layer differentiates convolution net
The active coating of network uses Sigmoid activation primitives, and other layers of pairing differentiate that the active coating of convolutional network uses ReLU functions.
A kind of image area conversion method based on production confrontation network, includes the following steps:
1), the U-shaped generation network of training establishes the network model of U-shaped generation network;Specifically include following steps:
A, the training image collection in domain to be converted is collected, training image concentrates the original area image for including matching one by one and target
Area image, the original area image concentrated to training image is normalized, defeated when the image after normalization is network training
Enter image Input, it is target area image target corresponding with input picture Input that training image, which concentrates aiming field image,;
B, input picture Input obtained by step A is converted into the network generation figure of training network through U-shaped generation network
As Output;
C, image is generated using input picture Input, target area image target and network obtained by step A and step B
Output carries out the training of multipair anti-discrimination natwork:The training of multipair anti-discrimination natwork includes the training of true and false discrimination natwork and matches
Training to discrimination natwork, wherein, the training of true and false discrimination natwork includes the following steps:
C11:The network weight of true and false discrimination natwork is initialized using random initializtion method;
C12:Network is generated into image Output as negative sample, target area image corresponding with input picture Input
Target is trained in true and false discrimination natwork as positive sample, using intersecting Mutual information entropy loss function and adam is excellent
Change the network weight that algorithm updates true and false discrimination natwork;
The training of pairing discrimination natwork includes the following steps:
C21:The network weight for matching discrimination natwork is initialized using random initializtion method:
C22:Network is generated into image Output and corresponding input picture Input as negative sample, by input picture
Input and corresponding target area image target is trained in discrimination natwork is matched as positive sample, uses intersection mutual trust
Cease the network weight of entropy loss function and adam optimization algorithms update pairing discrimination natwork;
D, step C is repeated, after anti-discrimination natwork training multipair twice, fixed true and false discrimination natwork and pairing differentiate net
The network weight of network;
E, using the multipair anti-discrimination natwork of the gained after step D training, U-shaped generation network is trained;Specifically
Include the following steps:
E1:The network weight of U-shaped generation network is initialized using Ha Weier random initializtions method;
E2:Network generation image Output is inputted into true and false discrimination natwork, true and false discrimination natwork exports true and false penalty values, and
The true and false penalty values of output are fed back into the decoding network in U-shaped generation network, for updating its network weight:True and false discriminating net
Network exports 30*30*1 images, and image Output is generated close to the penalty values of true picture for returning to network, wherein, true and false mirror
Ranging from 0 to the 1 of each image pixel point value of other network output, pixel point value represents input picture Input at this closer to 1
Pixel receptive field region represents input picture Input in the pixel sense closer to true image, pixel point value closer to 0
True image is more kept off by wild region;
E3:By input picture Input and corresponding network generation image Output input pairing discrimination natworks, pairing differentiates
Network output pairing penalty values, and the pairing penalty values of output are fed back into the decoding network in U-shaped generation network, for updating
Its network weight:Discrimination natwork output 30*30*1 images are matched, for returning to input picture Input and network generation image
The penalty values whether Output is input picture Input with target area image target pairings;Wherein, pairing discrimination natwork output
Each image pixel point value ranging from 0 to 1, closer to 1 expression input picture Input and network generation image Output get over
Matching represents more to mismatch closer to 0;
E4:The structure similar values between network generation image Output and target area image target are calculated, and will
The structural similarity numerical value calculated feeds back to the decoding network in U-shaped generation network as loss, for updating its network weight
Weight;Structural similarity numerical value includes SSIM loss functions result of calculation and L1 regularization result of calculations, wherein, SSIM loss letters
For number from SSIM algorithms, the output valve SSIM (x, y) of SSIM algorithms represents the similitude between two images, i.e. input picture
Structural similarity between x and target area image y, ranging from-the 1 to 1 of SSIM (x, y), represent the similar of two images close to 1
Degree is higher, and as input picture x and target area image y striking resemblances, the value of SSIM (x, y) is equal to 1;
The calculation formula of SSIM algorithm output valves is as follows:
In formula (1), x is input picture Input, and y is target area image target corresponding with input picture Input,
μxIt is the average value of x, μyIt is the average value of y,It is the variance of x,It is the variance of y, σxyIt is the covariance of x and y, c1=
(k1L)2And c2=(k2L)2It is for maintaining stable constant, L is the dynamic range of pixel value, k1=0.01,2=0.03;
F, step C~E is the weight training of primary U-shaped generation network, repeats step C~E, completes U-shaped generation net twice
After the weight training of network, that is, the training of U-shaped generation network is completed, gained generation network is the network model of U-shaped generation network;
2), image to be converted is inputted to the network model established through step 1) after normalized, you can complete figure to be converted
The image area conversion of picture:Image after normalization is as input picture Input input steps 1) in establish network model in, compile
Then the high dimensional feature of code network extraction input picture Input exports network by decoding network and generates image Output, output
Network generation image Output be through the transformed target area image of image area.
The whole loss function of U-shaped generation network is in the step E:
LGAN(G,D1,D2)=LD1+λ1LD2+λ2Lssim+λ3L1 (2)
The whole loss optimized the function is needed to be in entire production confrontation network:
G*=argminGANmaxD1maxD2(LGAN(G,D1,D2)+LD1+LD2) (3)
In formula (2) and (3),
Represent the true and false loss of true and false discrimination natwork output, Represent the pairing loss of pairing discrimination natwork output,It represents
The damage SSIM losses that SSIM loss functions calculate,Represent L1Regular terms loses, and x is represented
Input picture Input, y represent target area image target, λ corresponding with input picture Input1U-shaped life is accounted for for pairing loss
The parameter of weight, λ into network whole loss2Represent that SSIM losses account for the parameter of weight in U-shaped generation network whole loss, λ3
Represent L1The value of regular terms accounts for the parameter of weight in U-shaped generation network whole loss;
The training initial stage of U-shaped generation network, the ratio of true and false loss, pairing loss, SSIM losses and the loss of L1 regular terms
It is 1:1:4:1, with the increase of network training number, the ratio of true and false loss, pairing loss, SSIM losses and the loss of L1 regular terms
Example gradually becomes 1:1:0.5:1, i.e. SSSIM loss account for the parameter of weight in U-shaped generation network whole loss according to the total of setting
Body frequency of training continuously decreases.
The Mutual information entropy loss function that intersects in the step C is the intersection Mutual information entropy loss function with smooth item;
The formula of intersection Mutual information entropy loss function with smooth item is:
In formula (4), sizes of the i for batch, tiFor the sample value of prediction, yiFor true sample value, for the flat of addition
Sliding item, value be chosen for 0.005.
The generating process of the network generation image Output includes the following steps:
A), by image normalization to be converted be 256*256*3 pixels image, using the image after normalization as input
In image Input input coding networks, input picture Input by 8 layers of convolutional network in coding network, is finally exported successively
Data be 1*1*1024 characteristic image;The convolution kernel size of every layer of convolutional network is 3*3, step-length 2*2 in coding network;
B) characteristic image of 1*1*1024 generated in step a), is inputted into decoding network, characteristic image is successively by solution
8 layers of deconvolution network of code network, while the characteristic image after every layer of convolutional network operation in step a) is input to identical number
According to operation is carried out in the warp lamination of tensor size, ultimately generate complete network generation image Output, i.e. warp lamination
Input not only has the characteristic image from last layer de-convolution operation, also corresponds to the convolution algorithm characteristic image of tensor size;
Wherein, the deconvolution core size of every layer of deconvolution network is 3*3, step-length 2*2.
