CN114626335A - Character generation method, network training method, device, equipment and storage medium - Google Patents

Abstract

The application provides a character generation method, a network training device, a network training equipment and a storage medium for character generation, which can be applied to various scenes such as cloud technology, artificial intelligence, intelligent traffic, auxiliary driving and the like, wherein the character generation method comprises the following steps: acquiring at least two candidate characters with different font style information from a character set to be processed; generating font style information corresponding to at least two candidate characters based on a font style coding network; generating target characters corresponding to the font style information and the character content information based on a character generation network; the font style coding network and the character generation network are obtained by performing character generation training on a preset neural network, and the hidden space of the sample font style information used in the character generation training process is restricted to be in normal distribution. According to the embodiment of the application, the generation quality of the target characters can be improved, and the generation cost of the target characters is reduced.

Description

Character generation method, network training method, device, equipment and storage medium

technical field

The present application belongs to the field of computer technology, and specifically relates to a character generation method, a network training method, an apparatus, a device and a storage medium.

Background technique

In the related art, the stylistic representation of a font is usually learned from a subset of characters of a certain font, and then a whole set of characters corresponding to the new font is generated.

However, related technologies need to pre-set some texts under this new font, and then input these texts into a text generation model to learn style features, and then obtain other text results from the model, and finally obtain a complete set of texts corresponding to the new fonts. This process will consume more system resources, resulting in higher cost of text generation. In addition, the text generation model in the related art is prone to the problem of missing strokes or adhesion of strokes, resulting in low quality of text generation.

SUMMARY OF THE INVENTION

In order to solve the above technical problems, the present application provides a character generation method, a network training method, an apparatus, a device and a storage medium.

On the one hand, the present application proposes a method for generating text, the method comprising:

Obtain at least two candidate characters with different font style information from the character set to be processed;

Generate font style information corresponding to each of the at least two candidate characters based on a font style coding network;

Based on the text generation network, the target text corresponding to the font style information and the text content information is generated; the text content information represents the content of the text in the to-be-processed text set;

The font style encoding network and the text generation network are obtained by performing text generation training on a preset neural network, and the latent space of the sample font style information used in the text generation training process is constrained to be a normal distribution .

On the other hand, the present application provides a network training method for text generation, the method comprising:

Extract the first sample text and the second sample text from the sample text set;

Based on the sample font style information of the first sample text and the sample text content information of the second sample text, text generation training is performed on a preset neural network, and during the text generation training process, the sample font The latent space constraint of style information is normal distribution, and the font style encoding network and text generation network are obtained.

On the other hand, an embodiment of the present application provides a character generation device, the device comprising:

A text acquisition module, used for acquiring at least two candidate texts with different font style information from the to-be-processed text set;

a font style information generation module, configured to generate font style information corresponding to each of the at least two candidate characters based on a font style coding network;

A target text generation module, configured to generate a target text corresponding to the font style information and text content information based on a text generation network; the text content information represents the content of the text in the to-be-processed text set;

The font style encoding network and the text generation network are obtained by performing text generation training on a preset neural network, and the latent space of the sample font style information used in the text generation training process is constrained to be a normal distribution .

On the other hand, the present application provides a network training device for text generation, the device comprising:

The sample text acquisition module is used to extract the first sample text and the second sample text from the sample text set;

A training module, configured to perform text generation training on a preset neural network based on the sample font style information of the first sample text and the sample text content information of the second sample text, and during the text generation training process, Constraining the latent space of the sample font style information to a normal distribution, a font style encoding network and a character generation network are obtained.

On the other hand, the present application proposes an electronic device, the electronic device includes a processor and a memory, the memory stores at least one instruction or at least one program, and the at least one instruction or at least one program is loaded and executed by the processor to realize As described above, the text generation method or the text generation network training method.

On the other hand, the present application proposes a computer-readable storage medium, in which at least one instruction or at least one program is stored, and the at least one instruction or the at least one program is loaded by a processor and stored therein. Execute to implement the text generation method or the text generation network training method as described above.

On the other hand, the present application proposes a computer program product, including a computer program, when the computer program is executed by a processor, the above-mentioned character generation method or character generation network training method is implemented.

The text generation method, network training method, device, equipment and storage medium proposed in the embodiments of the present application use a trained font style coding network to generate font style information corresponding to the at least two candidate characters, and use the trained text to generate font style information. The network generates the target text corresponding to the font style information and text content information. Since the latent space of the sample font style information is constrained to a normal distribution during the training process, the font style information is compressed and the difference between different font style information is shortened. The spacing between the font styles makes the change of the style latent space smoother, avoiding the network encountering discontinuities, thus improving the generation quality of the target text; in addition, using the trained font style encoding network and text generation network can also reduce the target text generation process. Consumption of system resources, thereby reducing the cost of generating target text.

Description of drawings

In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present application or in the prior art, the following briefly introduces the accompanying drawings used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description The drawings are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings without any creative effort.

FIG. 1 is a schematic diagram of an implementation environment of a text generation method according to an exemplary embodiment.

FIG. 2 is a schematic flowchart of a method for generating characters according to an exemplary embodiment.

Fig. 3 is a flow chart showing a process of standardizing style and content according to an exemplary embodiment.

Fig. 4 is a flow chart of a network training method for text generation according to an exemplary embodiment.

Fig. 5 is a schematic diagram of a preset neural network according to an exemplary embodiment.

Fig. 6 is a flow chart of a lattice coding network and a character generation network obtained by training according to an exemplary embodiment.

FIG. 7 is a schematic diagram of obtaining sample font style information according to an exemplary embodiment.

FIG. 8 is a schematic diagram illustrating a method of standardizing sample font style information and sample text content information according to an exemplary embodiment.

FIG. 9 is a schematic diagram of a target text generated by using the text generation method according to an embodiment of the present application, according to an exemplary embodiment.

FIG. 10 is a comparison diagram of fusion effects according to an exemplary embodiment.

FIG. 11 is a schematic diagram of a similar font obtained by detecting the font of a target text by a font similarity detection model according to an exemplary embodiment.

Fig. 12 shows a text generating apparatus according to an exemplary embodiment.

Fig. 13 shows a network training apparatus for text generation according to an exemplary embodiment.

FIG. 14 is a hardware structural block diagram of a server for character generation or character generation network training provided by an embodiment of the present application.

Detailed ways

Artificial Intelligence (AI) is a theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technique of computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can respond in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines, so that the machines have the functions of perception, reasoning and decision-making.

