CN116935406A - Chinese character style generation method with natural writing property - Google Patents
Chinese character style generation method with natural writing property Download PDFInfo
- Publication number
- CN116935406A CN116935406A CN202310883227.4A CN202310883227A CN116935406A CN 116935406 A CN116935406 A CN 116935406A CN 202310883227 A CN202310883227 A CN 202310883227A CN 116935406 A CN116935406 A CN 116935406A
- Authority
- CN
- China
- Prior art keywords
- style
- chinese character
- skeleton
- data set
- training
- 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
- 238000000034 method Methods 0.000 title claims abstract description 33
- 238000012549 training Methods 0.000 claims abstract description 28
- 238000013461 design Methods 0.000 claims abstract description 12
- 230000000694 effects Effects 0.000 claims abstract description 7
- 238000005457 optimization Methods 0.000 claims abstract description 5
- 239000011159 matrix material Substances 0.000 claims description 12
- 239000013598 vector Substances 0.000 claims description 10
- 230000008569 process Effects 0.000 claims description 9
- 230000009467 reduction Effects 0.000 claims description 4
- 238000009877 rendering Methods 0.000 claims description 4
- 238000013527 convolutional neural network Methods 0.000 claims description 2
- 230000004044 response Effects 0.000 claims description 2
- 238000013507 mapping Methods 0.000 abstract description 5
- 238000010276 construction Methods 0.000 abstract description 3
- 238000012015 optical character recognition Methods 0.000 abstract description 3
- 238000013508 migration Methods 0.000 description 21
- 230000005012 migration Effects 0.000 description 21
- 230000006870 function Effects 0.000 description 17
- 230000000007 visual effect Effects 0.000 description 3
- 238000013459 approach Methods 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 2
- 230000008901 benefit Effects 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 239000000284 extract Substances 0.000 description 2
- 230000007246 mechanism Effects 0.000 description 2
- 230000004075 alteration Effects 0.000 description 1
- 239000000969 carrier Substances 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 125000004122 cyclic group Chemical group 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000009792 diffusion process Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Classifications
-
- 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
- G06V30/19—Recognition using electronic means
- G06V30/191—Design or setup of recognition systems or techniques; Extraction of features in feature space; Clustering techniques; Blind source separation
- G06V30/19173—Classification techniques
-
- 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
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- 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
- G06N3/0475—Generative 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
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- 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
- G06N3/094—Adversarial learning
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Biophysics (AREA)
- Evolutionary Computation (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Software Systems (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Multimedia (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Controls And Circuits For Display Device (AREA)
Abstract
The application provides a Chinese character style generation method with natural writing property, which comprises the following steps: s1, constructing a data set; s2, generating structural design and training of a model; s3, inputting a model after training optimization to generate a single-line style Chinese character with natural writability. The application realizes the mapping from the Chinese character skeleton to the style Chinese character through the training of the countermeasure generation network by constructing the full stroke connection Chinese character skeleton and style Chinese character data set of the common Chinese character sum 8876, realizes the real meaning of 'stroke break meaning connection and word break gas connection', has wide application value in the construction of the data set in the art design field and the optical character recognition field, and has great promotion effect in the natural writing Chinese character style generation field.
Description
Technical Field
The application relates to a Chinese character style generation method with natural writing property.
Background
The Chinese characters are used as carriers of Chinese civilization, and unique artistic aesthetic feeling is created by various writing styles. In contemporary visual design, fonts are the basic design elements and are also the necessary expression tools in the design. For a font user, a proper style of font can enable the whole design work to fully embody the concept of a designer, so that the uniformity of the style is maintained.
With the development of model generation in the field of computer vision, style migration and model generation gradually become a great means for replacing the traditional design mode. The model extracts style characteristics from the given style font data set and completes style migration to other not designed Chinese characters. However, the existing Chinese character generating method is based on the parts of the reference Chinese character, the generating result of the same Chinese character parts is fixed and rigid like letter printing, and the correspondence between strokes and parts is lacking. Secondly, the application object of the Chinese character component is a single character, so that the generated Chinese character is stiff like type printing, lacks the correspondence between the characters, and cannot be compared with the celebrity. In addition, as the latest picture generation model, the diffusion model cannot be well fitted with the Chinese character skeleton, and cannot be applied to the field of Chinese character generation.
