CN110309889A - A kind of Old-Yi character symbol restorative procedure of double arbiter GAN - Google Patents
A kind of Old-Yi character symbol restorative procedure of double arbiter GAN Download PDFInfo
- Publication number
- CN110309889A CN110309889A CN201910599249.1A CN201910599249A CN110309889A CN 110309889 A CN110309889 A CN 110309889A CN 201910599249 A CN201910599249 A CN 201910599249A CN 110309889 A CN110309889 A CN 110309889A
- Authority
- CN
- China
- Prior art keywords
- character
- arbiter
- model
- image
- accords
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
- G06V30/10—Character recognition
- G06V30/24—Character recognition characterised by the processing or recognition method
- G06V30/248—Character recognition characterised by the processing or recognition method involving plural approaches, e.g. verification by template match; Resolving confusion among similar patterns, e.g. "O" versus "Q"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
- G06V30/10—Character recognition
- G06V30/28—Character recognition specially adapted to the type of the alphabet, e.g. Latin alphabet
- G06V30/293—Character recognition specially adapted to the type of the alphabet, e.g. Latin alphabet of characters other than Kanji, Hiragana or Katakana
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Multimedia (AREA)
- Character Discrimination (AREA)
- Image Analysis (AREA)
Abstract
The present invention provides the Old-Yi character symbol restorative procedure of double arbiter GAN a kind of, the described method includes: construction is accorded with the DCGAN network model that generator and Yi character symbol arbiter form by Yi character, increase a screening arbiter in the network models, forms a kind of double arbiter GAN models;The probability distribution that Yi character accords with image is obtained from the related hand-written Yi nationality's text of multi-source by double arbiter GAN model realizations, goes to predict that Yi character to be repaired accords with image by acquired probability distribution, Yi character symbol is completed according to forecast image and is repaired.Double arbiter GAN of the invention are on the basis of depth convolution generates confrontation network, increase Yi nationality's text screening arbiter, generation model is accorded with to Yi character by the training in two stages and is iterated optimization, to obtain the complete image of Yi nationality's text to be repaired, the efficiency that Old-Yi character symbol is repaired can be effectively promoted.
Description
Technical field
The invention mainly relates to Ancient Books in Yi Language recovery technique correlative technology field, the Gu of specifically a kind of double arbiter GAN
Yi character accords with restorative procedure.
Background technique
In China, Ancient Books in Yi Language document is increasingly lost and damages seriously, and the researcher due to understanding thoroughly Gu Yiwen lacks,
So that ancient books repair progress is very slow.The development in image text field of artificial intelligence technology is literature of ancient book
Being automatically repaired offer may.
Traditional character image reparation, professional researcher are carried out by language ambience information and perception information, that is, pass through figure
As surrounding pixel and comprehensive standard text in each characteristic element complete.But Yi nationality's text does not have grapholect by computer
Referring to study, there are uncertain problems for hand-written Yi nationality's text.Simultaneous computer is also difficult the language ambience information for having people, for text
Cognition, needs many background knowledges, these are that accumulation is formed the mankind for a long time in certain environment, is difficult systematically to be described
And tissue.
In recent years, deep learning shows in fields such as image, semantic reparation, emotion perception, pattern-recognition and tagsorts
Exciting prospect out shows superior performance especially in terms of image generation.Image based on deep learning, which generates, to be calculated
Method can capture the advanced features of more images relative to traditional generating algorithm based on structure and texture, be usually used in carrying out line
Reason synthesis and image stylization migration.Production confrontation network (GAN) proposed by Goodfellow in 2014 is generated in image
Field achieves initiative progress, and during image generates, production fights network relative to traditional encoding-decoder
For can preferably fitting data, and fast speed, the sample of generation is sharper keen, but there is also deficiencies for this method, in full
It is unstable according to training, model is freely uncontrollable, training collapse the problems such as.2016, Radford et al. proposed that depth convolution generates
Formula fights network DCGAN.DCGAN merges CNN and GAN, by designing unique network structure, so that training is more stable.
It is limited by the development of current Ancient Books in Yi Language recovery technique, how depth convolution production is fought into network DCGAN skill
Art is applied becomes the technical problem that those skilled in the art need to solve in Ancient Books in Yi Language repair process.
Summary of the invention
For the deficiency for solving current technology, the present invention combination prior art provides a kind of double differentiations from practical application
The Old-Yi character of device GAN accords with restorative procedure, and double arbiter GAN of the invention are to generate the basis of confrontation network in depth convolution
On, increase Yi nationality's text screening arbiter, generation model is accorded with to Yi character by the training in two stages and is iterated optimization, from
And the complete image of Yi nationality's text to be repaired is obtained, it can effectively promote the efficiency that Old-Yi character symbol is repaired.
To achieve the above object, technical scheme is as follows:
A kind of Old-Yi character symbol restorative procedure of double arbiter GAN, which is characterized in that the described method includes:
Construction is accorded with the DCGAN network model that generator and Yi character symbol arbiter form by Yi character, in the network model
On the basis of increase a screening arbiter, form a kind of double arbiter GAN models;
The probability that Yi character accords with image is obtained from the related hand-written Yi nationality's text of multi-source by double arbiter GAN model realizations
Distribution goes to predict that Yi character to be repaired accords with image by acquired probability distribution, completes Yi character symbol according to forecast image and repairs
It is multiple.