Artificial intelligence technology is a comprehensive discipline, involving a wide range of fields, including both hardware-level technology and software-level technology. The basic technologies of artificial intelligence generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation/interaction systems, and mechatronics. Artificial intelligence software technology mainly includes computer vision technology, speech processing technology, natural language processing technology, and machine learning/deep learning.

Specifically, the embodiments of the present application relate to artificial neural network technology in deep learning.

In order to make those skilled in the art better understand the solutions of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only The embodiments are part of the present application, but not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by persons of ordinary skill in the art without creative work shall fall within the protection scope of the present application.

It should be noted that the terms "first", "second", etc. in the description and claims of the present application and the above drawings are used to distinguish similar objects, and are not necessarily used to describe a specific sequence or sequence. It is to be understood that data so used may be interchanged under appropriate circumstances so that the embodiments of the application described herein can be practiced in sequences other than those illustrated or described herein. Furthermore, the terms "comprising" and "having" and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product or server comprising a series of steps or units is not necessarily limited to those expressly listed Rather, those steps or units may include other steps or units not expressly listed or inherent to these processes, methods, products or devices.

FIG. 1 is a schematic diagram of an implementation environment of a text generation method according to an exemplary embodiment. As shown in FIG. 1 , the implementation environment may include at least a terminal 01 and a server 02 , and the terminal 01 and the server 02 may be directly or indirectly connected through wired or wireless communication, which is not limited in this application.

Specifically, the terminal can be used to collect the text set to be processed and the sample text set. Optionally, the terminal 01 may include, but is not limited to, a mobile phone, a computer, an intelligent voice interaction device, a smart home appliance, a vehicle-mounted terminal, an aircraft, and the like. The embodiments of the present application may be applied to various scenarios, including but not limited to cloud technology, artificial intelligence, smart transportation, and assisted driving.

Specifically, the server 02 can be used to train a font style encoding network and a character generation network, and generate target characters based on the font style encoding network and the character generation network. Optionally, the server 02 may be an independent physical server, a server cluster or a distributed system composed of multiple physical servers, or a cloud service, cloud database, cloud computing, cloud function, cloud storage, network Cloud servers for basic cloud computing services such as services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms.

It should be noted that FIG. 1 is only an example. In other scenarios, other implementation environments may also be included, for example, the implementation environment may include a terminal, a font style encoding network and a character generation network are obtained through terminal training, and target characters are generated based on the font style encoding network and the character generation network.

FIG. 2 is a schematic flowchart of a method for generating characters according to an exemplary embodiment. This method can be used in the implementation environment of FIG. 1 . This specification provides method operation steps as described in the embodiments or flow charts, but more or less operation steps may be included based on routine or non-creative work. The sequence of steps enumerated in the embodiments is only one of the execution sequences of many steps, and does not represent the only execution sequence. When an actual system or server product is executed, it can be executed sequentially or in parallel (for example, in a parallel processor or multi-threaded processing environment) according to the embodiments or the methods shown in the accompanying drawings. Specifically, as shown in Figure 2, the method may include:

S101. Acquire at least two candidate characters with different font style information from the character set to be processed.

Specifically, the font style information may include, but is not limited to, italics, Songs, Fangzheng cartoon, Hei, Chinese Xingkai, Fangzheng peach, and the like. Optionally, the to-be-processed character set may include all characters of various font style information. For example, it includes all the characters in italics, all the characters in Song style, and all the characters in Fangzheng cartoons, etc.

In this embodiment of the present application, at least two candidate characters with different font style information may be obtained from the character set to be processed in various ways, which are not specifically limited herein.

In one manner, the characters in the processed character set may be classified according to the font style information to obtain a plurality of font style information categories, and each font style information category corresponds to a plurality of characters. At least two candidate font style information categories may be determined from a plurality of font style information categories, and one character is respectively extracted from the characters corresponding to the at least two candidate font style information categories to obtain at least two candidate characters. For example, it is possible to determine the italic type and the Song type category from multiple font style information categories, the italic type corresponds to multiple characters, the Song type category corresponds to multiple characters, extract a character from the multiple characters corresponding to the italic type, and extract a character from the Song type category One character is extracted from the corresponding multiple characters to obtain at least two candidate characters.

In another manner, a preset number of characters may be respectively extracted from the characters corresponding to the at least two candidate font style information categories to obtain at least two kinds of candidate characters. For example, it is possible to determine the italic type and the Song type category from a plurality of font style information categories, the italic type corresponds to a plurality of characters, the Song type category corresponds to a plurality of characters, extract a preset number of characters from the multiple characters corresponding to the italic type, and extract a preset number of characters from the multiple characters corresponding to the italic type. A preset number of characters are extracted from a plurality of characters corresponding to the italic type category, and at least two kinds of candidate characters are obtained.

S103. Generate font style information corresponding to each of the at least two candidate characters above based on a font style coding network.

S105. Based on the text generation network, generate the target text corresponding to the font style information and the text content information; the text content information represents the content of the text in the to-be-processed text set.

The font style encoding network and the text generation network are obtained by performing text generation training on a preset neural network, and the latent space of the sample font style information used in the text generation training process is constrained to be a normal distribution.

Exemplarily, the text content information may be spatial information of the text content, for example, the stroke structure of the text.

Exemplarily, the latent space can refer to the space where the latent variables (eg, noise Z) are located, and the latent space constraint can be understood as constraining the latent vector of the latent space, and by adjusting the sample fonts during the training process. The latent space of style information is constrained, so that in the process of text generation, text with better quality and effect can be generated.

In a specific embodiment, in the above step S103, at least two candidate characters may be input into a pre-trained font style coding network, and feature extraction is performed on the at least two candidate characters through the font style coding network, respectively, Obtain font style information corresponding to at least two candidate characters respectively. For example, through the pre-trained coding network, the font style information corresponding to the at least two candidate characters extracted are Kai type and Song type respectively.

Optionally, in the embodiment of the present application, the font style information corresponding to the at least two candidate characters above may be generated in various ways, which are not specifically limited herein.

In one way, in the above step S103, the generating the font style information corresponding to the at least two candidate characters based on the font style coding network may include: mapping the at least two candidate characters to the font style coding network based on the font style coding network. In the above normal distribution, font style information corresponding to each of the at least two candidate characters is obtained.

Since the latent space of the sample font style information is constrained to be a normal distribution in the above text generation training process, after inputting at least two candidate characters into the font style coding network, the font style coding network can use the at least two candidate characters to The candidate characters are mapped to the normal distribution, the style feature vectors corresponding to the at least two candidate characters are extracted from the normal distribution, and the style feature vectors corresponding to the at least two candidate characters are extracted, and the at least two candidate characters are obtained. The font style information corresponding to each candidate text. Since the latent space of the sample font style information is constrained to be a normal distribution, the font style information is compressed during the font style information extraction process, and the distance between the different font style information corresponding to at least two candidate characters is shortened. It makes the change of the style latent space smoother and avoids the network encountering discontinuous points, thereby improving the generation quality of font style information, thereby improving the generation quality of target text.