The Chinese patent with the application number of CN202010333081.2 discloses a Chinese character style migration method and a Chinese character style migration system based on a multitasking countermeasure learning network, wherein the Chinese character style migration method and the Chinese character style migration system comprise the following steps: acquiring a Chinese character image to be subjected to style migration; inputting the Chinese character image to be subjected to style migration into a trained multi-task countermeasure learning network; the trained multitask countermeasure learning network outputs the multiple font images after style migration. A unified encoder is used for learning a general visual mode of a reference font which is important for all target fonts, so that characteristic level information is transmitted to the greatest extent across tasks, task-specific characteristics are reserved in respective network channels, the multi-task training strategy enables Chinese character style migration network training to be more stable, the generalization capability of the network is improved, the generated font style is more consistent with the target fonts, and stroke boundaries are clear.
The Chinese patent with the application number of CN202011564611.0 proposes a Chinese font style migration method, which is based on an original generation type countermeasure network, and adds two auxiliary networks on a generator formed by circularly generating the countermeasure network, wherein firstly, the two auxiliary networks simultaneously extract structural features of an original image and a picture generated by the generator through a Chinese character classification recognition residual network, and secondly, the style features generated by the generator are extracted by utilizing a style encoder so as to ensure style consistency.
Although the prior method well constrains style characteristics during style migration and Chinese character skeletons of Chinese characters to be migrated, the prior method still has the following defects: 1. all style migration results are built on a single word, namely the whole end-to-end generation process is not only that the strokes generated in the single word are fixed, the continuous stroke structures among strokes in the word and between words are not considered, the generation result is single, and the change is absent; 2. the reference Chinese characters are required to be standard fonts such as Song body and the like contained in a training set, and corresponding forward books are required to be provided in actual application scenes, so that the robustness is lacked, and the limitation is high.
Disclosure of Invention
In order to solve the defect that the results of the conventional style migration method are rigid and fixed, realize the natural writing performance of generating Chinese characters by style migration and achieve the effects of' continuous meaning of pen and continuous meaning of word and word, the application provides a Chinese character style generation method with natural writing performance, which has wide application value in the construction of data sets in the field of artistic design and the field of optical character recognition.
The technical scheme is as follows:
a Chinese character style generation method with natural writing property comprises the following steps: processing 8876 regular script Chinese character Bezier curves in a data preprocessing stage and fully connecting strokes to construct a data set of Chinese character skeleton lines and corresponding style Chinese character pictures in pairs; and in the style migration stage, training a neural network based on a cyclic countermeasure generation network, adding a attention mechanism, coding the marked continuous information, and splicing with an embedded layer code. Meanwhile, the continuous pen consistency loss function for measuring the natural writing property is creatively designed, so that the generator focuses on the generation of the natural writing property. In summary, the application provides a Chinese character style generation method with natural writing property, which realizes the mapping process of full-connection skeleton Chinese characters to style Chinese characters, thereby generating single-line style Chinese characters with natural writing property.
A Chinese character style generation method with natural writing property comprises the following steps:
s1, constructing a data set;
s2, generating structural design and training of a model;
s3, inputting a model after training optimization to generate a single-line style Chinese character with natural writability.
The technical scheme of the application has the advantages that:
according to the technical scheme, a generating type countermeasure network with a focusing attention mechanism for connecting strokes is provided by constructing a continuous stroke data set of common Chinese characters, and a creatively provided continuous stroke consistency loss function is assisted, so that the mapping process from the continuous stroke Chinese character skeleton data set to style Chinese characters is trained, and style migration from Chinese character skeletons to any style fonts and having natural writing performance is realized. And by utilizing a refinement algorithm of the extracted skeleton, style migration among any style of Chinese characters can be realized. Compared with the prior style migration method, the method has remarkable improvement in the aspect of natural writing property.
The application realizes the mapping from the Chinese character skeleton to the style Chinese character through the training countermeasure generation network by constructing the full stroke connection Chinese character skeleton and style Chinese character data set of 8876 of common and secondary common Chinese characters, and realizes the 'stroke break meaning connection and word break gas connection' in the real sense, thereby having wide application value in the construction of the data set in the art design field and the optical character recognition field. The application has great promotion effect in the field of Chinese character style generation with natural writing property.