For double arbiter GAN models, DCGAN network is trained using Yi character symbol first, makes the network
Yi character symbol generator can generate hand-written Yi character symbol image at random;
Then the Yi character of generation symbol is screened by screening arbiter, by the Yi character symbol of generation with it is to be repaired
The difference of Yi character symbol establishes loss function, optimizes to double arbiter GAN models, finally constrains Yi character and accords with generator,
The Yi character symbol infinite approach for generating it Yi character symbol image to be repaired.
When being trained to the DCGAN network, the input that Yi character accords with Maker model is to obey equally distributed 100
The random number of dimensional vector obtains one 64 × 64 matrix by the forward-propagating of model, which is sent into Yi character symbol
Arbiter model, the result and 0 difference, the penalty values loss1 of arbiter model is accorded with for Yi character;
Further, true Yi character is accorded with into image, input Yi character accords with arbiter, obtains a knot by forward-propagating
Fruit, the result with 1 difference be Yi character accord with arbiter model penalty values loss2,
It is optimized finally, according with arbiter model to Yi character by loss1+loss2.
After being optimized to DCGAN network, 100 dimensional vector data are generated by screening arbiter model, which is made
The input of generator is accorded with for Yi character, is carried out forward-propagating, is obtained one 64 × 64 matrix;The matrix is sent into Yi character symbol
Arbiter model, obtain one differentiate as a result, the result with 1 difference be screening arbiter model penalty values loss1;
The difference of this 64 × 64 matrix and Yi character symbol image to be repaired is sought simultaneously, which is loss2;
Screening arbiter model is optimized finally by loss1+loss2.
The Yi character symbol generator and Yi character symbol arbiter correspond to mathematic(al) representation are as follows:
D:y=d (x, θD) (2)
Wherein, G indicates generator;θGFor parameter to be optimized;G () is the nonlinear mapping function that need to be advanced optimized;z
For the input data of g (), i.e., double precision random number between -1~1, length 100;X is that true Yi character accords with data,It is model output as a result, represent 64 × 64 image pixel value size, and numberical range is between -1~1;D indicates to differentiate
Device;θDFor parameter to be optimized;Y is the output of d () as a result, input data is judged as genuine probability, and y ∈ [0,1];
The objective function Equation of arbiter is
The objective function Equation of generator is
The training of model carries out after the objective function for designing formula (4), (5), using gradient descent method to (θG,
θD) parameter alternative optimization.
It is described screening arbiter objective function Equation be
lossz=max { ∑ log (D (z))+abs (A-B) } (6)
In above formula, A is image to be repaired, and B is that the image B, abs (A-B) of generator G (Z) are the pixel of two images
Difference, represent the difference that two images are shown in probability distribution;
Z is optimized by gradient descent method, makes z wirelessly close to desired value, image B is obtained by generator with this
Yi character symbol is repaired.
Beneficial effects of the present invention:
The present invention fights network using DCGAN production to obtain the distribution probability of Yi character symbol, and in the net of DCGAN
Increase a screening arbiter on network structure, form a double arbiters confrontation and generate network, the model realization is from multi-source phase
The probability distribution for obtaining Yi character symbol image in hand-written Yi nationality's text is closed, goes to predict Yi character to be repaired by acquired probability distribution
Image is accorded with, the reparation of Yi character symbol is completed.The experimental results showed that the present invention proposes method for repairing below character incompleteness one third
Multiple rate reaches 77.3%, can effectively promote the efficiency that Old-Yi character symbol is repaired.
Detailed description of the invention
Attached drawing 1 is used double arbiter structure charts;
Attached drawing 2 is that Yi character of the present invention accords with Maker model schematic diagram;
Attached drawing 3 is that Yi character of the present invention accords with arbiter model schematic;
Attached drawing 4 is that Yi character of the present invention accords with arbiter model detailed structure view;
Attached drawing 5 is that Yi character of the present invention accords with Maker model detailed structure view;
Attached drawing 6 is Yi character of the present invention symbol screening arbiter model detailed structure view;
Attached drawing 7 is font sample schematic diagram in the embodiment of the present invention;
Attached drawing 8 is Yi character body library schematic diagram in the embodiment of the present invention;
Attached drawing 9 is Yi nationality's text hand-written data collection individual data figure in the embodiment of the present invention;
Attached drawing 10 is that generator exports image schematic diagram in the embodiment of the present invention;
Attached drawing 11 is the sample schematic diagram in the embodiment of the present invention in data set B;
Attached drawing 12 is Yi nationality's texts and pictures schematic diagram to be repaired in the embodiment of the present invention;
Attached drawing 13 is to export image by Maker model in the embodiment of the present invention;
Attached drawing 14 is penalty values change curve in the embodiment of the present invention;
Attached drawing 15 is that training obtains the raw obtained image of z ' input generator in the embodiment of the present invention;
Attached drawing 16 is the image after repairing in the embodiment of the present invention;
Attached drawing 17 is partial data repairing effect figure in the embodiment of the present invention.