In the step b), in the characteristic image of the deconvolution network inputs of three first layers, rolled up by every layer in step a)
Characteristic image after product network operations, which is input in the warp lamination of identical data tensor size, added in during operation
Dropout is operated;Wherein, it is 0.2 that the parameter of Dropout operations, which uses, i.e., closes in two articulamentums 20% connection at random
Node.
Using the sliding window form of convolution kernel, sliding window size is 7*7 for the calculating of SSIM algorithms in the step E4.
The present invention has the following advantages:
(1) network is generated by including the confrontation type of U-shaped generation network, pairing discrimination natwork and true and false discrimination natwork, built
The network model of vertical U-shaped generation network, then the conversion of the image area for the network model realization topography for passing through foundation, are filled up
Instantly it without the vacancy of the confrontation type generation network for topography's conversion, improves confrontation type generation network and turns in image area
The use scope in field is changed, and improves the effect and unfailing performance of image area conversion;
(2) generation network network model training process in, using pairing discrimination natwork and true and false discrimination natwork it is more
Confrontation mode carries out, in order to improve network training initial stage discrimination natwork judgement it is poor the problem of, increase SSIM loss letter
Number, using SSIM algorithms calculate image similarity, and using its result of calculation as lose to more newly-generated network weight,
Using the result of calculation of SSIM algorithms as loss, the problem of confrontation network initial stage antagonism is low can be made up so that generation
Network can preferably restrain, and never obtain better image domain conversion effect;
(3) by using intersection Mutual information entropy loss function in the training process of multipair anti-discrimination natwork so that multipair
The training of anti-discrimination natwork is more stablized, since traditional cross entropy loss function contains log operations, at the initial stage of counting loss
Stage so that the fluctuation of loss is big, so that in the training stage it is possible that there is the result that loss function is 0 occurring, causes to instruct
Experienced failure, by increasing smooth item, fluctuation when reducing trained, it is therefore prevented that trained failure scenarios occur, and improve depth
The stability of degree confrontation network training;
(4) by being arranged in generation network, the input of warp lamination not only has the spy from last layer de-convolution operation
Image is levied, also corresponds to the convolution algorithm characteristic image of tensor size so that the information of image is protected to the full extent
It stays, more preferably, more completely preserves the characteristic information in original image, improve the effect and authenticity of image area conversion, Er Qie
The three first layers convolutional layer of coding network adds in Dropout operations into the operation of corresponding warp lamination input feature vector image, effectively
The image unification obtained after decoding is prevented, further improves image area conversion quality;
(3) by using Leak ReLU active coatings and by the parameter of Leak ReLU be 0.2, preferably remain generation net
The information of network artwork image field, and network is made to retain residual information as far as possible in backpropagation, improve the complete of conversion image
Whole degree, ensure that conversion effect;
Description of the drawings
Fig. 1 is the network structure of the present invention;
Fig. 2 is U-shaped generation network structure in Fig. 1;
Fig. 3 is the network structure of true and false discrimination natwork in Fig. 1;
Fig. 4 is the network structure that discrimination natwork is matched in Fig. 1;
Fig. 5 is U-shaped generation network training figure in Fig. 1;
Fig. 6 is the network training figure of true and false discrimination natwork in Fig. 1;
Fig. 7 is the network training figure that discrimination natwork is matched in Fig. 1.
Specific embodiment
For a better understanding of the present invention, technical scheme of the present invention is described further below in conjunction with the accompanying drawings.
As shown in Figure 1, the present invention includes U-shaped generation network U-net, true and false discrimination natwork D1-net and pairing discrimination natwork
D2-net, further comprises structural similarity numerical computations part, and structural similarity numerical computations are calculated including SSIM loss functions
Part and L1 regularizations part;U-shaped generation network U-net is used for the conversion of image area, and true and false discrimination natwork D1-net is to be used for
Judge that network generates the whether true discriminators of image Output, discrimination natwork D2-net is for judging that network generates image
The discriminator whether matched of Output and original image;
As shown in Fig. 2, U-shaped generation network U-net includes coding network F-net and decoding network G-net, coding network F-
Net carries out image the characteristic pattern that convolution operation exports its higher-dimension, and decoding network G-net uses deconvolution network, by spy
Sign figure carries out the generation that image is completed in deconvolution.
The output terminal connection decoding of input terminal access the input picture Input, coding network F-net of coding network F-net
The input terminal of network G-net, the output terminal generation network generation image Output of decoding network G-net.
Coding network F-net includes eight layers of convolutional network, and the convolution kernel size of every layer of convolutional network is 3*3, step-length 2*
2, every layer of convolutional network includes convolutional layer, Normalization layers of Batch and Leak ReLU active coatings, due to generating net
Network needs reservation original image domain information as possible and in order to which network is made to retain residual information as far as possible in backpropagation,
Therefore the present invention activates the selection of every layer of activation primitive of coding network F-net in generation network using Leak ReLU
Function, the wherein alpha parameter of Leak ReLU active coatings are 0.2.
Decoding network G-net includes eight layers of deconvolution network, and the deconvolution core size of every layer of deconvolution network is 3*3, is walked
A length of 2*2, every layer of deconvolution network include warp lamination, Normalization layers of Batch and active coating, and first to layer 7
The active coating of deconvolution network uses ReLU active coatings, and the active coating of the 8th layer of deconvolution network uses tanh active coatings.
As shown in figure 3, true and false discrimination natwork D1-net is for differentiating whether the image of generation is true image, therefore its
It inputs as a sub-picture, it is true and false to export;True and false discrimination natwork D1-net includes the true and false discriminating convolution net that multilayer is transmitted successively
Network, every layer of true and false discriminating convolutional network include convolutional layer, Normalization layers of Batch and active coating.
ReLU activation primitives can effectively transmit residual error and can keep nonlinear fitting, and tanh activation primitive output valves are -1
To between 1, and the value of sigmoid activation primitives value output is convenient for being calculated with label between 0 to 1, therefore true and false discriminating net
The excitation function used except remaining every layer of last layer of network D1-net is ReLU functions, last layer of output layer uses
Sigmoid activation primitives, i.e., last layer true and false active coating for differentiating convolutional network is using Sigmoid activation primitives, other layers
The true and false active coating for differentiating convolutional network uses ReLU functions.
The design of every layer of true and false convolution kernel for differentiating convolutional network follows the principle of small convolution kernel, using the volume of 3*3
Product core removes the true and false discriminating convolutional networks of 32*32*256 to the true and false every layer of step-length stride between differentiating convolutional network of 30*30*1
For 1 convolution algorithm, remaining step-length stride is 2.The generation of gradient diffusing phenomenon in order to prevent adds in every layer of network
Enter Normalization layers of Bacth, while play the role of pond for 2 in view of the step-length of part layer convolutional calculation, because
Pooling layers have been not added in this network.