Drawings
FIG. 1 is a flow chart of the application;
FIG. 2 is a pair data set of the "skeleton of Chinese character-Zhao Meng line book" of the embodiment;
FIG. 3 is an embodiment model architecture;
FIG. 4 is a representation of the sushi style font generation in an embodiment (paired words, left to generate results, right to target);
FIG. 5 is a Shu Tong style font generation (left to generate results, right to target) in an embodiment;
fig. 6 is Zhao Meng style font generation (left as generated result, right as target) in an embodiment.
Detailed Description
The technical scheme provided by the application is further described below with reference to specific embodiments and attached drawings. The advantages and features of the present application will become more apparent in conjunction with the following description.
The overall flow is shown in fig. 1.
S1, constructing a data set:
the application constructs the full-stroke Chinese character skeleton data set by editing the closed curve in the Bezier curve and then rendering.
The Bezier curves are divided into first-order, second-order and third-order curves according to the orders, and are respectively marked as:
m (a, b) L (c, d) (a first order bezier curve from point (a, b) to point (c, d);
m (a, b) C (x, y, C, d) (a second order bezier curve starting from point (a, b) to point (C, d) and controlling the point coordinates to (x, y);
M(a，b)Q(x 1 ,y 1 ,x 2 ,y 2 c, d) (from point (a, b) to point (c, d), and the control point coordinates are (x) 1 ,y 1 ) And (x) 2 ,y 2 ) A third-order bezier curve).
In the data set of the application, the Chinese character skeleton consists of a plurality of lines, wherein the number of the lines is the stroke number of the Chinese character. Each line is formed by connecting the three curves.
And a set of fully-connected Chinese character skeleton data set is constructed by connecting the pen collection of each pen of the Bezier curve and the pen start of the next pen. The adding mode is that a line segment "< path stroke-width=" 3 "stroke=" #000 "d=" MaBLcd "fill=" none "stroke-linecap=" round ">/path >" is added between each stroke line, wherein a, b, c, d are respectively the end point coordinates of the last line segment and the start coordinates of the next line segment. Then, a pair of skeleton-style Chinese character data sets are constructed by rendering the style Chinese character image from the ttf font file and corresponding to the skeleton data sets one by one according to unicode codes, as shown in fig. 2.
S2 structural design and training of generative model
S2.1 building model Structure
Firstly, in order to ensure the consistency of styles, a convolutional neural network is utilized to construct a style encoder, a series of Chinese characters with the same style are input into the encoder, and the extracted high-dimensional style features are subjected to dimension reduction by utilizing PCA to obtain style feature vectors.
Then, in order to ensure that the skeleton of the original Chinese character can be reserved for generating Chinese characters and focus on the call-for-response relation among strokes, bezier curves of the skeleton of the Chinese character are connected end to end so as to generate the skeleton of the Chinese character with full strokes and input the skeleton into a structure encoder, and the obtained structural characteristics are subjected to PCA dimension reduction to obtain structural characteristic vectors.
To ensure that the model can focus on the generation of continuous strokes, the application creatively provides a module focusing on continuous stroke information in training. A coded attention matrix is established and used as a label of a data set to be input into a model together. The concrete expression mode is as follows: dividing the Chinese character skeleton rendered into pictures into 8 x 8 sub-pictures, and if the original Bezier curve covers continuous strokes, setting the sub-picture as 1, otherwise setting the sub-picture as 0, so as to obtain a 1024-dimensional single-heat coding matrix. And flattening the attention coded matrix, respectively splicing the attention coded matrix with the two characteristic vectors, and finally inputting the attention coded matrix into a generator. In the training process of the generator, the generation quality of the generation result and the generation quality of the training set at the continuous stroke position are measured, and a continuous stroke consistency loss function is defined as a training direction.
The objective of the arbiter is to authenticate the paired data sets provided in the training set and the generated result of the generator, and to judge the former as false and the latter as true. And (3) carrying out back propagation through the comprehensive loss function, and continuously training and optimizing the discriminator and the generator.
The overall architecture of the model is shown in fig. 3.