Specific embodiment
With reference to the drawings and specific embodiments, the invention will be further described.It should be understood that these embodiments are merely to illustrate
The present invention rather than limit the scope of the invention.In addition, it should also be understood that, after reading the content taught by the present invention, this field
Technical staff can make various changes or modifications the present invention, and such equivalent forms equally fall within range defined herein.
The present invention proposes one screening arbiter model of increase on depth convolution production confrontation network (DCGAN),
A kind of double arbiter GAN models are formed, Yi nationality's text is repaired using the model.Double arbiter GAN model realizations are from multi-source phase
The probability distribution for obtaining Yi character symbol image in hand-written Yi nationality's text is closed, goes to predict Yi character to be repaired by acquired probability distribution
Image is accorded with, image repair task is completed according to forecast image.Its basic procedure are as follows: accorded with first using Yi character to DCGAN network
It is trained, so that the Yi character of the network is accorded with Maker model and generate hand-written Yi character symbol image at random, then establish one
A screening arbiter model screens the Yi character symbol generated, passes through Yi character symbol and the Yi character to be repaired of generation
The difference of symbol establishes loss function, and optimizes to double arbiter models, can finally constrain Maker model, generates it
Yi character symbol is being not random character, but the Yi character symbol image that infinite approach is to be repaired, completes Yi character symbol with this and repairs.
Double arbiter network structures corresponding to the method for the present invention are as shown in Figure 1, the network is differentiated by a Yi character symbol
Device, a screening arbiter and a Yi character symbol generator composition.The training of network is divided into two stages:
First stage accords with Maker model to Yi character and Yi character symbol arbiter model is trained.Maker model
Input be to obey the random number of equally distributed 100 dimensional vector to obtain one 64 × 64 by the forward-propagating of model
Matrix.The matrix is sent into Yi character and accords with arbiter model, the result and 0 difference, is the penalty values of arbiter model
loss1.(training arbiter when, to it is all generate come as a result, arbiter tries hard to be determined as vacation).It further, will be true
Yi character accord with image, input arbiter, by forward-propagating obtain one as a result, the result with 1 difference be arbiter mould
The penalty values loss2 of type.Finally, being optimized by loss1+loss2 to arbiter model.
Second stage, generator and arbiter all no longer optimize.100 dimensional vector numbers are generated by screening washer model
According to using the data as the input of generator, progress forward-propagating obtains one 64 × 64 matrix.The matrix is sent into Yi nationality
Chinese character arbiter model, obtain one differentiate as a result, the result with 1 difference be screening model penalty values loss1 (generate
Out as a result, being desirable to arbiter it can be determined that true).This 64 × 64 matrix and Yi character to be repaired are asked simultaneously
The difference of image is accorded with, which is loss2, is optimized finally by loss1+loss2 to screening model.
By above-mentioned two step, the instruction of double arbiter productions confrontation network structure for Yi character symbol image is completed
Practice, obtains a stable model.Yi character accords with arbiter and Yi character symbol generator schematic diagram is as shown in Figures 2 and 3.
The effect of Yi character symbol Maker model is so that generating distribution of " puppet " image as far as possible with " nature " image
Unanimously;The effect of arbiter model is that correct judgement is made between generation " puppet " image and " nature " image, i.e., two classification
Device.The method for realizing whole network training is that the two networks is allowed to vie each other, and ultimately generates model and passes through study " nature "
The intrinsic propesties of data generate new data similar with " nature " sample to depict the distribution probability of " nature " sample,
Mathematical expression are as follows:
D:y=d (x, θD) (2)
Wherein, G indicates generator;θGFor parameter to be optimized;G () is the nonlinear mapping function that need to be advanced optimized;z
For the input data of g (), i.e., double precision random number between -1~1, length 100;X is that true Yi character accords with data,It is model output as a result, represent 64 × 64 image pixel value size, and numberical range is between -1~1;D indicates to differentiate
Device;θDFor parameter to be optimized;Y is the output of d () as a result, input data is judged as genuine probability, and y ∈ [0,1].
Arbiter modelling is that natural data is judged as to genuine probability and generation data are determined as pseudo- probability to want
Height, the objective function Equation of arbiter are as follows:
To log (d (x, θ in formula (4)D)) penalty values to take negative physical interpretation be that x is judged as genuine uncertainty
The smaller the better, optimum state is d (x)=1;And it is rightPenalty values take negative physical interpretation for willIt is judged as that pseudo- uncertainty is the smaller the better, i.e., willIt is judged as that pseudo- probability is the bigger the better;By the uncertain of all judgements
Property is summed, and entropy is just obtained, according to entropy to the parameter θ of modelDIt optimizes.
For generator when arbiter model is fixed, the distribution character for generating data is as much as possible literary with sample Yi nationality
Character data is consistent, i.e. maximization objective function as shown in formula (5):
The training of model (formula (4) and (5)) after designing loss function carries out, using gradient descent method to (θG,
θD) parameter alternative optimization, it is finally reached Nash Equilibrium.