In a specific embodiment, in the above step S105, the content of the text can be extracted from the text set to be processed as text content information, and the text content information and font style information can be input into the pre-trained text generation network , to generate the target text. Among them, the text generation network can include two branches: a style branch and a content branch. The input of the style branch is the font style information obtained by the font style encoding network, and the input of the content branch is the text content information.

In one method, the text content information of each character can be sequentially extracted from the text set to be processed, and the text content information of each character can be traversed in turn to obtain the target text corresponding to the text content information and font style information of each character. .

In another way, the text content information of all characters can also be extracted from the to-be-processed character set, and the text content information of all the characters can be traversed in parallel to obtain the target text corresponding to the text content information of all the characters and the font style information.

In an optional embodiment, the above method may further include: calculating an average value of the above font style information based on the above font style coding network to obtain the target font style information.

Correspondingly, in the above step S105, generating the target text corresponding to the font style information and the text content information based on the text generation network may include: generating the target font style information corresponding to the text content information based on the text generation network. of the above target text.

Specifically, since at least two candidate characters are input into the font style coding network, for each candidate character, the font style coding network will output a font style information. Therefore, the fonts corresponding to the at least two candidate characters can be obtained. The average value of the style information is used to obtain the target font style information. The target font style information can be understood as a new style feature, that is, the target text style feature is different from the font style information corresponding to the text in the text set to be processed. Exemplarily, since text can be treated as an image, and the style information of the image is usually represented by the mean and variance, the above calculation of the mean value of the font style information corresponding to at least two candidate texts can be understood as finding the corresponding font style information of at least two candidate texts. mean of the image.

Correspondingly, after obtaining the target font style information, the target font style information and the text content information can be input into the text generation network to obtain the target text corresponding to the text content information and the target font style information. In the embodiment of the present application, by calculating the average value of font style information corresponding to at least two candidate characters, new style features (ie, target font style information) can be obtained, and the accuracy of generating the target font style information is high. Inputting the target font style information and text content information with high accuracy into the text generation network can improve the generation quality of the target text; in addition, a whole set of new text corresponding to the target font style information can be generated through the text generation network (ie target text), there is no need to pre-set part of the text under a new font, which can reduce the consumption of system resources in the process of generating the target text, thereby reducing the cost of generating the target text.

Fig. 3 is a flow chart showing a process of standardizing style and content according to an exemplary embodiment. As shown in FIG. 3, in an optional embodiment, the above method may further include:

S201. Decouple the above font style information.

S203. Standardize the decoupled font style information to obtain standard font style information.

S205. Standardize the above text content information to obtain standard text content information.

Correspondingly, in the above-mentioned step S105, the above-mentioned generation of the target text corresponding to the above-mentioned font style information and text content information based on the text generation network may include:

Based on the character generation network, the target character corresponding to the standard font style information and the standard character content information is generated.

In an optional embodiment, the structure of the text generation network may be a SPADE+AdaIN structure. Among them, SPADE is a generative model that converts segmentation graphs into reality graphs. As a backbone network, SPADE can better preserve the structural information of text. AdaIN is the abbreviation of Adaptive Instance Normalization, and the Chinese name is Adaptive Instance Normalization.

Exemplarily, in the above steps S201-S203, the font style information can be input into a multi-layer perceptron (Multi-layer Perceptron, MLP) for decoupling, and the decoupled font style information can be standardized by means of AdaIN. After processing, the standard font style information after standardized processing is obtained, that is, the decoupled style is input into each layer of the network in the form of AdaIN, and the style of the generated image is controlled by controlling the feature statistics. Among them, MLP is an artificial neural network with a forward structure. For example, decoupling may refer to: decoupling the style itself, such as stroke thickness, stroke and so on.

The font style information in the embodiment of the present application is decoupled after multiple fully connected layers, and then sent to each layer of the network. In this way, the text generation network can extract font style information from various semantic levels.

In text generation, the content branch preserves the stroke structure of the text. The stroke structure of text (ie, text content information) is a key factor. Exemplarily, in the above step S205, the stroke structure of the text can be standardized in the form of Spatial Adaptive Normalization (SpatialAdaIN) to obtain standard text content information, that is, the stroke structure of the text is in the form of SpatialAdaIN. Inserted into the text generation network, the stroke structure of the text is better preserved in this way, that is, the spatial information of the original segmentation is preserved), and the generation quality of the target text is improved.

Fig. 4 is a flow chart of a network training method for text generation according to an exemplary embodiment. As shown in Figure 4, the method may include:

S301. Extract the first sample text and the second sample text from the sample text set.

Optionally, the first sample text may be a sample text set or a single text, and the second sample text may also be a sample text set or a single text.

即第一样本文字。其中，

d_s是风格嵌入的维度，K为大于或等于1的正整数。Illustratively, for any text x _ij in the sample text set, it contains a style attribute S _i ∈S and a content attribute C _j ∈C. In order to represent a style, K characters can be taken from the font whose style is _Si in the sample character set to form a reference set of this style

That is, the first sample text. in,

d _s is the dimension of the style embedding, and K is a positive integer greater than or equal to 1.

即第二样本文字。Exemplarily, in order to represent the content attribute C _j , some texts can be taken from the sample text set to form the content reference set

That is, the second sample text.

中的文字，作为第二样本文字。If the text of different styles and the same content is selected as the content reference set, it will increase the difficulty of network learning content. This is because different styles of text have great differences in position and stroke shape, and it is difficult for the network to abstract the correct content from them. stroke structure. Therefore, in an exemplary embodiment, in order to reduce the difficulty of learning the content on the network, a sample text whose font style information is Song type can be selected from the sample text set as the second sample text. Since Song type is a standard font, it is beneficial to the model Learn font structure. Further, in order to further reduce the difficulty of online learning content, a single Song Dynasty can also be selected from the sample text set, that is,

The text in , as the second sample text.

S303. Based on the sample font style information of the above-mentioned first sample text and the sample text content information of the above-mentioned second sample text, perform text generation training on a preset neural network, and in the above-mentioned text generation training process, the above-mentioned sample font style information The latent space constraint is normal distribution, and the font style encoding network and the text generation network are obtained.