S2.2 training of generative models
S2.2.1 Chinese character skeleton image is defined as c g The style Chinese character image is c f The skeleton of the Chinese character to be generated is c p And c f The other complete Chinese character image with the same style is c s ，c g And c f The image vector obtained by connection in the channel dimension is c v C extracted using a style encoder p Is coded as s p 。
In addition, character c x S for style coding x Representation, T for continuous coding x And is represented, wherein x can take all characters, the meaning of which is as described above.
S2.2.2 the integrated loss function includes the following four parts:
1) Countering the loss, for taking into account the distance of the result generated by the generator from the paired data sets;
wherein D is s (c) A representative discriminator for discriminating whether or not the image c having the style s is generated by the generator, G (c, s) representing the result of the generator by adding the style based on the skeleton of the Chinese character, s f Is a style Chinese character image c f Is a style encoding of (c).
In the training process, the generator and the discriminator are optimized simultaneously, and the purpose of the generator is to generate a generation result which can not be identified by the discriminator, and which is originally input and which is generated. The objective of the arbiter is to identify as much as possible which picture is generated by the generator and which is original.
2) L1 loss for taking into account the difference of the generated result from the standard result at the pixel level;
target average absolute error loss function:
wherein c v ，c f ，s s For style coding of Chinese character image to be migrated style, G (c) v ,s s ) Is C v And S is s The result after passing through the generator.
The target average absolute error loss function is used to optimize the generator. The more similar the Chinese character image generated by the generator is to the target image, the more ideal the generation is. The similarity of the two images is judged by calculating a pixel-by-pixel average absolute error loss function of the two images, and the loss is made to approach zero in the optimization process. The loss function ensures that the generated result of the generator has a visual similarity to the target result.
3) The style coding loss is used for considering the coding effect of the style coder;
style coding loss function:
wherein c v ，s s ，s f Is as defined above.
The style coding loss function is used to optimize both the generator and the style encoder. The working flow is as follows: a noise is extracted from random, and the noise is passed through a mapping network of a mapper to obtain a style code corresponding to a brand-new specific style. This style is then put into a generator together with the skeleton to generate a target image G. The style of the current image is extracted using a style encoder and compared with the style of noise generation. The average absolute error loss is calculated with the goal of making the loss approach zero. This ensures that the extracted style remains consistent with the original after passing through the generator. In the process, the accuracy of extracting the style codes by the style coder is optimized, and the utilization rate of the style codes by the generator is improved.
4) The consistency loss of the continuous strokes is used for focusing and generating the consistency degree of the continuous strokes of the Chinese character and the natural writing property of the target Chinese character;
definition S (c) represents extracting the text structure of the text image c using a concatenated code extractor and outputting it in the form of an embedded vector. And (3) placing the original image and the image of the target font generated by the generator into a residual neural network, calculating the mean square error loss of the output embedded vector, and minimizing a loss function through training, so that the final effect is consistent continuous coding of the original image and the target font on the embedded layer. The generator of the loss function optimization can be used for reserving continuous writing information of characters in the generating process, and natural writing property is improved.
The integrated loss function is employed to coordinate the four loss functions above. The integrated loss function of the model can be expressed as follows:
wherein lambda is gan ，λ L1 ，λ enc ，λ stroke To integrate the loss functions. When the network is trained, lambda is found in subsequent experiments gan Is set to a value of 10 lambda L1 、λ enc And lambda (lambda) stroke Set to 1, target formula used. The corner marks G, E, F, D represent the generator, the style encoder, the structure encoder, and the arbiter, respectively.
In order to verify that the model has the effect of generating the Chinese characters with natural writing property, a plurality of fonts are selected for example verification, and fig. 4, 5 and 6 show the results of migration of three representative example fonts with natural writing property, and a Su body, a Shu Tongti body and a Zhao Meng body are respectively selected as experimental example objects, wherein the picture on the left side is the migration result of the model after the Chinese character skeleton is input, and the picture on the right side is a target image in the test set.
The above description is only illustrative of the preferred embodiments of the application and is not intended to limit the scope of the application in any way. Any alterations or modifications of the application, which are obvious to those skilled in the art based on the teachings disclosed above, are intended to be equally effective embodiments, and are intended to be within the scope of the appended claims.