Based on above-mentioned modelling, can be described as the reparation problem of Yi character symbol defect: assuming that have one come
The image B of self-generator G (Z), it gives the reconstruct to lack part, allows generator to generate image B's if can find
It is distributed Z, so that it may repair to lack part according to image B.The effect of screening arbiter of the invention is exactly by Yi character
The range automatically generated is accorded with to be limited on the specific Yi character symbol for needing to repair.Assuming that image to be repaired is A, only have when initial
One random value z, generator G (z) image B ' generated may have no association with image A or correlation is little.Here with two
The difference abs (A-B ') of the pixel value of width image, the difference of probability distribution between Lai Daibiao two images.But only in this, as penalty values
Z is optimized not enough, it is also necessary to which limiting image B ' must be a Yi character symbol, so further by arbiter process decision chart
As A is one of the optimization aim that the probability that Yi character accords with also is used as z.Since it is desirable that arbiter determines that image B ' accords with for Yi character,
It is desired for very, so penalty values are log (D (z)), can be obtained by the function of optimization z, such as formula in conjunction with two penalty values
(6) shown in.
lossz=max { ∑ log (D (z))+abs (A-B) } (6)
Due to losszFor convex function, therefore z is optimized by gradient descent method, it will be able to so that z is infinitely close to
Desired Z obtains image B by generator with this and repairs to Yi character symbol.
Screening arbiter is most important part in double arbiter models.It, can by according with the training of generator to Yi character
It is random for being accorded with the Yi character of the mapping relations of the random number of 100 dimensional vectors of acquisition and Yi character symbol, but generator output,
It needs to select one to accord with closest prediction data with Yi character to be repaired to complete to repair by screening arbiter.Due to
Wish that the output of model G can be accorded with ad infinitum close to Yi character to be repaired, therefore is accorded with the output and Yi character to be repaired of G
Penalty values loss1 of the difference of image corresponding points as screening model;Also wish that the output of model G can be judged as by model D
Be Yi character symbol, therefore model D output with 1 difference as screen arbiter penalty values loss2.In conjunction with two penalty values
It can be obtained by the majorized function as shown in formula (6).The function is solved with gradient descent method (model D and G without
Optimization), constantly optimize F1 layers of parameter, screening arbiter is enabled to obtain a data z.Data z is positive by model G
An image with hand-written Yi character symbol infinite approach to be repaired can be exported by propagating.
Have for model structure of the invention as described below:
Yi character accords with arbiter model:
The input that Yi character accords with arbiter model is character picture, judges whether the image is Yi character by arbiter
Symbol.Model includes 1 input layer, 3 convolutional layers and 1 output layer.Its model structure is as shown in Figure 4.
Yi character accords with arbiter and is made of 4 layers of CNN (not including input layer), and INPUT is input layer in Fig. 6, is input
Initial data, the data accord with image from Yi character, and size is 64 × 64 pixels, because Yi character symbol is gray level image,
Therefore 3 channels of image are modified to single channel.OUT is output layer, only 1 node.Convolutional layer indicates with C, details
As shown in table 1.
1 arbiter model parameter table of table
Convolution kernel number | Convolution kernel size | Step-length | Characteristic pattern size | Number of parameters | |
C1 layers (convolutional layer) | 128 | 5×5 | 2 | 32×32 | 3328 |
C2 layers (convolutional layer) | 256 | 5×5 | 2 | 16×16 | 819456 |
C3 layers (convolutional layer) | 512 | 5×5 | 2 | 8×8 | 3277312 |
C4 layers (convolutional layer) | 1024 | 5×5 | 2 | 4×4 | 13108224 |
OUTPUT layers (output layer) | 1×1 | 16385 |
Then the OUTPUT layers of dot product calculated between input vector and weight vectors are transmitted along with a biasing
Result is exported to sigmoid function.Whether 1 node on behalf of this output layer is Yi character symbol, if the value of node is 1,
Indicate that the result of Network Recognition is considered that Yi character accords with, 0 opposite.
Yi character accords with Maker model:
Yi character Maker model be made of 4 layers of CNN (do not include input layer), and INPUT is input layer in Fig. 5, is inputted and is
Obey the random number of equally distributed 100 dimensional vector.OUT is output layer, the matrix that output data is 64 × 64 × 1, it is desirable to logical
It crosses and trains, which can indicate that Yi character accords with image, and model detailed maps are as shown in Figure 5.Warp lamination indicates with DC, entirely
Articulamentum indicates that details are as shown in table 2 with F.
2 Maker model parameter list of table
Convolution kernel number | Convolution kernel size | Step-length | Characteristic pattern size | Number of parameters | |
F1 layers (full articulamentum) | 1654784 | ||||
DC1 layers (convolutional layer) | 512 | 5×5 | 2 | 8×8 | 13107712 |
DC2 layers (convolutional layer) | 256 | 5×5 | 2 | 16×16 | 3277056 |
DC3 layers (convolutional layer) | 128 | 5×5 | 2 | 32×32 | 819328 |
DC4 layers (convolutional layer) | 1 | 5×5 | 2 | 64*64 | 3201 |
Screen arbiter model:
Screening arbiter model is made of input layer, 1 full articulamentum F1 and output layer, and G indicates that Yi character accords with generator
Model, D indicate that Yi character accords with arbiter model.INPUT is input layer in figure, is inputted to obey equally distributed 100 dimensional vector
Random number.OUT is the output combination that output layer output data is model G and D.Model structure is as shown in Figure 6.