Whether images of the pairing discrimination natwork D2-net for discriminating generation matches, therefore its input is with the image of input
Two images;Pairing discrimination natwork D2-net includes the Concat layers transmitted successively and multilayer matches and differentiates convolutional network, every layer
Pairing differentiates that convolutional network includes convolutional layer, Normalization layers of Batch and active coating, and the pairing of last layer differentiates volume
For the active coating of product network using Sigmoid activation primitives, other layers of pairing differentiate that the active coating of convolutional network uses ReLU functions.
As shown in figure 4, pairing discrimination natwork D2-net is similar with the structure of true and false discrimination natwork D1-net, only in first layer
Input when more a sub-pictures, i.e. tensor of the input of first layer for 256*256*6, it is similary that ReLU and sigmoid is selected to swash
Function living, using Normalization layers of Bacth, and the loss function used equally carries smooth item with D1-net.
If it is target area image target with the true target area image that input picture Input is matched one by one;Network is given birth to
The negative sample of true and false discrimination natwork D1-net is inputted as the training negative sample of true and false discrimination natwork D1-net into image Output
Input terminal, target area image target input true and false discrimination natwork D1- as the training positive sample of true and false discrimination natwork D1-net
The positive sample input terminal of net, the value of true and false discrimination natwork D1-net outputs feed back to decoding network G-net as true and false penalty values
True and false loss input terminal;Network generates image Output and corresponding input picture Input as training negative sample input pairing
The negative sample input terminal of discrimination natwork D2-net, target area image target and corresponding input picture Input are as the positive sample of training
The positive sample input terminal of this input pairing discrimination natwork D2-net, the value of pairing discrimination natwork D2-net outputs are lost as pairing
Value feeds back to the pairing loss input terminal of decoding network G-net;By network generation image Output and target area image target
Between structural similarity numerical value as compensation penalty values feed back to decoding network G-net compensation lose input terminal, wherein, knot
The calculating of structure similarity figure includes SSIM loss functions and L1 regularizations.
The invention also includes a kind of image area conversion methods based on production confrontation network, include the following steps:
1), the U-shaped generation network U-net of training establishes the network model of U-shaped generation network U-net;Specifically include following step
Suddenly:
A, the training image collection in domain to be converted is collected, training image concentrates the original area image for including matching one by one and target
Area image, the original domain image normalization that training image is concentrated are the image of 256*256*3 pixels, and the image after normalization is
Input picture Input during for network training, it is target corresponding with input picture Input that training image, which concentrates aiming field image,
Area image target;
B, input picture Input obtained by step A is converted into the network of training network through U-shaped generation network U-net
Generate image Output;
C, image is generated using input picture Input, target area image target and network obtained by step A and step B
Output carries out the training of multipair anti-discrimination natwork:The training of multipair anti-discrimination natwork includes the instruction of true and false discrimination natwork D1-net
Practice and match the training of discrimination natwork D2-net;
As shown in fig. 6, the training of true and false discrimination natwork D1-net includes the following steps:
C11:The network weight of true and false discrimination natwork D1-net is initialized using random initializtion method;
C12:Network is generated into image Output as negative sample, target area image corresponding with input picture Input
Target carries out two classification based trainings as positive sample in true and false discrimination natwork D1-net, mutual using the intersection with smooth item
Comentropy loss function and adam optimization algorithms update the network weight of true and false discrimination natwork D1-net;
Intersecting Mutual information entropy can be larger in the incipient stage value of counting loss, while is led in the training stage it is possible that occurring 0
Failure to train is caused, increases fluctuation when smooth item reduces trained, it is therefore prevented that trained failure scenarios occur, improved with flat
The intersection Mutual information entropy function of sliding item improves the stability of depth confrontation network training;
To be the label that positive and negative samples generate one-hot types, then according to last first in two classification based trainings
0 to 1 value of layer sigmoid activation primitive outputs calculates cross entropy using the SigmoidCrossEntropy for adding smooth item
Loss finally updates true and false discrimination natwork D1-net weights according to the loss of feedback;
Wherein, the formula of the intersection Mutual information entropy loss function with smooth item is
In formula (1), sizes of the i for batch, tiFor the sample value of prediction, yiFor true sample value, for the flat of addition
Sliding item, value be chosen for 0.005;
As shown in fig. 7, the training of pairing discrimination natwork D2-net includes the following steps:
C21:The network weight for matching discrimination natwork D2-net is initialized using random initializtion method:
C22:Network is generated into image Output and corresponding input picture Input as negative sample, by input picture
Input and corresponding target area image target carries out the instruction of two classification as positive sample in discrimination natwork D2-net is matched
Practice, use the intersection Mutual information entropy loss function with smooth item and adam optimization algorithms update pairing discrimination natwork D2-net
Network weight;
D, step C is repeated, after anti-discrimination natwork training multipair twice, fixed true and false discrimination natwork D1-net and pairing
The network weight of discrimination natwork D2-net:It is true using multiple training in order to which the training for making U-shaped generation network U-net is more stablized
False discrimination natwork D1-net and pairing discrimination natwork D2-net, the strategy of the U-shaped generation network U-net of retraining;
E, using the multipair anti-discrimination natwork of the gained after step D training, U-shaped generation network U-net is trained;
As shown in figure 5, the training of U-shaped generation network U-net includes the following steps:
E1:The network weight of U-shaped generation network U-net is initialized using Ha Weier random initializtions method;
E2:Network generation image Output is inputted into true and false discrimination natwork D1-net, true and false discrimination natwork D1-net outputs
True and false penalty values, and the true and false penalty values of output are fed back into the decoding network G-net in U-shaped generation network U-net, for more
Its new network weight:True and false discrimination natwork D1-net exports 30*30*1 images, is approached for returning to network generation image Output
The penalty values of true picture, wherein, ranging from 0 to the 1 of each image pixel point value of true and false discrimination natwork D1-net outputs, as
Vegetarian refreshments value closer to 1 represent input picture Input in the pixel receptive field region closer to true image, pixel point value is got over
Close to 0 expression input picture Input true image is more kept off in the pixel receptive field region;
E3:By input picture Input and corresponding network generation image Output input pairing discrimination natwork D2-net, match
Pairing penalty values are exported to discrimination natwork D2-net, and the pairing penalty values of output are fed back in U-shaped generation network U-net
Decoding network G-net, for updating its network weight:Discrimination natwork D2-net output 30*30*1 images are matched, it is defeated for returning
Enter image Input and network generation image Output whether be input picture Input and target area image target pairings damage
Mistake value;Wherein, ranging from 0 to the 1 of each image pixel point value of pairing discrimination natwork D2-net outputs represents defeated closer to 1
Enter image Input and network generation image Output is more matched, represent more to mismatch closer to 0;
E4:The structural similarity numerical value between network generation image Output and target area image target is calculated, and
The decoding network G-net in U-shaped generation network U-net is fed back to using the structural similarity numerical value calculated as loss, is used for
Update its network weight;The calculating of structural similarity numerical value includes SSIM loss functions and L1 regularizations, wherein, SSIM losses
Function derive from SSIM algorithms, be it is a kind of weigh two images similitude index, output valve SSIM (x, y) table of SSIM algorithms
Show the similitude between two images, i.e. structural similarity between input picture x and target area image y, the model of SSIM (x, y)
It is -1 to 1 to enclose, and represents that the similarity of two images is higher close to 1, as input picture x and target area image y striking resemblances,
The value of SSIM (x, y) is equal to 1;Using SSIM as loss function, enable to generation network that can preferably restrain, so as to obtain
Better image domain conversion effect;
The calculation formula of SSIM algorithm output valves is as follows:
In formula (2), x is input picture Input, and y is target area image target corresponding with input picture Input,
μxIt is the average value of x, μyIt is the average value of y,It is the variance of x,It is the variance of y, σxyIt is the covariance of x and y, c1=
(k1L)2With2=(k2L)2It is for maintaining stable constant, L is the dynamic range of pixel value, k1=0.01,2=0.03;
F, step C~E is the weight training of primary U-shaped generation network U-net, repeats step C~E, completes U-shaped life twice
Into after the weight training of network U-net, that is, the training of U-shaped generation network U-net is completed, gained generation network is U-shaped generation
The network model of network U-net;
However, needing to be converted to the form of sliding window when calculating two images SSIM values, cause to select different size of cunning
The difference of window and parameter δ, obtained result have differences.The SSIM algorithms that Wang et.al et al. are initially proposed use
Be 11*11 sliding window, the image generated however, as generation network in initial stage differs larger with target image, big ruler
SSIM values between the two images that very little sliding window calculates are in close proximity to 0 in initial period numerical value so that loss can not have
Effect passes back to generation network, leads to the failure to train of confrontation type generation network G AN.Given this problem and in view of network inputs
It is the photo of 256*256 pixels, the calculating of SSIM algorithms is using the sliding window form of convolution kernel, the sliding window of selection in the final present invention
Size is 7*7.