Claims (5)
1. A Chinese character style generation method with natural writing performance is characterized by comprising the following steps:
s1, constructing a data set;
s2, generating structural design and training of a model;
s3, inputting a model after training optimization to generate a single-line style Chinese character with natural writability.
2. The method for generating a style of chinese characters with natural writing property as in claim 1, wherein,
s1, constructing a data set: the full-continuous Chinese character skeleton data set is constructed by editing a closed curve in the Bezier curve and then rendering, and specifically comprises the following steps:
the Bezier curves are divided into first-order, second-order and third-order curves according to the orders, and are respectively marked as:
m (a, b) L (c, d) represents a first order bezier curve from point (a, b) to point (c, d);
m (a, b) C (x, y, C, d) represents a second order bezier curve starting from point (a, b) to point (C, d) and controlling the coordinates of the point (x, y);
M(a，b)Q(x 1 ,y 1 ,x 2 ,y 2 c, d) from point (a, b) to point (c, d), and the control point coordinates are (x 1 ,y 1 ) And (x) 2 ,y 2 ) A third-order bezier curve;
in the data set, the Chinese character skeleton consists of a plurality of lines, wherein the number of the lines is the stroke number of the Chinese character; each line is formed by connecting the three curves;
and a set of fully-connected Chinese character skeleton data set is constructed by connecting the pen collection of each pen of the Bezier curve and the pen start of the next pen.
3. The method for generating a style of chinese characters with natural writing property as in claim 2, wherein,
the method constructs a set of fully-connected Chinese character skeleton data set by connecting the pen collection of each pen of the Bezier curve and the pen start of the next pen, and the adding mode is as follows:
a line segment "< path stroke-width=" 3 "stroke=" #000 "is added between each stroke line"
d= "MabLcd" fill= "none" stroke-linecap= "round" >/path > "wherein a, b, c, d are the end coordinates of the last line segment and the start coordinates of the next line segment respectively;
then, a paired skeleton-style Chinese character data set is constructed by rendering a style Chinese character image from the ttf font file and corresponding to the skeleton data set one by one according to unicode codes.
4. The method for generating a style of chinese characters with natural writing property as in claim 2, wherein,
s2 specifically comprises the following steps:
s2.1 building model Structure
Firstly, in order to ensure the consistency of styles, a convolutional neural network is utilized to construct a style encoder, a series of Chinese characters with the same style are input into the encoder, and the extracted high-dimensional style features are subjected to PCA dimension reduction to obtain style feature vectors;
next, in order to ensure that the skeleton of the original Chinese character can be reserved for generating Chinese characters and focus on the call-for-response relation among strokes, bezier curves of the skeleton of the Chinese character are connected end to end so as to generate the skeleton of the Chinese character with full strokes and input the skeleton into a structure encoder, and the obtained structural features are subjected to PCA dimension reduction to obtain structural feature vectors;
finally, ensuring that the model can focus on the generation of continuous strokes, and providing a module focusing on continuous stroke information in training; setting up a coded attention matrix connected with a pen, and inputting the coded attention matrix serving as a label of a data set into a model;
the aim of the discriminator is to discriminate the paired data sets provided in the training set and the generated result of the generator, and the former is judged as false, and the latter is judged as true; counter propagation is carried out through the comprehensive loss function, and the discriminant and the generator are continuously trained and optimized;
s2.2 training of generative models
S2.2.1 Chinese character skeleton image is defined as c g The style Chinese character image is c f The skeleton of the Chinese character to be generated is c p And c f The other complete Chinese character image with the same style is c s ，c g And c f The image vector obtained by connection in the channel dimension is c v C extracted using a style encoder p Is coded as s p ；
S2.2.2 the integrated loss function includes the following four parts:
1) Countering the loss, for taking into account the distance of the result generated by the generator from the paired data sets;
wherein D is s (c) A representative discriminator for discriminating whether or not the image c having the style s is generated by the generator, G (c, s) representing the result of the generator by adding the style based on the skeleton of the Chinese character, s f The style coding is used for style Chinese character images;
2) L1 loss for taking into account the difference of the generated result from the standard result at the pixel level;
target average absolute error loss function:
wherein s is s For style coding of Chinese character image to be migrated style, G (c) v ,s s ) Is C v And S is s Results after passing through the generator;
3) The style coding loss is used for considering the coding effect of the style coder;
style coding loss function:
4) The consistency loss of the continuous strokes is used for focusing and generating the consistency degree of the continuous strokes of the Chinese character and the natural writing property of the target Chinese character;
comprehensive loss function of model:wherein lambda is gan ，λ L1 ，λ enc ，λ stroke Each parameter is used for integrating the loss function; the corner marks G, E, F, D represent the generator, the style encoder, the structure encoder, and the arbiter, respectively.