Wherein, F1 layers (full articulamentum):
The number of every layer of neuron weight of full articulamentum neural network, Param=(input data dimension+1) × nerve
First number adds 1 to be because each neuron has a bias.Input data dimension is 100, which has used 1*64*64
A node, so number of parameters is (100+1) × 1 × 100=10100.
G (Yi nationality's text Maker model):
Input of F1 layers of the output as G obtains one 64 × 64 matrix through forward-propagating.
D (Yi nationality's text arbiter model):
Input of the output of model G as D obtains 1 dimension data through forward-propagating.
Embodiment:
Restorative procedure is accorded with based on Old-Yi character provided by the present invention, the samples sources of the present embodiment are in " the west of 370,000 words
Southern Yi nationality's will " in 2142 common Old-Yi characters symbols choosing, copied by Yi nationality, distributed over Yunnan, Sichuan and Guizhou academics and students, provided 1200 parts and adopted
Collect table, wherein 800 parts, soft 200 parts of style acquisition tables of Yi nationality's text roman acquisition tables, 200 parts of hard-tipped pen style acquisition tables, such as Fig. 7 institute
Show, 151200 font samples have been obtained.Meanwhile for the ease of post-processing analysis, corresponding fontlib is devised (as schemed
Shown in 8) and ancient Yi nationality text input method.
Sample in Yi character body library is converted to the picture of 64 × 64 pixels composition, each pixel is with 0~255
Gray value indicates, is divided into training set A and test set B in the ratio of each text 7: 3, part sample is as shown in Figure 9.
Experiment is trained model using training set A, and training is using 32 samples as small lot number every time.Training process
Are as follows:
(1) the Yi nationality's text data randomly choosed in 32 training set A pass through Yi nationality as the input of Yi nationality's text discrimination model
The forward-propagating of literary discrimination model, obtain 32 differentiate as a result, i.e. input data whether be Yi nationality's text probability, by this 32 generally
Rate does mean square deviation with 1 respectively and obtains penalty values d1;
(2) 32 are randomly generated and obeys equally distributed 100 dimensional vector data, generates the defeated of model as Yi nationality's text
Enter, forward direction is passed through using this data as the input of Yi nationality's text discrimination model by the matrix that forward-propagating obtains 32 64*64*1
Propagation also obtains 32 differentiation results D32, D32 and does mean square deviation with 0 respectively obtaining penalty values d2;
(3) d1 and d2 are averaged to obtain the penalty values loss of Yi nationality's text arbiterd, to lossdYi nationality's text is sentenced with Adam algorithm
The parameter of other model optimizes.
(4) the result D32 that 32 data of Yi character symbol generator output are obtained by Yi nationality's text arbiter, as Yi nationality's text
Also the penalty values of Yi nationality's text generator are used as while the penalty values d2 of arbiter.Because being contemplated to be all lifes to Yi nationality's text generator
At the data gone out, Yi nationality's text arbiter is all determined as very, so D32 does mean square deviation with 1 respectively and obtains the damage of Yi nationality's text generator
Mistake value lossg, to lossgModel is generated to Yi nationality's text with Adam algorithm to optimize.
By constantly repeating above-mentioned optimization process, Yi nationality's text arbiter and Yi nationality's text Maker model are trained.It is instructing
During white silk, the penalty values of every 100 records Yi nationality's text arbiter and Yi nationality's text generator, after training by 9300 times, Yi nationality
Literary Maker model and Yi nationality's text arbiter model constantly carry out game in the training process.When the penalty values of Yi nationality's text arbiter model
When reduction, indicate that Yi nationality's text that Yi nationality's text arbiter identifies that generator generates is that false probability increases.And Yi nationality's text Maker model
When penalty values reduce, indicate that Yi nationality's text that Yi nationality's text generator generates is judged as true by Yi nationality's text arbiter maximum probability.Therefore two moulds
The penalty values of type alternately rise, this also illustrates that two models in continuous game, alternately optimize model parameter.
Input of the vector data of 1000 100 dimensions as generator is randomly generated, takes second place in 2 model trainings 9300
Afterwards, 1000 Yi character symbols are obtained, Jing Yiwen expert judgments generate result and the form of true Yi character symbol is close.Partial data
As shown in Figure 10.
In character repairing phase, Yi nationality's text Maker model and Yi nationality's text arbiter model are no longer optimized, but to Yi nationality's text
Screening arbiter is trained.A sample is chosen from Yi nationality text data set B as object is repaired, as shown in figure 11.Because should
Sample is complete Yi nationality's text, has no missing, is verification the verifying results, using one 20 × 20 all 1's matrix to the middle section of image
It is covered, as shown in figure 12.The input of Yi nationality's text generator be no longer by random signal, but Yi nationality's text screening arbiter output
Z accords with the output G (z) of generator by the available Yi character of forward-propagating, as shown in figure 13.