In the above-mentioned step E being trained to U-shaped generation network U-net, the whole loss of U-shaped generation network U-net
Function is:
LGAN(G,D1,D2)=LD1+λ1LD2+λ2Lssim+λ3L1 (3)
The whole loss optimized the function is needed to be in entire production confrontation network:
G*=argminGANmaxD1maxD2(LGAN(G,D1,D2)+LD1+LD2) (4)
In formula (3) and (4),
Represent the true and false loss of true and false discrimination natwork D1-net outputs,Represent pairing discrimination natwork D2-
The pairing loss of net outputs,Represent the SSIM damages that SSIM loss functions calculate
It loses,Represent L1Regular terms loses, and x represents input picture Input, and y is represented and input
Image Input corresponding target area image target, λ1It accounts in the decoding network G-net of whole generation network and weighs for pairing loss
The parameter of weight, λ2Represent that SSIM losses account for the parameter of weight in the whole decoding network G-net for generating network, λ3Represent L1Canonical
The value of item accounts for the parameter of weight in the whole decoding network G-net for generating network.
At the training initial stage of U-shaped generation network U-net, true and false loss, pairing loss, SSIM losses and the loss of L1 regular terms
Ratio be 1:1:4:1, with the increase of network training number, true and false loss, pairing loss, SSIM losses and L1 regular terms damage
The ratio of mistake gradually becomes 1:1:0.5:1, i.e. SSSIM loss account for it is U-shaped generation network U-net whole loss in weight parameter by
It is continuously decreased according to the overall frequency of training of setting.
It can make when network training initial stage true and false discrimination natwork D1-net is low with pairing discrimination natwork D2-net distinguishing abilities
With the residual error of the U-shaped generation network U-net of SSIM loss functions feedback, target area image can be effectively generated, when true and false mirror
Other network D1-net and pairing discrimination natwork D2-net then reduces SSIM loss functions anti-after ability is constantly promoted in training
Shared weight in feedback generation network residual error so that generate the residual error major part of network comes from true and false discrimination natwork D1-net
The loss fed back with pairing discrimination natwork D2-net, generation more preferable than the effect of conventional images domain conversion method to obtain
The more true effect of image.
2), by image to be converted by normalized, the pixel after normalization is 256*256, the figure after normalization
The network model established as input through step 1), you can complete the image area conversion of image to be converted:Image after normalization is made
For input picture Input input steps 1) in the network model established, coding network F-net extractions input picture Input's
Then high dimensional feature exports network by decoding network G-net and generates image Output, the network generation image Output of output
As through the transformed target area image of image area.
In the image area transfer process of U-shaped generation network, need image input coding network F-net first carrying out convolution
Operation, then carries out deconvolution operation, realizes the conversion of image area again;But;Previous U-shaped generation network, artwork is in convolution
The partial information of image is difficult to retain in the process, therefore, in order to preferably, more completely preserve the feature letter in original image
Breath, the generating process of inventive network generation image Output include the following steps:
A), by image normalization to be converted be 256*256*3 pixels image, using the image after normalization as input
In image Input input coding networks F-net, input picture Input is successively by 8 layers of convolution net in coding network F-net
Network, the data finally exported are the characteristic image of 1*1*1024;The convolution kernel size of every layer of convolutional network in coding network F-net
For 3*3, step-length 2*2;
B) characteristic image of 1*1*1024 generated in step a), is inputted into decoding network G-net, characteristic image passes through successively
8 layers of deconvolution network of decoding network G-net are crossed, while the characteristic image after every layer of convolutional network operation in step a) is inputted
Operation is carried out into the warp lamination of identical data tensor size, ultimately generates complete network generation image Output, i.e., instead
The input of convolutional layer not only has the characteristic image from last layer de-convolution operation, and the convolution algorithm for also corresponding to tensor size is special
Levy image;Wherein, the deconvolution core size of every layer of deconvolution network is 3*3, step-length 2*2.
In the characteristic image of the deconvolution network inputs of three first layers, after by every layer of convolutional network operation in step a)
Characteristic image, which is input in the warp lamination of identical data tensor size, to carry out adding in Dropout operations during operation;Its
In, it is 0.2 that the parameter of Dropout operations, which uses, i.e., closes in two articulamentums 20% connecting node at random.
The input of warp lamination not only has the characteristic image from last layer de-convolution operation, also corresponds to tensor size
Convolution algorithm characteristic image so that the information of image is retained to the full extent, more preferably, more completely preserves original image
In characteristic information, improve the effect and authenticity of image area conversion, and the figure that decoding network G-net is obtained in order to prevent
As unification, added in the three first layers convolutional layer of coding network F-net into the operation of corresponding warp lamination input feature vector image
Dropout is operated, and is effectively prevent the image unification obtained after decoding, is further improved image area conversion quality.
Finally it should be noted that:The above embodiments are merely illustrative of the technical solutions of the present invention, rather than its limitations;Although
The present invention is described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that, still may be used
Either which part or all technical features are equal with modifying to the technical solution recorded in previous embodiment
It replaces, and these modifications or replacement, the model for technical solution of the embodiment of the present invention that it does not separate the essence of the corresponding technical solution
It encloses.
Claims (10)
1. a kind of image area switching network based on production confrontation network, it is characterised in that:Including U-shaped generation network, true and false
Discrimination natwork and pairing discrimination natwork, U-shaped generation network include coding network and decoding network, and the input terminal of coding network accesses
Input picture Input, the input terminal of the output terminal connection decoding network of coding network, the output terminal output network life of decoding network
Into image Output;If it is target area image target with the true target area image that input picture Input is matched one by one;Net
Network generates the negative sample input terminal that image Output inputs true and false discrimination natwork as the training negative sample of true and false discrimination natwork, mesh
The positive sample input terminal that area image target inputs true and false discrimination natwork as the training positive sample of true and false discrimination natwork is marked, it is true and false
The value of discrimination natwork output feeds back to the true and false loss input terminal of decoding network as true and false penalty values;Network generates image
Negative sample input terminals of the Output and corresponding input picture Input as training negative sample input pairing discrimination natwork, aiming field
Positive sample input terminals of the image target and corresponding input picture Input as training positive sample input pairing discrimination natwork, matches
Input terminal is lost in the pairing that decoding network is fed back to as pairing penalty values to the value that discrimination natwork exports;Network is generated into image
Structural similarity value feeds back to the compensation damage of decoding network as compensation penalty values between Output and target area image target
Lose input terminal.