In the embodiment of the present application, the sample font style information of the first sample text and the sample text content information of the second sample text may be input into a preset neural network for text generation training, and in the above text generation training process, the above The latent space constraint of the sample font style information is normal distribution, so as to obtain the font style encoding network and the text generation network. During the training process, the latent space of the sample font style information is constrained to a normal distribution, which compresses the font style information, narrows the distance between different font style information, makes the change of the style latent space smoother, and avoids the need for network When encountering discontinuities, the training accuracy of the font style encoding network and the text generation network is improved; in addition, based on the sample font style information of the first sample text and the sample text content information of the second sample text, the preset neural network is trained for text generation , the font style encoding network and the text generation network can be obtained, and the consumption of system resources during the training process is less, thereby reducing the training difficulty and cost of the network.

Fig. 5 is a schematic diagram of a preset neural network according to an exemplary embodiment. As shown in FIG. 5 , the preset neural network may include a preset font style encoding network, a preset text generation network and a preset discrimination network, and the preset neural network is a generative adversarial nets (GAN) as a whole. Structure. In addition, the preset neural network uses SpatialAdaIN and AdaIN to input spatial information of sample text content information and sample font style information of different levels to each layer of the generator respectively.

The following describes the process of using the preset neural network in Figure 5 to train the font style encoding network and the text generation network:

Suppose, the first sample text is "Gong, ..., Men, Zheng" of "Founder Cartoon", which is used to represent the font style information of "Founder Cartoon", and the second sample text is Song type "De", which is used for Indicates the text content of "de", input the first sample text into the preset style coding network to obtain sample font style information, input the sample font style information into the preset text generation network in the form of AdaIN, and convert the second sample text into the preset text generation network. The sample text content information is inserted into the preset text generation network in the form of SpatialAdaIN, and the preset text generation network outputs the text "De" (that is, the reference text) under "Founder Cartoon", and the preset discriminator network is used to determine the output of the training. Refer to the text "De" (that is, the reference text) under "Founder cartoon" in the sample text set to obtain the loss information, and continuously adjust the parameters of the network during the training process until the loss information meets the preset conditions. Stop the training process, and get Trained font style encoding network and text generation network.

Fig. 6 is a flow chart of a lattice coding network and a character generation network obtained by training according to an exemplary embodiment. As shown in FIG. 6, in the above S303, based on the sample font style information of the first sample text and the sample text content information of the second sample text, the preset neural network is subjected to text generation training, and the text generation In the training process, the latent space of the above-mentioned sample font style information is constrained to a normal distribution, and the font style encoding network and the text generation network are obtained, which can include:

S3031. Map the first sample text to the latest normal distribution based on the preset font style coding network to obtain current sample font style information.

)。Optionally, in the above step S3031, the preset font style coding network may be a forward propagation network, that is, the forward propagation network takes the first sample text (ie, the style reference set R _S ) as input, and outputs the current sample font. style information (i.e. style embedding vector

).

和

上述两个向量表示多元正态分布N(μ_s,σ_s)的参数μ_s和σ_s。Exemplarily, the style embedding vector Z _S may be obtained by random sampling. The forward propagation network can map the style reference set R _S to the latest normal distribution, and output two vectors

and

The above two vectors represent the parameters μ _s and σ _s of the multivariate normal distribution N(μ _s ,σ _s ).

S3033. Process the standard normal distribution and the current sample font style information based on the preset font style coding network to obtain the sample font style information.

In the embodiment of the present application, in the above step S3033, the preset font style coding network can make the current sample font style information align with the multivariate standard normal distribution, so as to obtain the sample font style information.

FIG. 7 is a schematic diagram of obtaining sample font style information according to an exemplary embodiment. As shown in FIG. 7, in the above step S3033, the above-mentioned standard normal distribution and the above-mentioned current sample font style information are processed based on the above-mentioned preset font style coding network to obtain the above-mentioned sample font style information, which may include:

S30331. Based on the above-mentioned preset font style coding network, randomly obtain a feature vector corresponding to the above-mentioned current sample font style information from the above-mentioned standard normal distribution.

S30333. Based on the above-mentioned preset font style coding network, through the difference information between the above-mentioned feature vector and the above-mentioned current sample font style information, the above-mentioned latest normal distribution is updated, and the updated latest normal distribution is re-used as the above-mentioned latest normal distribution.

S30335. Based on the above-mentioned preset font style coding network, in the above-mentioned mapping of the above-mentioned first sample text to the latest normal Tai distribution, obtain the current sample font style information, and the above-mentioned updated latest normal Tai distribution is re-used as the above-mentioned latest normal Tai distribution. It repeats from time to time until the difference information above satisfies the preset condition.

S30337. Based on the above-mentioned preset font style coding network, use the current sample font style information when the above-mentioned difference information satisfies the preset condition as the above-mentioned sample font style information.

)向多元标准正态分布看齐，预设字体风格编码网络可以从标准正态分布中随机确定一个与该当前样本字体风格信息对应的特征向量。在上述步骤S20333-S20337中，该预设编码网络可以计算该特征向量和上述当前样本字体风格信息之间的差异信息，并将该差异信息作为分布损失值，基于该分布损失值可以对预设字体风格编码网络中的最新正太分布进行调整更新，以使最新正太分布被不断调整为能够用于得到质量较高的样本字体风格信息，即最新正太分布被不断调整使得差异信息满足预设条件，并将差异信息满足预设条件时的当前样本字体风格信息，作为该样本字体风格信息。Exemplarily, in the above step S30331, during training, in order to make the current sample font style information (ie the style embedding vector

) is aligned with the multivariate standard normal distribution, and the preset font style encoding network can randomly determine a feature vector corresponding to the font style information of the current sample from the standard normal distribution. In the above steps S20333-S20337, the preset coding network can calculate the difference information between the feature vector and the above-mentioned current sample font style information, and use the difference information as a distribution loss value. The latest orthodox distribution in the font style coding network is adjusted and updated, so that the latest orthodox distribution is continuously adjusted to be able to obtain high-quality sample font style information, that is, the latest orthodox distribution is continuously adjusted so that the difference information satisfies the preset conditions, The current sample font style information when the difference information satisfies the preset condition is used as the sample font style information.

In one approach, the following constraints can be imposed on the preset style encoder:

Among them, N represents multivariate normal distribution, KL refers to KL divergence, and KL divergence is an indicator to measure the degree of matching of two probability distributions. The greater the difference between the two distributions, the greater the KL divergence.