The G (z) of generator output is a random Yi character symbol, with Yi character symbol and onrelevant to be repaired.By G
(z) it is compared with Yi character symbol to be repaired, penalty values loss is obtained according to formula (6)z.Here losszDeclined with gradient
Method optimization, to screening arbiter be trained, it is expected that screen arbiter optimization after can export a z ', make G (z ') with to
The Yi character of reparation accords with infinite approach, but possesses incomplete part font, i.e., consistent with Figure 11.Learning rate, which is arranged, is
0.001, screening arbiter model is trained, is trained 2000 times altogether.The change curve of its penalty values is such as in the training process
Shown in Figure 14, it can be seen that after training 750 times, penalty values gradually drop within the scope of one, that is, indicate to be currently generated
Character and character to be repaired between difference also drop to a certain range.After 2000 suboptimization, screening arbiter is generated
Z ', by Yi character accord with generator in forward-propagating obtain G (z '), as shown in figure 15.It makes discovery from observation, in completion pair
After 2000 training of z ', a Yi character symbol closely similar with Yi nationality's text to be repaired is obtained.It will be lacked in Yi nationality's text to be repaired
Part, be filled with G (z '), the image after being repaired, as shown in figure 16.Horizontal axis represents frequency of training, and vertical pivot represents
Penalty values size, unit are px (pixel unit).
Above-mentioned experiment repeats 1000 times, extracted from Yi nationality text test set B at random each time a sample be fabricated to it is to be repaired
Yi nationality text, then repaired by methods herein, to Yi character symbol repair rate be 77.3%, as shown in table 3.
3 Old-Yi character of table accords with reparation ratio
Number | Accounting | |
It repairs completely | 523 | 52% |
It repairs part | 246 | 25% |
It does not complete and repairs | 231 | 23% |
It is total | 1000 | 100% |
Figure 17 is the partial data in 1000 experiments, from left to right respectively Yi nationality's text original sample image, generate to
Yi nationality's texts and pictures picture of reparation, generator generate the image come, the effect picture after repairing.
Claims (6)
1. the Old-Yi character of double arbiter GAN a kind of accords with restorative procedure, which is characterized in that the described method includes:
Construction is accorded with the DCGAN network model that generator and Yi character symbol arbiter form by Yi character, in the network models
One screening arbiter of upper increase, forms a kind of double arbiter GAN models;
The probability point that Yi character accords with image is obtained from the related hand-written Yi nationality's text of multi-source by double arbiter GAN model realizations
Cloth goes to predict that Yi character to be repaired accords with image by acquired probability distribution, completes Yi character symbol according to forecast image and repairs.
2. a kind of Old-Yi character of double arbiter GAN as described in claim 1 accords with restorative procedure, which is characterized in that for institute
Double arbiter GAN models are stated, DCGAN network is trained using Yi character symbol first, the Yi character of the network is accorded with and generates
Device can generate hand-written Yi character symbol image at random;
Then it is screened by Yi character symbol of the screening arbiter to generation, passes through Yi character symbol and the Yi nationality's text to be repaired of generation
The difference of character establishes loss function, optimizes to double arbiter GAN models, finally constrains Yi character and accords with generator, makes it
The Yi character that the Yi character symbol infinite approach of generation is to be repaired accords with image.
3. a kind of Old-Yi character of double arbiter GAN as claimed in claim 2 accords with restorative procedure, which is characterized in that described
When DCGAN network is trained, the input of Yi character symbol Maker model is the random of equally distributed 100 dimensional vector of obedience
Number, by the forward-propagating of model, obtains one 64 × 64 matrix, which is sent into Yi character and accords with arbiter model, should
As a result the difference with 0 accords with the penalty values loss1 of arbiter model for Yi character;
Further, by true Yi character accord with image, input Yi character accord with arbiter, by forward-propagating obtain one as a result,
The result with 1 difference be Yi character accord with arbiter model penalty values loss2,
It is optimized finally, according with arbiter model to Yi character by loss1+loss2.
4. a kind of Old-Yi character of double arbiter GAN as claimed in claim 3 accords with restorative procedure, which is characterized in that DCGAN
After network optimizes, 100 dimensional vector data are generated by screening arbiter model, accord with generator for the data as Yi character
Input, carry out forward-propagating, obtain one 64 × 64 matrix;The matrix is sent into Yi character and accords with arbiter model, is obtained
One differentiate as a result, the result with 1 difference be screening arbiter model penalty values loss1;
The difference of this 64 × 64 matrix and Yi character symbol image to be repaired is sought simultaneously, which is loss2;
Screening arbiter model is optimized finally by loss1+loss2.
5. a kind of Old-Yi character of double arbiter GAN as claimed in claim 4 accords with restorative procedure, which is characterized in that the Yi nationality
Chinese character generator and Yi character symbol arbiter correspond to mathematic(al) representation are as follows:
D:y=d (x, θD) (2)
Wherein, G indicates generator;θGFor parameter to be optimized;G () is the nonlinear mapping function that need to be advanced optimized;Z is g
The input data of (), i.e., the double precision random number between -1~1, length 100;X is that true Yi character accords with data,For
Model output as a result, represent 64 × 64 image pixel value size, and numberical range is between -1~1;D indicates arbiter;
θDFor parameter to be optimized;Y is the output of d () as a result, input data is judged as genuine probability, and y ∈ [0,1];
The objective function Equation of arbiter is
The objective function Equation of generator is
The training of model carries out after the objective function for designing formula (4), (5), using gradient descent method to (θG, θD) ginseng
Number alternative optimization.