2. the image area switching network according to claim 1 based on production confrontation network, it is characterised in that:The volume
Code network includes eight layers of convolutional network, and the convolution kernel size of every layer of convolutional network is 3*3, step-length 2*2, and every layer of convolutional network is equal
Including convolutional layer, Normalization layers of Batch and Leak ReLU active coatings, the alpha parameter of Leak ReLU active coatings is
0.2;The decoding network includes eight layers of deconvolution network, and the deconvolution core size of every layer of deconvolution network is 3*3, step-length 2*
2, every layer of deconvolution network includes warp lamination, BatchNormalization layers and active coating, and first to layer 7 deconvolution
The active coating of network uses ReLU active coatings, and the active coating of the 8th layer of deconvolution network uses tanh active coatings.
3. the image area switching network according to claim 2 based on production confrontation network, it is characterised in that:It is described true
False discrimination natwork includes the true and false discriminating convolutional network that multilayer is transmitted successively, and every layer of true and false discriminating convolutional network includes convolution
Layer, Normalization layers of Batch and active coating, last layer of true and false active coating for differentiating convolutional network use Sigmoid
Activation primitive, other layer of true and false active coating for differentiating convolutional network use ReLU functions.
4. the image area switching network according to claim 3 based on production confrontation network, it is characterised in that:It is described to match
Include the Concat layers transmitted successively to discrimination natwork and multilayer pairing differentiates convolutional network, every layer of pairing differentiates that convolutional network is equal
Including convolutional layer, Normalization layers of Batch and active coating, last layer of pairing differentiates that the active coating of convolutional network uses
Sigmoid activation primitives, other layers of pairing differentiate that the active coating of convolutional network uses ReLU functions.
5. a kind of image area conversion method of the image area switching network based on production confrontation network described in claim 4,
It is characterized in that:Include the following steps:
1), the U-shaped generation network of training establishes the network model of U-shaped generation network;Specifically include following steps:
A, the training image collection in domain to be converted is collected, training image concentrates the original area image for including matching one by one and aiming field figure
Picture, the original area image concentrated to training image are normalized, the input figure when image after normalization is network training
As Input, it is target area image target corresponding with input picture Input that training image, which concentrates aiming field image,;
B, the networks for being converted into training network through U-shaped generation network of the input picture Input obtained by step A are generated into image
Output;
C, input picture Input, target area image target obtained by step A and step B and network generation image Output are utilized
Carry out the training of multipair anti-discrimination natwork:The training of multipair anti-discrimination natwork includes the training of true and false discrimination natwork and pairing differentiates
The training of network, wherein, the training of true and false discrimination natwork includes the following steps:
C11:The network weight of true and false discrimination natwork is initialized using random initializtion method;
C12:Network is generated into image Output as negative sample, target area image target corresponding with input picture Input
It as positive sample, is trained in true and false discrimination natwork, uses intersection Mutual information entropy loss function and adam optimization algorithms
Update the network weight of true and false discrimination natwork;
The training of pairing discrimination natwork includes the following steps:
C21:The network weight for matching discrimination natwork is initialized using random initializtion method:
C22:Using network generation image Output and corresponding input picture Input as negative sample, by input picture Input with
Corresponding target area image target is trained in discrimination natwork is matched, is damaged using Mutual information entropy is intersected as positive sample
Lose the network weight of function and adam optimization algorithms update pairing discrimination natwork;
D, step C is repeated, after anti-discrimination natwork training multipair twice, fixed true and false discrimination natwork and pairing discrimination natwork
Network weight;
E, using the multipair anti-discrimination natwork of the gained after step D training, U-shaped generation network is trained;It specifically includes
Following steps:
E1:The network weight of U-shaped generation network is initialized using Ha Weier random initializtions method;
E2:Network generation image Output is inputted into true and false discrimination natwork, true and false discrimination natwork exports true and false penalty values, and will be defeated
The true and false penalty values gone out feed back to the decoding network in U-shaped generation network, for updating its network weight:True and false discrimination natwork is defeated
Go out 30*30*1 images, image Output is generated close to the penalty values of true picture for returning to network, wherein, true and false discriminating net
Ranging from 0 to the 1 of each image pixel point value of network output, pixel point value represent input picture Input in the pixel closer to 1
Point receptive field region represents input picture Input in the pixel receptive field closer to true image, pixel point value closer to 0
More keep off true image in region;
E3:By input picture Input and corresponding network generation image Output input pairing discrimination natworks, discrimination natwork is matched
Output pairing penalty values, and the pairing penalty values of output are fed back into the decoding network in U-shaped generation network, for updating its net
Network weight:Discrimination natwork output 30*30*1 images are matched, for returning to input picture Input and network generation image Output
The penalty values for whether being input picture Input with target area image target pairings;Wherein, the output of pairing discrimination natwork is each
Ranging from 0 to the 1 of image pixel point value is more matched closer to 1 expression input picture Input and network generation image Output,
It represents more to mismatch closer to 0;
E4:The structure similar values between network generation image Output and target area image target are calculated, and will be calculated
The structural similarity numerical value gone out feeds back to the decoding network in U-shaped generation network as loss, for updating its network weight;Knot
Structure similarity figure includes SSIM loss functions result of calculation and L1 regularization result of calculations, wherein, SSIM loss functions come
Derived from SSIM algorithms, the output valve SSIM (x, y) of SSIM algorithms represents the similitude between two images, i.e. input picture x and
Structural similarity between target area image y, ranging from-the 1 to 1 of SSIM (x, y), the similarity for representing two images close to 1
Higher, as input picture x and target area image y striking resemblances, the value of SSIM (x, y) is equal to 1;
The calculation formula of SSIM algorithm output valves is as follows:
In formula (1), x is input picture Input, and y is target area image target, μ corresponding with input picture InputxIt is x
Average value, μyIt is the average value of y,It is the variance of x,It is the variance of y, σxyIt is the covariance of x and y, c1=(k1L)2With
c2=(k2L)2It is for maintaining stable constant, L is the dynamic range of pixel value, k1=0.01, k2=0.03;
F, step C~E is the weight training of primary U-shaped generation network, repeats step C~E, completes U-shaped generation network twice
After weight training, that is, the training of U-shaped generation network is completed, gained generation network is the network model of U-shaped generation network;
2), image to be converted is inputted to the network model established through step 1) after normalized, you can complete to wait to turn
Change the image area conversion of image:Image after normalization is as input picture Input input steps 1) in establish network model
In, then the high dimensional feature of coding network extraction input picture Input exports network by decoding network and generates image
Output, the network generation image Output of output is through the transformed target area image of image area.