Correspondingly, in the above-mentioned step S30333, the KL divergence between the feature vector and the above-mentioned current sample font style information can be calculated, and based on the KL divergence, the latest normal distribution in the preset font style coding network is adjusted and updated, so that The latest normal distribution is continuously adjusted so that it can be used to obtain higher-quality sample font style information, that is, the latest normal distribution is continuously adjusted so that the KL divergence is less than the preset divergence threshold, and the KL divergence is less than the preset divergence threshold. The current sample font style information at the time is used as the sample font style information.

In another way, in the above step S30333, the maximum average difference algorithm may also be used to calculate the difference information between the feature vector and the above-mentioned current sample font style information. Among them, the Maximum Mean Difference algorithm is used to measure the difference between two distributions.

这个空间里任意选取，从而压缩了风格嵌入Z_S的空间，拉近了不同风格之间的间距，使得风格隐空间的变化更加平滑，避免网络遭遇间断点，从而提高了字体风格编码网络的训练精度。In the embodiment of the present application, by imposing constraints on the preset font style encoding network, the current sample font style information (style embedding vector Z _S ) is sampled from the multivariate standard normal distribution, so each dimension of the style embedding vector Z _S is in near the mean of 0, but not at

This space is arbitrarily selected, thereby compressing the space of style embedding _ZS , narrowing the distance between different styles, making the change of style latent space smoother, avoiding the network encountering discontinuous points, thus improving the training of font style encoding network. precision.

S3035. Based on the preset text generation network, generate the current text corresponding to the sample font style information and the sample text content information.

)，而内容分支输入的是内容文字图像，因而预设文字生成网络生成的当前文字

In the embodiment of the present application, given sample font style information (style S _i ) and sample text content (content C _j ), the goal of the preset text generation network is to generate text x _ij corresponding to the style and content. The input of the preset text generation network includes two branches: the content branch and the style branch. The input of the style branch is the sample font style information (that is, the style embedding) obtained by the preset font style encoding network.

), and the content branch input is the content text image, so the preset text generation network generates the current text

S3037. Based on the preset discrimination network, perform discrimination processing on the current text, the reference text, the sample font style information, and the sample text content information to obtain loss information.

S3039. Train the preset font style coding network and the preset text generation network based on the loss information to obtain the font style coding network and the text generation network; the reference text represents the sample text content information under the sample font style information Text.

The preset discriminant network is an important part of GAN. Its goal is to judge whether the given current text is real enough. If it is real, the preset discriminant network should give a high score to the current text, otherwise it will give a low score. During training, the preset discrimination network usually considers that the current text generated by the preset text generation network is not real enough, and only the reference fonts in the sample text set are real. In this way, the preset text generation network can only generate more realistic images in order to fool the preset discrimination network. The preset text generation network and the preset discrimination network continuously improve their abilities in such a zero-sum game. In the end, the current text generated by the preset text generation network is close to the reference text, so that a higher-quality text generation can be trained. network.

The purpose of the preset text generation network is to deceive the preset discrimination network, that is, the preset text generation network does not have direct supervision information, but relies on the preset discrimination network to provide supervision information. In an optional embodiment, the preset discriminating network can use two preset discriminating networks D ₁ and D ₂ with the same structure to discriminate images of different sizes, and the images of small sizes can make the preset discriminating network have a larger size. Receptive field, and large-size images make the preset discriminant network pay more attention to details, and can also avoid the problem of overfitting to a certain extent.

In a specific embodiment, for each discriminative network, a hinge loss function (hingeloss) can be used as the loss function of GAN:

Among them, G, D, and E refer to the preset text generation network, preset discrimination network and preset font style encoding network respectively, and k refers to the number of preset discrimination networks.

In an exemplary embodiment, in order to enable the preset discrimination network to distinguish the content and style of fonts and to improve the network quality, in S3037, in addition to the current text and reference text, the input of the preset discrimination network may also include Sample font style information and sample text content information.

表示输入文字x在第k个判别器，第i层的feature，则featurematching loss可表示为：In another exemplary embodiment, in order to stabilize the training process, a feature matching loss (feature matching loss) can also be used to align the features of the current text and the reference text in each layer of the preset discriminant network, so that

Indicates that the input text x is the feature of the kth discriminator and the ith layer, then the featurematching loss can be expressed as:

T表示预设判别网络的卷积层数量，N_t表示预设判别网络第t层feature的元素数量。in,

T represents the number of convolutional layers of the preset discriminant network, and _Nt represents the number of elements of the t-th layer feature of the preset discriminant network.

To sum up, after constructing each part of the preset neural network, the overall optimization goal of the entire preset neural network is:

Among them, the hyperparameters λ _FM =10, λ _VAE =0.05, G, D, and E respectively refer to the preset text generation network, the preset discrimination network and the preset font style encoding network.

The preset neural network in the embodiment of the present application may include a preset font style encoding network, a preset text generation network, and a preset discrimination network, and the latent space of the sample font style information is constrained to a normal distribution through the preset font style encoding network , thereby compressing the font style information, narrowing the distance between different font style information, making the change of the style latent space smoother, avoiding the network encountering discontinuities, thereby improving the font style encoding network and text generation network training accuracy, The training cost and difficulty are reduced; in addition, the current text corresponding to the sample font style information and the sample text content information is generated through the preset text generation network, and the current text, reference text, sample font style information and sample text content information are generated through the preset discrimination network. The sample text content information is discriminated to obtain the loss information, and based on the loss information, the font style coding network and the text generation network are obtained by training, which further improves the training accuracy of the font style coding network and the text generation network and reduces the training cost. .

FIG. 8 is a schematic diagram illustrating a method of standardizing sample font style information and sample text content information according to an exemplary embodiment. As shown in FIG. 8, in an optional embodiment, the above method may further include:

S401. Decouple the above sample font style information.

S403. Standardize the decoupled sample font style information to obtain standard sample font style information.

S405. Standardize the above-mentioned sample text content information to obtain standard sample text content information.

Correspondingly, in the above step S3035, based on the above-mentioned preset text generation network, the above-mentioned sample font style information and the current text corresponding to the above-mentioned sample text content information are generated, including:

Based on the preset text generation network, the current text corresponding to the standard sample text content information and the standard sample font style information is generated.

In an optional embodiment, continuing as shown in FIG. 5 , the structure of the preset text generation network may be a SPADE+AdaIN structure. Among them, SPADE is a generative model that converts segmentation graphs into reality graphs. As a backbone network, SPADE can better preserve the structural information of text.