6. a kind of Old-Yi character of double arbiter GAN as claimed in claim 5 accords with restorative procedure, which is characterized in that the sieve
The objective function Equation for selecting arbiter is
lossz=max { ∑ log (D (z))+abs (A-B) } (6)
In above formula, A is image to be repaired, and B is the difference that the image B, abs (A-B) of generator G (Z) are the pixel of two images,
Represent the difference that two images are shown in probability distribution;
Z is optimized by gradient descent method, makes z wirelessly close to desired value, image B is obtained to Yi nationality by generator with this
Chinese character is repaired.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910599249.1A CN110309889A (en) | 2019-07-04 | 2019-07-04 | A kind of Old-Yi character symbol restorative procedure of double arbiter GAN |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910599249.1A CN110309889A (en) | 2019-07-04 | 2019-07-04 | A kind of Old-Yi character symbol restorative procedure of double arbiter GAN |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110309889A true CN110309889A (en) | 2019-10-08 |
Family
ID=68079137
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910599249.1A Pending CN110309889A (en) | 2019-07-04 | 2019-07-04 | A kind of Old-Yi character symbol restorative procedure of double arbiter GAN |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110309889A (en) |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110852326A (en) * | 2019-11-06 | 2020-02-28 | 贵州工程应用技术学院 | Handwriting layout analysis and multi-style ancient book background fusion method |
CN111008604A (en) * | 2019-12-09 | 2020-04-14 | 上海眼控科技股份有限公司 | Prediction image acquisition method and device, computer equipment and storage medium |
CN111260620A (en) * | 2020-01-14 | 2020-06-09 | 三峡大学 | Image anomaly detection method and device and electronic equipment |
CN112001427A (en) * | 2020-08-04 | 2020-11-27 | 中国科学院信息工程研究所 | Image conversion method and device based on analogy learning |
CN112307343A (en) * | 2020-11-05 | 2021-02-02 | 重庆邮电大学 | Cross-E-book city user alignment method based on double-layer iterative compensation and full-face representation |
CN112381725A (en) * | 2020-10-16 | 2021-02-19 | 广东工业大学 | Image restoration method and device based on deep convolution countermeasure generation network |
CN114494003A (en) * | 2022-04-01 | 2022-05-13 | 华南理工大学 | Ancient character generation method combining shape transformation and texture transformation |
CN115797216A (en) * | 2022-12-14 | 2023-03-14 | 齐鲁工业大学 | Inscription character restoration model and restoration method based on self-coding network |
Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107563385A (en) * | 2017-09-02 | 2018-01-09 | 西安电子科技大学 | License plate character recognition method based on depth convolution production confrontation network |
CN107945118A (en) * | 2017-10-30 | 2018-04-20 | 南京邮电大学 | A kind of facial image restorative procedure based on production confrontation network |
CN108460720A (en) * | 2018-02-01 | 2018-08-28 | 华南理工大学 | A method of changing image style based on confrontation network model is generated |
CN108460760A (en) * | 2018-03-06 | 2018-08-28 | 陕西师范大学 | A kind of Bridge Crack image discriminating restorative procedure fighting network based on production |
CN108520503A (en) * | 2018-04-13 | 2018-09-11 | 湘潭大学 | A method of based on self-encoding encoder and generating confrontation network restoration face Incomplete image |
CN109191402A (en) * | 2018-09-03 | 2019-01-11 | 武汉大学 | The image repair method and system of neural network are generated based on confrontation |
CN109377448A (en) * | 2018-05-20 | 2019-02-22 | 北京工业大学 | A kind of facial image restorative procedure based on generation confrontation network |
CN109785258A (en) * | 2019-01-10 | 2019-05-21 | 华南理工大学 | A kind of facial image restorative procedure generating confrontation network based on more arbiters |
-
2019
- 2019-07-04 CN CN201910599249.1A patent/CN110309889A/en active Pending
Patent Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107563385A (en) * | 2017-09-02 | 2018-01-09 | 西安电子科技大学 | License plate character recognition method based on depth convolution production confrontation network |
CN107945118A (en) * | 2017-10-30 | 2018-04-20 | 南京邮电大学 | A kind of facial image restorative procedure based on production confrontation network |
CN108460720A (en) * | 2018-02-01 | 2018-08-28 | 华南理工大学 | A method of changing image style based on confrontation network model is generated |
CN108460760A (en) * | 2018-03-06 | 2018-08-28 | 陕西师范大学 | A kind of Bridge Crack image discriminating restorative procedure fighting network based on production |
CN108520503A (en) * | 2018-04-13 | 2018-09-11 | 湘潭大学 | A method of based on self-encoding encoder and generating confrontation network restoration face Incomplete image |
CN109377448A (en) * | 2018-05-20 | 2019-02-22 | 北京工业大学 | A kind of facial image restorative procedure based on generation confrontation network |
CN109191402A (en) * | 2018-09-03 | 2019-01-11 | 武汉大学 | The image repair method and system of neural network are generated based on confrontation |