6. the image area conversion method according to claim 5 based on production confrontation network, it is characterised in that:The step
The whole loss function of U-shaped generation network is in rapid E:
LGAN(G, D1, D2)=LD1+λ1LD2+λ2Lssim+λ3L1 (2)
The whole loss optimized the function is needed to be in entire production confrontation network:
G*=arg minGANmaxD1maxD2(LGAN(G, D1, D2)+LD1+LD2) (3)
In formula (2) and (3),Table
Show the true and false loss of true and false discrimination natwork output, Represent the pairing loss of pairing discrimination natwork output,Table
Show the damage SSIM losses that SSIM loss functions calculate,Represent L1Regular terms loses, x
Represent that input picture Input, y represent target area image target, λ corresponding with input picture Input1U is accounted for for pairing loss
The parameter of weight, λ in type generation network whole loss2Represent that SSIM losses account for the ginseng of weight in U-shaped generation network whole loss
Number, λ3Represent L1The value of regular terms accounts for the parameter of weight in U-shaped generation network whole loss;
At the training initial stage of U-shaped generation network, the ratio of true and false loss, pairing loss, SSIM losses and the loss of L1 regular terms is 1: 1
: 4: 1, with the increase of network training number, it is true and false loss, pairing loss, SSIM loss and L1 regular terms loss ratio by
1: 1: 0.5: 1 is faded to, i.e. the parameter that SSSIM losses account for weight in U-shaped generation network whole loss is instructed according to the totality of setting
Practice number to continuously decrease.
7. the image area conversion method according to claim 5 based on production confrontation network, it is characterised in that:The step
The Mutual information entropy loss function that intersects in rapid C is the intersection Mutual information entropy loss function with smooth item;Friendship with smooth item
Fork Mutual information entropy loss function formula be:
In formula (4), sizes of the i for batch, tiFor the sample value of prediction, yiFor true sample value, EPS is the smooth of addition
, the value of EPS is chosen for 0.005.
8. the image area conversion method according to claim 5 based on production confrontation network, it is characterised in that:Described
The generating process of network generation image Output includes the following steps:
A), by image normalization to be converted be 256*256*3 pixels image, using the image after normalization as input picture
In Input input coding networks, input picture Input is successively by 8 layers of convolutional network in coding network, the number finally exported
According to the characteristic image for 1*1*1024;The convolution kernel size of every layer of convolutional network is 3*3, step-length 2*2 in coding network;
B) characteristic image of 1*1*1024 generated in step a), is inputted into decoding network, characteristic image is successively by decoding net
8 layers of deconvolution network of network, while the characteristic image after every layer of convolutional network operation in step a) is input to identical data
It measures in the warp lamination of size and carries out operation, ultimately generate the input of complete network generation image Output, i.e. warp lamination
Not only there is the characteristic image from last layer de-convolution operation, also correspond to the convolution algorithm characteristic image of tensor size;Wherein,
The deconvolution core size of every layer of deconvolution network is 3*3, step-length 2*2.
9. the image area conversion method according to claim 8 based on production confrontation network, it is characterised in that:The step
It is rapid b) in, in the characteristic image of the deconvolution network inputs of three first layers, after by every layer of convolutional network operation in step a)
Characteristic image, which is input in the warp lamination of identical data tensor size, to carry out adding in Dropout operations during operation;Its
In, it is 0.2 that the parameter of Dropout operations, which uses, i.e., closes in two articulamentums 20% connecting node at random.
10. the image area conversion method according to claim 5 based on production confrontation network, it is characterised in that:It is described
Using the sliding window form of convolution kernel, sliding window size is 7*7 for the calculating of SSIM algorithms in step E4.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711273921.5A CN108171320B (en) | 2017-12-06 | 2017-12-06 | Image domain conversion network and conversion method based on generative countermeasure network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711273921.5A CN108171320B (en) | 2017-12-06 | 2017-12-06 | Image domain conversion network and conversion method based on generative countermeasure network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN108171320A true CN108171320A (en) | 2018-06-15 |
CN108171320B CN108171320B (en) | 2021-10-19 |
Family
ID=62525151
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201711273921.5A Active CN108171320B (en) | 2017-12-06 | 2017-12-06 | Image domain conversion network and conversion method based on generative countermeasure network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108171320B (en) |
Cited By (32)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109166144A (en) * | 2018-07-20 | 2019-01-08 | 中国海洋大学 | A kind of image depth estimation method based on generation confrontation network |
CN109190524A (en) * | 2018-08-17 | 2019-01-11 | 南通大学 | A kind of human motion recognition method based on generation confrontation network |
CN109215014A (en) * | 2018-07-02 | 2019-01-15 | 中国科学院深圳先进技术研究院 | Training method, device, equipment and the storage medium of CT image prediction model |
CN109272486A (en) * | 2018-08-14 | 2019-01-25 | 中国科学院深圳先进技术研究院 | Training method, device, equipment and the storage medium of MR image prediction model |
CN109543674A (en) * | 2018-10-19 | 2019-03-29 | 天津大学 | A kind of image copy detection method based on generation confrontation network |
CN109639710A (en) * | 2018-12-29 | 2019-04-16 | 浙江工业大学 | A kind of network attack defence method based on dual training |
CN109671018A (en) * | 2018-12-12 | 2019-04-23 | 华东交通大学 | A kind of image conversion method and system based on production confrontation network and ResNets technology |
CN109711254A (en) * | 2018-11-23 | 2019-05-03 | 北京交通大学 | The image processing method and device of network are generated based on confrontation |
CN109740682A (en) * | 2019-01-08 | 2019-05-10 | 南京大学 | A kind of image-recognizing method converting and generate model based on domain |
CN110210422A (en) * | 2019-06-05 | 2019-09-06 | 哈尔滨工业大学 | It is a kind of based on optical imagery auxiliary naval vessel ISAR as recognition methods |
CN110211035A (en) * | 2019-04-18 | 2019-09-06 | 天津中科智能识别产业技术研究院有限公司 | Merge the image super-resolution method of the deep neural network of mutual information |
CN110225260A (en) * | 2019-05-24 | 2019-09-10 | 宁波大学 | A kind of three-dimensional high dynamic range imaging method based on generation confrontation network |
CN110276811A (en) * | 2019-07-02 | 2019-09-24 | 厦门美图之家科技有限公司 | Image conversion method, device, electronic equipment and readable storage medium storing program for executing |
CN110472528A (en) * | 2019-07-29 | 2019-11-19 | 江苏必得科技股份有限公司 | A kind of metro environment target training set creation method and system |