Exemplarily, in the above steps S401-S403, the sample font style information can be input into the ML for decoupling, and the decoupled sample font style information can be standardized through AdaIN to obtain a standardized standard. Sample font style information, and input the decoupled style into each layer of the network in the form of AdaIN, and control the style of the generated image by controlling the feature statistics. Font style information is decoupled after multiple fully connected layers, and then fed into each layer of the network. In this way, the preset text generation network can extract font style information from various semantic levels. Furthermore, SPADE uses AdaIN as the normalization method, which can well preserve the spatial information of the original segmentation. That is, through the structure of the generator of SPADE+AdaIN, the quality of font fusion can be improved.

During the text generation process, the content branch preserves the stroke structure of the text. The stroke structure of text (ie, text content information) is a key factor. Exemplarily, in the above step S405, the sample text content information can be inserted into the network in the form of SpatialAdaIN, to standardize the sample text content information to obtain standard sample text content information, thereby retaining the stroke structure of the text ( That is, the spatial information of the original segmentation), which improves the generation quality of the current text.

In a specific embodiment, the specific process of standardizing processing by means of AdaIN can be as follows:

AdaIN takes a content input x and a style input s and normalizes it by aligning the channel-level (C) mean and standard deviation of x onto s. AdaIN does not need to learn affine parameters, it can adaptively calculate affine parameters from the style input, and the calculation formula is as follows:

AdaIN(x,s)=σ(s)(σ(x)x-μ(x))+μ(s),

Among them, IN in AdaIN refers to calculating the mean and standard deviation in space (H, W) for each feature channel (C) of each sample (N).

In a specific embodiment, the specific process of standardization processing by means of Spatial AdaIN may be as follows:

Similar to AdaIN, SpatialAdaIN has no radiation parameters to learn and is an adaptive standardization method. In particular, the mean and standard deviation of SpatialAdaIN statistics are pixel-level, not channel-level. Pixel-level statistics can improve the complexity of the model, and can also better preserve the spatial information of the image, thereby better preserving the stroke structure of the text.

In a feasible embodiment, low-resolution sample text content information may also be used as the input of the preset text generation network. Starting from low-resolution text, the preset text generation network only needs to slightly adjust the position, thickness and other details of each stroke according to the style, and then the text of the corresponding style can be generated, which greatly reduces the difficulty of text generation.

In a feasible embodiment, the embodiment of the present application further provides a font quality detection model and a font similarity detection model, which are used to measure the font quality and innovation degree of the generated target text.

Exemplarily, the font quality detection model is mainly used to evaluate whether the font is complete and whether it is difficult to identify. Text and real text, negative samples are low-quality text generated by the model. The test shows that the model has a classification accuracy rate of 93%, and has strong quality discrimination ability. Further, in order to prevent the generated text from being too similar to the training text, a font similarity detection model can also be trained to evaluate the similarity between the font of the generated text and the font of the training text. The evaluation index can be Cosine Similarity. .

In a specific embodiment, a preset number of font styles may be selected, and each font contains a preset number of commonly used characters (contents) as a sample character set. And train the sample text set according to the above training process. FIG. 9 is a schematic diagram of a target text generated by using the text generation method according to an embodiment of the present application, according to an exemplary embodiment. As shown in FIG. 9 , characters with different styles can be generated by using the font style coding network and the character generation network provided by the embodiments of the present application, including square and straight characters, characters with abstract strokes and artistic effects, and characters with different thicknesses. text, and the same text can maintain its own style. Under the premise of ensuring the diversity of styles, the generated text has good integrity and high quality, and there is basically no problem of missing and merged strokes.

In a feasible embodiment, an empirical mode decomposition model (Empirical Mode Decomposition, EMD) may also be trained, and texts are generated through the EMD. Table 1 is a comparison table of the fusion effect between the text generated by EMD and the target text generated by the method in the embodiment of the present application. FIG. 10 is a comparison diagram of fusion effects according to an exemplary embodiment.

As shown in Table 1, the fusion effect of the target text generated by the method in the embodiment of the present application is better. Among them, FID is the abbreviation of Frechet Inception Distance score, and the Chinese name is distance score. Yield refers to the ratio between the generated text with better quality and the total generated text.

Table 1 Comparison of fusion effects

Model FID↓ Yield rate↑ EMD 32.66 63.30% this application 28.10 91.49%

As shown in Figure 10, the characters fused by EMD are prone to problems of incomplete fonts such as missing strokes and merging of strokes. For example, the word "True" in the first line is missing a horizontal line, while the words "Lian" and "Jie" have strokes. The merge in turn resulted in a typo. However, using the target subgraph generated by the embodiment of the present application, the problem of lack of strokes and merging rarely occurs, the structure of the text can be well preserved, typos rarely occur, and the quality of the font is greatly improved.

FIG. 11 is a schematic diagram of a similar font obtained by detecting the font of a target text by a font similarity detection model according to an exemplary embodiment. As shown in Figure 11, it can be seen from the first three lines that the font of the generated target text has obvious style differences with the font of the text in the training set (ie, the closest font and the next closest font in Figure 11). However, the font height of the text displayed in the fourth row is similar to the font height of the text in the font library, and the text can be automatically filtered out to filter out the inappropriate text, which further improves the quality of the generated target text.

In a feasible embodiment, as disclosed in the present application for the text generation method and the text generation network training method, font style information, text content information, etc. can be stored on the blockchain.

Fig. 12 shows a text generating apparatus according to an exemplary embodiment. As shown in Figure 12, the device may include at least:

The character acquisition module 501 may be configured to acquire at least two candidate characters with different font style information from the to-be-processed character set.

The font style information generating module 503 may be configured to generate font style information corresponding to each of the at least two candidate characters above based on the font style coding network.

The target text generation module 505 can be configured to generate target text corresponding to the font style information and text content information based on the text generation network; the text content information represents the content of the text in the to-be-processed text set.

Wherein, the above-mentioned font style coding network and text generation network are obtained by performing text generation training on a preset neural network, and the latent space of the sample font style information used in the text generation training process is constrained to be a normal distribution.

In an exemplary embodiment, the font style information generation module 503 can be configured to map the at least two candidate characters to the normal distribution based on the font style coding network, so as to obtain the corresponding correspondence between the at least two candidate characters. font style information.

In an exemplary embodiment, the above-mentioned apparatus may further include:

The target font style information determination module may be configured to calculate the average value of the above font style information based on the above font style coding network to obtain the target font style information.

Correspondingly, the target text generation module 505 may be configured to generate the target text corresponding to the target font style information and the text content information based on the text generation network.

In an exemplary embodiment, the above-mentioned apparatus may further include:

The first decoupling module can be used to decouple the above font style information.

The first standardization processing module can be used to standardize the decoupled font style information to obtain standard font style information.

The second standardization processing module may be configured to perform standardization processing on the above-mentioned text content information to obtain standard text content information.