CN109785258A (en) * | 2019-01-10 | 2019-05-21 | 华南理工大学 | A kind of facial image restorative procedure generating confrontation network based on more arbiters |
Non-Patent Citations (2)
Title |
---|
唐骞: "基于对抗学习的图像修复", 《计算机产品与流通》, 15 January 2019 (2019-01-15), pages 210 * |
李克文 等: "多尺度生成式对抗网络图像修复算法", 《计算机科学与探索》, vol. 14, no. 1, 17 June 2019 (2019-06-17), pages 159 - 170 * |
Cited By (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110852326A (en) * | 2019-11-06 | 2020-02-28 | 贵州工程应用技术学院 | Handwriting layout analysis and multi-style ancient book background fusion method |
CN110852326B (en) * | 2019-11-06 | 2022-11-04 | 贵州工程应用技术学院 | Handwriting layout analysis and multi-style ancient book background fusion method |
CN111008604A (en) * | 2019-12-09 | 2020-04-14 | 上海眼控科技股份有限公司 | Prediction image acquisition method and device, computer equipment and storage medium |
CN111260620A (en) * | 2020-01-14 | 2020-06-09 | 三峡大学 | Image anomaly detection method and device and electronic equipment |
CN111260620B (en) * | 2020-01-14 | 2023-08-18 | 三峡大学 | Image anomaly detection method and device and electronic equipment |
CN112001427A (en) * | 2020-08-04 | 2020-11-27 | 中国科学院信息工程研究所 | Image conversion method and device based on analogy learning |
CN112381725B (en) * | 2020-10-16 | 2024-02-02 | 广东工业大学 | Image restoration method and device based on depth convolution countermeasure generation network |
CN112381725A (en) * | 2020-10-16 | 2021-02-19 | 广东工业大学 | Image restoration method and device based on deep convolution countermeasure generation network |
CN112307343B (en) * | 2020-11-05 | 2023-04-07 | 重庆邮电大学 | Cross-E-book city user alignment method based on double-layer iterative compensation and full-face representation |
CN112307343A (en) * | 2020-11-05 | 2021-02-02 | 重庆邮电大学 | Cross-E-book city user alignment method based on double-layer iterative compensation and full-face representation |
CN114494003A (en) * | 2022-04-01 | 2022-05-13 | 华南理工大学 | Ancient character generation method combining shape transformation and texture transformation |
CN115797216A (en) * | 2022-12-14 | 2023-03-14 | 齐鲁工业大学 | Inscription character restoration model and restoration method based on self-coding network |
CN115797216B (en) * | 2022-12-14 | 2024-05-24 | 齐鲁工业大学 | Self-coding network-based steganography character restoration model and restoration method |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110309889A (en) | A kind of Old-Yi character symbol restorative procedure of double arbiter GAN | |
Meng et al. | Adavit: Adaptive vision transformers for efficient image recognition | |
CN110378985B (en) | Animation drawing auxiliary creation method based on GAN | |
CN109992779B (en) | Emotion analysis method, device, equipment and storage medium based on CNN | |
CN104463209B (en) | Method for recognizing digital code on PCB based on BP neural network | |
CN109614979A (en) | A kind of data augmentation method and image classification method based on selection with generation | |
CN109829541A (en) | Deep neural network incremental training method and system based on learning automaton | |
CN102314614B (en) | Image semantics classification method based on class-shared multiple kernel learning (MKL) | |
CN109389037A (en) | A kind of sensibility classification method based on depth forest and transfer learning | |
Jeon et al. | T-gd: Transferable gan-generated images detection framework | |
CN113128620B (en) | Semi-supervised domain self-adaptive picture classification method based on hierarchical relationship | |
CN108596243B (en) | Eye movement gaze prediction method based on hierarchical gaze view and conditional random field | |
US20150147728A1 (en) | Self Organizing Maps (SOMS) for Organizing, Categorizing, Browsing and/or Grading Large Collections of Assignments for Massive Online Education Systems | |
CN113807176B (en) | Small sample video behavior recognition method based on multi-knowledge fusion | |
Zhai et al. | Asian female facial beauty prediction using deep neural networks via transfer learning and multi-channel feature fusion | |
CN109543749A (en) | Drawing sentiment analysis method based on deep learning | |
CN114283083B (en) | Aesthetic enhancement method of scene generation model based on decoupling representation | |
Gao et al. | Wallpaper texture generation and style transfer based on multi-label semantics | |
Niu et al. | Exploring causes for the dropout on massive open online courses | |
CN107169958A (en) | Machine learning, background suppress with perceiving the vision significance detection method that positive feedback is combined | |
CN110472746A (en) | A kind of coding prediction technique and system based on artificial intelligence | |
CN110378882A (en) | A kind of Chinese medicine tongue nature method for sorting colors of multi-layer depth characteristic fusion | |
Martins et al. | Evotype: from shapes to glyphs | |
CN113689514A (en) | Theme-oriented image scene graph generation method | |
CN113408418A (en) | Calligraphy font and character content synchronous identification method and system |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20191008 |
|
RJ01 | Rejection of invention patent application after publication |