5. The method for generating a style of chinese characters with natural writing property as recited in claim 4, wherein,
the method is characterized in that a coding attention matrix connected with a pen is established and is used as a label of a data set to be input into a model together, and the specific expression mode is as follows: dividing a Chinese character skeleton rendered into pictures into 8 x 8 sub-pictures, and if the original Bezier curve covers continuous strokes, setting the sub-picture as 1, otherwise setting the sub-picture as 0, so as to obtain a 1024-dimensional single-heat coding matrix; flattening the matrix of the attention code, respectively splicing the matrix with the two characteristic vectors, and finally inputting the matrix into a generator; in the training process of the generator, the generation quality of the generation result and the generation quality of the training set at the continuous stroke position are measured, and a continuous stroke consistency loss function is defined as a training direction.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310883227.4A CN116935406A (en) | 2023-07-18 | 2023-07-18 | Chinese character style generation method with natural writing property |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310883227.4A CN116935406A (en) | 2023-07-18 | 2023-07-18 | Chinese character style generation method with natural writing property |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116935406A true CN116935406A (en) | 2023-10-24 |
Family
ID=88391885
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310883227.4A Pending CN116935406A (en) | 2023-07-18 | 2023-07-18 | Chinese character style generation method with natural writing property |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116935406A (en) |
-
2023
- 2023-07-18 CN CN202310883227.4A patent/CN116935406A/en active Pending
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110503598B (en) | Font style migration method for generating countermeasure network based on conditional cycle consistency | |
| Jiang et al. | Scfont: Structure-guided chinese font generation via deep stacked networks | |
| CN112232149B (en) | Document multimode information and relation extraction method and system | |
| CN109635883A (en) | The Chinese word library generation method of the structural information guidance of network is stacked based on depth | |
| CN110728219A (en) | 3D face generation method based on multi-column multi-scale graph convolution neural network | |
| CN110033054B (en) | Personalized handwriting migration method and system based on collaborative stroke optimization | |
| CN113901894A (en) | Video generation method, device, server and storage medium | |
| CN111508048A (en) | Automatic generation method for human face cartoon with interactive arbitrary deformation style | |
| CN112036137A (en) | Deep learning-based multi-style calligraphy digital ink simulation method and system | |
| CN110473151B (en) | Partition convolution and correlation loss based dual-stage image completion method and system | |
| CN115063799B (en) | Print form mathematical formula identification method and device and storage medium | |
| CN117058266B (en) | Handwriting word generation method based on skeleton and outline | |
| CN110097615B (en) | Stylized and de-stylized artistic word editing method and system | |
| CN112418235A (en) | Point cloud semantic segmentation method based on expansion nearest neighbor feature enhancement | |
| Jung et al. | Learning free-form deformation for 3D face reconstruction from in-the-wild images | |
| CN114419174A (en) | On-line handwritten text synthesis method, device and storage medium | |
| CN113065417A (en) | Scene text recognition method based on generation countermeasure style migration | |
| CN112597998A (en) | Deep learning-based distorted image correction method and device and storage medium | |
| CN113096133A (en) | Method for constructing semantic segmentation network based on attention mechanism | |
| CN116485892A (en) | Six-degree-of-freedom pose estimation method for weak texture object | |
| CN116935406A (en) | Chinese character style generation method with natural writing property | |
| CN114861595B (en) | Vector line transformation-based individual font generation method | |
| CN113554549B (en) | Text image generation method, device, computer equipment and storage medium | |
| CN113052156B (en) | Optical character recognition method, device, electronic equipment and storage medium | |
| CN113298814A (en) | Indoor scene image processing method based on progressive guidance fusion complementary network |
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 |