Correspondingly, the target text generation module 505 may be configured to generate the target text corresponding to the standard font style information and the standard text content information based on the text generation network.

Fig. 13 shows a network training apparatus for text generation according to an exemplary embodiment. As shown in Figure 13, the device may also include:

The sample text acquisition module 601 is used for extracting the first sample text and the second sample text from the sample text set.

The training module 603 is configured to perform text generation training on the preset neural network based on the sample font style information of the above-mentioned first sample text and the sample text content information of the above-mentioned second sample text. The latent space constraint of the sample font style information is normal distribution, and the font style encoding network and the text generation network are obtained.

In an exemplary embodiment, the above-mentioned preset neural network includes a preset font style encoding network, a preset text generation network and a preset discrimination network, and the above-mentioned training module 603 may include:

The mapping unit may be configured to map the above-mentioned first sample text to the latest normal distribution based on the above-mentioned preset font style coding network to obtain current sample font style information.

The sample font style information generating unit may be configured to process the standard normal distribution and the current sample font style information based on the preset font style coding network to obtain the sample font style information.

The current text generation unit may be configured to generate the current text corresponding to the sample font style information and the sample text content information based on the preset text generation network.

The loss information determining unit may be configured to perform a discrimination process on the current text, the reference text, the sample font style information and the sample text content information based on the preset discrimination network to obtain loss information.

The network generation unit can be used to train the preset font style coding network and the preset text generation network based on the loss information, and obtain the font style coding network and the text generation network; the reference text represents the content information of the sample text in the above Text under sample font style information.

In an exemplary embodiment, the above-mentioned sample font style information generating unit may include:

The feature vector obtaining subunit may be configured to randomly obtain the feature vector corresponding to the font style information of the current sample from the standard normal distribution based on the preset font style coding network.

The update subunit can be used to determine the difference information between the above-mentioned feature vector and the above-mentioned current sample font style information based on the above-mentioned preset font style coding network, update the above-mentioned latest normal distribution based on the above-mentioned difference information, and update the updated latest font style information. The normal distribution is re-assigned as the latest normal distribution above.

The repeating subunit can be used to map the above-mentioned first sample text to the latest normal distribution based on the above-mentioned preset font style coding network to obtain the current sample font style information, and the above-mentioned updated latest normal distribution is re-assigned. The above-mentioned latest normal distribution is repeated until the above-mentioned difference information satisfies the preset condition.

The sample font style information determination subunit may be configured to use the current sample font style information when the difference information meets the preset condition based on the preset font style coding network as the sample font style information.

In an exemplary embodiment, the above-mentioned apparatus may further include:

The second decoupling module can be used to decouple the above-mentioned sample font style information.

The third standardization processing module can be used to standardize the decoupled sample font style information to obtain standard sample font style information.

The fourth standardization processing module may be configured to perform standardization processing on the above-mentioned sample text content information to obtain standard sample text content information.

Correspondingly, the above-mentioned current character generation unit may be configured to generate the above-mentioned current character corresponding to the above-mentioned standard sample text content information and the above-mentioned standard sample font style information based on the above-mentioned preset character generation network.

It should be noted that the apparatus embodiments provided in the embodiments of the present application and the above-mentioned method embodiments are based on the same inventive concept.

An embodiment of the present application further provides an electronic device, the electronic device includes a processor and a memory, the memory stores at least one instruction or at least one program, and the at least one instruction or at least one program is loaded and executed by the processor to achieve the above The text generation method or the network training method for text generation provided by any of the embodiments.

Embodiments of the present application further provide a computer-readable storage medium, which can be set in a terminal to store information about a text generation method or a network training method for text generation in the method embodiment. At least one instruction or at least one segment of program of the , at least one instruction or at least one segment of program is loaded and executed by the processor to implement the character generation method or the character generation network training method provided by the above method embodiments.

Optionally, in this embodiment of the present specification, the storage medium may be located in at least one network server among multiple network servers of a computer network. Optionally, in this embodiment, the above-mentioned storage medium may include but is not limited to: a U disk, a read-only memory (Read-Only Memory, ROM), a random access memory (Random Access Memory, RAM), a mobile hard disk, a magnetic Various media that can store program codes, such as a disc or an optical disc.

The memory in the embodiments of this specification can be used to store software programs and modules, and the processor executes various functional application programs and data processing by running the software programs and modules stored in the memory. The memory may mainly include a stored program area and a stored data area, wherein the stored program area may store the operating system, application programs required for functions, etc.; the stored data area may store data created according to the use of the device, and the like. Additionally, the memory may include high-speed random access memory, and may also include non-volatile memory, such as at least one magnetic disk storage device, flash memory device, or other volatile solid state storage device. Accordingly, the memory may also include a memory controller to provide processor access to the memory.

Embodiments of the present application also provide a computer program product or computer program, where the computer program product or computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the character generation method or the character generation network training method provided by the above method embodiments.

The embodiments of the character generation method or the character generation network training method provided by the embodiments of the present application may be executed in a terminal, a computer terminal, a server, or a similar computing device. Taking running on a server as an example, FIG. 14 is a hardware structural block diagram of a server for character generation or character generation network training provided by an embodiment of the present application. As shown in FIG. 14 , the server 700 may vary greatly due to different configurations or performances, and may include one or more central processing units (Central Processing Units, CPU) 710 (the central processing unit 710 may include, but is not limited to, a microcomputer). A processor MCU or a processing device such as a programmable logic device FPGA), a memory 730 for storing data, and one or more storage media 720 (eg, one or more mass storage devices) for storing application programs 723 or data 722. Among them, the memory 730 and the storage medium 720 may be short-term storage or persistent storage. The program stored in the storage medium 720 may include one or more modules, and each module may include a series of instructions to operate on the server. Furthermore, the central processing unit 710 may be configured to communicate with the storage medium 720 to execute a series of instruction operations in the storage medium 720 on the server 700 . Server 700 may also include one or more power supplies 760, one or more wired or wireless network interfaces 750, one or more input and output interfaces 740, and/or, one or more operating systems 721, such as Windows Server™, MacOS X™ , UnixTM, LinuxTM, FreeBSDTM and more.

Input-output interface 740 may be used to receive or transmit data via a network. The specific example of the above-mentioned network may include a wireless network provided by the communication provider of the server 700 . In one example, the I/O interface 740 includes a network adapter (Network Interface Controller, NIC), which can be connected to other network devices through the base station so as to communicate with the Internet. In one example, the input/output interface 740 may be a radio frequency (Radio Frequency, RF) module, which is used to communicate with the Internet in a wireless manner.