CN110517705A (en) * | 2019-08-29 | 2019-11-29 | 北京大学深圳研究生院 | A kind of binaural sound sources localization method and system based on deep neural network and convolutional neural networks |
CN110544239A (en) * | 2019-08-19 | 2019-12-06 | 中山大学 | Multi-modal MRI conversion method, system and medium for generating countermeasure network based on conditions |
CN110660020A (en) * | 2019-08-15 | 2020-01-07 | 天津中科智能识别产业技术研究院有限公司 | Image super-resolution method of countermeasure generation network based on fusion mutual information |
CN110674335A (en) * | 2019-09-16 | 2020-01-10 | 重庆邮电大学 | Hash code and image bidirectional conversion method based on multi-generation and multi-countermeasure |
CN111047525A (en) * | 2019-11-18 | 2020-04-21 | 宁波大学 | Method for translating SAR remote sensing image into optical remote sensing image |
CN111161239A (en) * | 2019-12-27 | 2020-05-15 | 上海联影智能医疗科技有限公司 | Medical image analysis method, device, storage medium and computer equipment |
CN111210409A (en) * | 2019-12-30 | 2020-05-29 | 浙江大学 | Condition-based generation confrontation network structure damage identification method |
CN111241725A (en) * | 2019-12-30 | 2020-06-05 | 浙江大学 | Structure response reconstruction method for generating countermeasure network based on conditions |
CN111241614A (en) * | 2019-12-30 | 2020-06-05 | 浙江大学 | Engineering structure load inversion method based on condition generation confrontation network model |
CN111325661A (en) * | 2020-02-21 | 2020-06-23 | 京工数演(福州)科技有限公司 | Seasonal style conversion model and method for MSGAN image |
CN111597946A (en) * | 2020-05-11 | 2020-08-28 | 腾讯科技(深圳)有限公司 | Processing method of image generator, image generation method and device |
CN112001397A (en) * | 2020-08-25 | 2020-11-27 | 广东光速智能设备有限公司 | Method and system for generating identification card character recognition training data of intelligent construction site |
CN112183325A (en) * | 2020-09-27 | 2021-01-05 | 哈尔滨市科佳通用机电股份有限公司 | Road vehicle detection method based on image comparison |
CN112465115A (en) * | 2020-11-25 | 2021-03-09 | 科大讯飞股份有限公司 | GAN network compression method, device, equipment and storage medium |
CN112633306A (en) * | 2019-09-24 | 2021-04-09 | 杭州海康威视数字技术股份有限公司 | Method and device for generating confrontation image |
CN113221897A (en) * | 2020-02-06 | 2021-08-06 | 马上消费金融股份有限公司 | Image correction method, image text recognition method, identity verification method and device |
CN113837048A (en) * | 2021-09-17 | 2021-12-24 | 南京信息工程大学 | Vehicle weight recognition method based on less sample attention |
CN114638745A (en) * | 2022-03-16 | 2022-06-17 | 江南大学 | Medical image intelligent conversion method based on multiple reference information |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20140279506A1 (en) * | 2013-03-14 | 2014-09-18 | Bank Of America Corporation | User interface for mobile payment via transfer network |
CN106845471A (en) * | 2017-02-20 | 2017-06-13 | 深圳市唯特视科技有限公司 | A kind of vision significance Forecasting Methodology based on generation confrontation network |
CN107123151A (en) * | 2017-04-28 | 2017-09-01 | 深圳市唯特视科技有限公司 | A kind of image method for transformation based on variation autocoder and generation confrontation network |
CN107368752A (en) * | 2017-07-25 | 2017-11-21 | 北京工商大学 | A kind of depth difference method for secret protection based on production confrontation network |
-
2017
- 2017-12-06 CN CN201711273921.5A patent/CN108171320B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20140279506A1 (en) * | 2013-03-14 | 2014-09-18 | Bank Of America Corporation | User interface for mobile payment via transfer network |
CN106845471A (en) * | 2017-02-20 | 2017-06-13 | 深圳市唯特视科技有限公司 | A kind of vision significance Forecasting Methodology based on generation confrontation network |
CN107123151A (en) * | 2017-04-28 | 2017-09-01 | 深圳市唯特视科技有限公司 | A kind of image method for transformation based on variation autocoder and generation confrontation network |
CN107368752A (en) * | 2017-07-25 | 2017-11-21 | 北京工商大学 | A kind of depth difference method for secret protection based on production confrontation network |
Cited By (54)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109215014A (en) * | 2018-07-02 | 2019-01-15 | 中国科学院深圳先进技术研究院 | Training method, device, equipment and the storage medium of CT image prediction model |
CN109215014B (en) * | 2018-07-02 | 2022-03-04 | 中国科学院深圳先进技术研究院 | Training method, device and equipment of CT image prediction model and storage medium |
CN109166144B (en) * | 2018-07-20 | 2021-08-24 | 中国海洋大学 | Image depth estimation method based on generation countermeasure network |
CN109166144A (en) * | 2018-07-20 | 2019-01-08 | 中国海洋大学 | A kind of image depth estimation method based on generation confrontation network |
CN109272486A (en) * | 2018-08-14 | 2019-01-25 | 中国科学院深圳先进技术研究院 | Training method, device, equipment and the storage medium of MR image prediction model |
CN109272486B (en) * | 2018-08-14 | 2022-07-08 | 中国科学院深圳先进技术研究院 | Training method, device and equipment of MR image prediction model and storage medium |
CN109190524A (en) * | 2018-08-17 | 2019-01-11 | 南通大学 | A kind of human motion recognition method based on generation confrontation network |
CN109190524B (en) * | 2018-08-17 | 2021-08-13 | 南通大学 | Human body action recognition method based on generation of confrontation network |
CN109543674B (en) * | 2018-10-19 | 2023-04-07 | 天津大学 | Image copy detection method based on generation countermeasure network |
CN109543674A (en) * | 2018-10-19 | 2019-03-29 | 天津大学 | A kind of image copy detection method based on generation confrontation network |
CN109711254A (en) * | 2018-11-23 | 2019-05-03 | 北京交通大学 | The image processing method and device of network are generated based on confrontation |
CN109711254B (en) * | 2018-11-23 | 2020-12-15 | 北京交通大学 | Image processing method and device based on countermeasure generation network |
CN109671018A (en) * | 2018-12-12 | 2019-04-23 | 华东交通大学 | A kind of image conversion method and system based on production confrontation network and ResNets technology |
CN109639710B (en) * | 2018-12-29 | 2021-02-26 | 浙江工业大学 | Network attack defense method based on countermeasure training |
CN109639710A (en) * | 2018-12-29 | 2019-04-16 | 浙江工业大学 | A kind of network attack defence method based on dual training |
CN109740682A (en) * | 2019-01-08 | 2019-05-10 | 南京大学 | A kind of image-recognizing method converting and generate model based on domain |
CN109740682B (en) * | 2019-01-08 | 2020-07-28 | 南京大学 | Image identification method based on domain transformation and generation model |
CN110211035A (en) * | 2019-04-18 | 2019-09-06 | 天津中科智能识别产业技术研究院有限公司 | Merge the image super-resolution method of the deep neural network of mutual information |
CN110211035B (en) * | 2019-04-18 | 2023-03-24 | 天津中科智能识别产业技术研究院有限公司 | Image super-resolution method of deep neural network fusing mutual information |
CN110225260B (en) * | 2019-05-24 | 2021-02-19 | 宁波大学 | Three-dimensional high dynamic range imaging method based on generation countermeasure network |
CN110225260A (en) * | 2019-05-24 | 2019-09-10 | 宁波大学 | A kind of three-dimensional high dynamic range imaging method based on generation confrontation network |
CN110210422A (en) * | 2019-06-05 | 2019-09-06 | 哈尔滨工业大学 | It is a kind of based on optical imagery auxiliary naval vessel ISAR as recognition methods |
CN110210422B (en) * | 2019-06-05 | 2021-04-27 | 哈尔滨工业大学 | Ship ISAR image identification method based on optical image assistance |
CN110276811B (en) * | 2019-07-02 | 2022-11-01 | 厦门美图之家科技有限公司 | Image conversion method and device, electronic equipment and readable storage medium |