Those of ordinary skill in the art can understand that the structure shown in FIG. 14 is only a schematic diagram, which does not limit the structure of the above-mentioned electronic device. For example, server 700 may also include more or fewer components than those shown in FIG. 14 , or have a different configuration than that shown in FIG. 14 .

It should be noted that: the above-mentioned order of the embodiments of the present application is only for description, and does not represent the advantages and disadvantages of the embodiments. And the foregoing describes specific embodiments of the present specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims can be performed in an order different from that in the embodiments and still achieve desirable results. Additionally, the processes depicted in the figures do not necessarily require the particular order shown, or sequential order, to achieve desirable results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

Each embodiment in this specification is described in a progressive manner, and the same and similar parts between the various embodiments may be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the apparatus and server embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and for related parts, please refer to the partial descriptions of the method embodiments.

Those of ordinary skill in the art can understand that all or part of the steps of implementing the above embodiments can be completed by hardware, or can be completed by instructing relevant hardware through a program, and the program can be stored in a computer-readable storage medium. The storage medium can be read-only memory, magnetic disk or optical disk, etc.

The above are only preferred embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application. Inside.

Claims (13)

1. A method for generating words, the method comprising:

acquiring at least two candidate characters with different font style information from a character set to be processed;

generating font style information corresponding to the at least two candidate characters based on a font style coding network;

generating target characters corresponding to the font style information and the character content information based on a character generation network; the text content information represents the content of the text in the text set to be processed;

the font style coding network and the character generation network are obtained by performing character generation training on a preset neural network, and the hidden space of the sample font style information used in the character generation training process is constrained to be in normal distribution.

2. The method of claim 1, wherein generating font style information corresponding to each of the at least two candidate words based on the font style encoding network comprises:

and mapping the at least two candidate characters to the normal distribution based on the font style coding network to obtain font style information corresponding to the at least two candidate characters.

3. The method of claim 1, further comprising:

calculating the average value of the font style information based on the font style coding network to obtain target font style information;

generating the target characters corresponding to the font style information and the character content information based on the character generation network, wherein the generating comprises:

and generating the target characters corresponding to the target font style information and the character content information based on the character generation network.

4. The method according to any one of claims 1 to 3, further comprising:

decoupling said font style information;

standardizing the decoupled font style information to obtain standard font style information;

standardizing the text content information to obtain standard text content information;

correspondingly, the generating the target text corresponding to the font style information and the text content information based on the text generation network includes:

and generating the target characters corresponding to the standard font style information and the standard character content information based on the character generation network.

5. A network training method for generating characters is characterized by comprising the following steps:

extracting a first sample word and a second sample word from the sample word set;

and performing character generation training on a preset neural network based on the sample font style information of the first sample character and the sample character content information of the second sample character, and constraining the hidden space of the sample font style information into normal distribution in the character generation training process to obtain a font style coding network and a character generation network.

6. The method according to claim 5, wherein the preset neural network includes a preset font style coding network, a preset character generation network, and a preset discrimination network, the character generation training is performed on the preset neural network based on the sample font style information of the first sample character and the sample character content information of the second sample character, and in the character generation training process, the hidden space of the sample font style information is constrained to be normally distributed to obtain the font style coding network and the character generation network, including:

mapping the first sample font to the latest normal distribution based on the preset font style coding network to obtain the current sample font style information;

processing the standard normal distribution and the current sample font style information based on the preset font style coding network to obtain the sample font style information;

generating the current characters corresponding to the sample font style information and the sample character content information based on the preset character generation network;

based on the preset discrimination network, performing discrimination processing on the current characters, the reference characters, the sample font style information and the sample character content information to obtain loss information;

training the preset font style coding network and the preset character generation network based on the loss information to obtain the font style coding network and the character generation network; the reference characters represent characters of the sample character content information under the sample font style information.

7. The method according to claim 6, wherein the processing the standard normal distribution and the current sample font style information based on the preset font style encoding network to obtain the sample font style information comprises:

based on the preset font style coding network, randomly acquiring a feature vector corresponding to the font style information of the current sample from the standard normal distribution;

determining difference information between the feature vector and the current sample font style information based on the preset font style encoding network, updating the latest positive distribution based on the difference information, and taking the updated latest positive distribution as the latest positive distribution again;

based on the preset font style coding network, repeating the mapping of the first sample text word to the latest positive-too-distribution to obtain the current sample font style information, and the renewed latest positive-too-distribution as the latest positive-too-distribution until the difference information meets the preset condition;

and taking the current sample font style information when the difference information meets the preset condition as the sample font style information based on the preset font style coding network.

8. The method of claim 6, further comprising:

decoupling the sample font style information;

standardizing the decoupled sample font style information to obtain standard sample font style information;

standardizing the sample text content information to obtain standard sample text content information;

correspondingly, the generating the current text corresponding to the sample font style information and the sample text content information based on the preset text generation network includes:

and generating the current characters corresponding to the standard sample character content information and the standard sample font style information based on the preset character generation network.

9. A text generation apparatus, the apparatus comprising:

the character acquisition module is used for acquiring at least two candidate characters with different font style information from the character set to be processed;

the font style information generating module is used for generating font style information corresponding to the at least two candidate characters based on a font style coding network;

the target character generation module is used for generating target characters corresponding to the font style information and the character content information based on a character generation network; the text content information represents the content of the text in the text set to be processed;

the font style coding network and the character generation network are obtained by performing character generation training on a preset neural network, and the hidden space of the sample font style information used in the character generation training process is constrained to be in normal distribution.

10. A network training apparatus for character generation, the apparatus comprising:

the sample character acquisition module is used for extracting a first sample character and a second sample character from the sample character set;

and the training module is used for performing character generation training on a preset neural network based on the sample font style information of the first sample characters and the sample character content information of the second sample characters, and in the character generation training process, the hidden space of the sample font style information is restricted to normal distribution to obtain a font style coding network and a character generation network.

11. An electronic device, comprising a processor and a memory, wherein at least one instruction or at least one program is stored in the memory, and the at least one instruction or the at least one program is loaded by the processor and executed to implement the method for generating words according to any one of claims 1 to 4 or the method for training a network for generating words according to any one of claims 5 to 8.

12. A computer-readable storage medium, having at least one instruction or at least one program stored therein, which is loaded and executed by a processor to implement the method for generating words according to any one of claims 1 to 4 or the method for network training of word generation according to any one of claims 5 to 8.

13. A computer program product comprising a computer program, wherein the computer program, when executed by a processor, implements the method of generating a word according to any one of claims 1 to 4 and the method of network training for word generation according to any one of claims 5 to 8.

Images