CN110276811A (en) * | 2019-07-02 | 2019-09-24 | 厦门美图之家科技有限公司 | Image conversion method, device, electronic equipment and readable storage medium storing program for executing |
CN110472528B (en) * | 2019-07-29 | 2021-09-17 | 江苏必得科技股份有限公司 | Subway environment target training set generation method and system |
CN110472528A (en) * | 2019-07-29 | 2019-11-19 | 江苏必得科技股份有限公司 | A kind of metro environment target training set creation method and system |
CN110660020B (en) * | 2019-08-15 | 2024-02-09 | 天津中科智能识别产业技术研究院有限公司 | Image super-resolution method of antagonism generation network based on fusion mutual information |
CN110660020A (en) * | 2019-08-15 | 2020-01-07 | 天津中科智能识别产业技术研究院有限公司 | Image super-resolution method of countermeasure generation network based on fusion mutual information |
CN110544239B (en) * | 2019-08-19 | 2021-12-17 | 中山大学 | Multi-modal MRI conversion method, system and medium for generating countermeasure network based on conditions |
CN110544239A (en) * | 2019-08-19 | 2019-12-06 | 中山大学 | Multi-modal MRI conversion method, system and medium for generating countermeasure network based on conditions |
CN110517705B (en) * | 2019-08-29 | 2022-02-18 | 北京大学深圳研究生院 | Binaural sound source positioning method and system based on deep neural network and convolutional neural network |
CN110517705A (en) * | 2019-08-29 | 2019-11-29 | 北京大学深圳研究生院 | A kind of binaural sound sources localization method and system based on deep neural network and convolutional neural networks |
CN110674335B (en) * | 2019-09-16 | 2022-08-23 | 重庆邮电大学 | Hash code and image bidirectional conversion method based on multiple generation and multiple countermeasures |
CN110674335A (en) * | 2019-09-16 | 2020-01-10 | 重庆邮电大学 | Hash code and image bidirectional conversion method based on multi-generation and multi-countermeasure |
CN112633306A (en) * | 2019-09-24 | 2021-04-09 | 杭州海康威视数字技术股份有限公司 | Method and device for generating confrontation image |
CN112633306B (en) * | 2019-09-24 | 2023-09-22 | 杭州海康威视数字技术股份有限公司 | Method and device for generating countermeasure image |
CN111047525A (en) * | 2019-11-18 | 2020-04-21 | 宁波大学 | Method for translating SAR remote sensing image into optical remote sensing image |
CN111161239A (en) * | 2019-12-27 | 2020-05-15 | 上海联影智能医疗科技有限公司 | Medical image analysis method, device, storage medium and computer equipment |
CN111161239B (en) * | 2019-12-27 | 2024-02-27 | 上海联影智能医疗科技有限公司 | Medical image analysis method, device, storage medium and computer equipment |
CN111241725A (en) * | 2019-12-30 | 2020-06-05 | 浙江大学 | Structure response reconstruction method for generating countermeasure network based on conditions |
CN111241614A (en) * | 2019-12-30 | 2020-06-05 | 浙江大学 | Engineering structure load inversion method based on condition generation confrontation network model |
CN111210409A (en) * | 2019-12-30 | 2020-05-29 | 浙江大学 | Condition-based generation confrontation network structure damage identification method |
CN113221897A (en) * | 2020-02-06 | 2021-08-06 | 马上消费金融股份有限公司 | Image correction method, image text recognition method, identity verification method and device |
CN111325661A (en) * | 2020-02-21 | 2020-06-23 | 京工数演(福州)科技有限公司 | Seasonal style conversion model and method for MSGAN image |
CN111325661B (en) * | 2020-02-21 | 2024-04-09 | 京工慧创(福州)科技有限公司 | Seasonal style conversion model and method for image named MSGAN |
CN111597946A (en) * | 2020-05-11 | 2020-08-28 | 腾讯科技(深圳)有限公司 | Processing method of image generator, image generation method and device |
CN112001397A (en) * | 2020-08-25 | 2020-11-27 | 广东光速智能设备有限公司 | Method and system for generating identification card character recognition training data of intelligent construction site |
CN112183325A (en) * | 2020-09-27 | 2021-01-05 | 哈尔滨市科佳通用机电股份有限公司 | Road vehicle detection method based on image comparison |
CN112465115A (en) * | 2020-11-25 | 2021-03-09 | 科大讯飞股份有限公司 | GAN network compression method, device, equipment and storage medium |
CN113837048B (en) * | 2021-09-17 | 2023-08-01 | 南京信息工程大学 | Vehicle re-identification method based on less sample attention |
CN113837048A (en) * | 2021-09-17 | 2021-12-24 | 南京信息工程大学 | Vehicle weight recognition method based on less sample attention |
CN114638745A (en) * | 2022-03-16 | 2022-06-17 | 江南大学 | Medical image intelligent conversion method based on multiple reference information |
CN114638745B (en) * | 2022-03-16 | 2023-08-18 | 江南大学 | Medical image intelligent conversion method based on multi-borrowing information |
Also Published As
Publication number | Publication date |
---|---|
CN108171320B (en) | 2021-10-19 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108171320A (en) | A kind of image area switching network and conversion method based on production confrontation network | |
CN107220657B (en) | A kind of method of high-resolution remote sensing image scene classification towards small data set | |
CN108564119A (en) | A kind of any attitude pedestrian Picture Generation Method | |
CN108304826A (en) | Facial expression recognizing method based on convolutional neural networks | |
CN108921123A (en) | A kind of face identification method based on double data enhancing | |
CN105095870B (en) | Pedestrian based on transfer learning recognition methods again | |
CN109858466A (en) | A kind of face critical point detection method and device based on convolutional neural networks | |
CN106504064A (en) | Clothes classification based on depth convolutional neural networks recommends method and system with collocation | |
CN106503654A (en) | A kind of face emotion identification method based on the sparse autoencoder network of depth | |
CN108171266A (en) | A kind of learning method of multiple target depth convolution production confrontation network model | |
CN109063724A (en) | A kind of enhanced production confrontation network and target sample recognition methods | |
CN105005774A (en) | Face relative relation recognition method based on convolutional neural network and device thereof | |
CN106651915B (en) | The method for tracking target of multi-scale expression based on convolutional neural networks | |
CN110378208B (en) | Behavior identification method based on deep residual error network | |
CN108052884A (en) | A kind of gesture identification method based on improvement residual error neutral net | |
CN106485214A (en) | A kind of eyes based on convolutional neural networks and mouth state identification method | |
CN107423678A (en) | A kind of training method and face identification method of the convolutional neural networks for extracting feature | |
CN106991408A (en) | The generation method and method for detecting human face of a kind of candidate frame generation network | |
CN111476241B (en) | Character clothing conversion method and system | |
CN108985464A (en) | The continuous feature generation method of face for generating confrontation network is maximized based on information | |
CN109598341A (en) | A kind of detection of convolutional neural networks training result and method for visualizing based on genetic algorithm | |
CN110288029A (en) | Image Description Methods based on Tri-LSTMs model | |
CN110175248A (en) | A kind of Research on face image retrieval and device encoded based on deep learning and Hash | |
CN107423727A (en) | Face complex expression recognition methods based on neutral net | |
CN109583942A (en) | A kind of multitask convolutional neural networks customer behavior analysis method based on dense network |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
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 |