Chinese character generation method and device, electronic equipment and storage medium

Abstract

The invention provides a Chinese character generation method and device, electronic equipment and a storage medium, relates to the technical field of artificial intelligence, and is suitable for the fields of financial science and technology and medical health. The method comprises the following steps: carrying out feature conversion on a sample random noise vector, and then carrying out hydraulic feature decomposition to obtain sample noise sub-hydraulic features; performing cross-scale feature coupling on the at least two sample noise sub-hydraulic features through an initial cross-scale coupler to obtain sample multi-scale hydraulic features, and further generating a sample random Chinese character image; performing image discrimination on the sample random Chinese character image to generate a pressure gradient signal, and updating parameters of the initial cross-scale coupler according to the pressure gradient signal to obtain a target cross-scale coupler; and finally, generating a target Chinese character in combination with the target cross-scale coupler. The Chinese characters of different structures can be adaptively generated, the Chinese character generation accuracy is improved, and particularly, the situation that the Chinese characters are broken or adhered can be reduced.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and is applicable to the fields of fintech and healthcare, and particularly relates to a Chinese character generation method and apparatus, an electronic device, and a storage medium. Background Art

[0002] The Chinese character generation technology is a technology that can automatically generate Chinese characters based on algorithms or rules. For example, a neural network (such as a generative adversarial network or a transformation model) is used to learn the features of Chinese characters, so as to generate new Chinese characters or simulate writing. The Chinese character generation technology can be applied to scenarios such as advertisement design and poster generation in the fintech field. For example, in the insurance business of the advertisement design scenario, through the Chinese character generation technology, advertisement pictures with specific styles (such as characters like "benefit", "purchase", "reduce", etc.) can be generated for insurance agents to market insurance products. The Chinese character generation technology can also be applied to scenarios such as medical label annotation and medical material generation in the healthcare field. For example, in the medical label annotation scenario, through the Chinese character generation technology, standardized and clear Chinese medical orders or drug instructions can be automatically generated to facilitate patients to obtain information.

[0003] Existing Chinese character generation technologies usually use generative adversarial networks. Although they can improve the naturalness and diversity of the generated fonts to a certain extent, there are still obvious deficiencies when dealing with high-resolution outputs and complex structures. For example, existing Chinese character generation technologies are prone to breakage or adhesion when generating complex structures (such as characters like "鬱" and "龘").

[0004] Therefore, there is a problem of low accuracy in Chinese character generation in related technologies. Summary of the Invention

[0005] The main purpose of the embodiments of this application is to propose a Chinese character generation method and apparatus, an electronic device, and a storage medium, which can adaptively generate Chinese characters with different structures, improve the accuracy of Chinese character generation, and in particular, can reduce the occurrence of breakage or adhesion of Chinese characters.

[0006] To achieve the above object, in the first aspect of the embodiments of this application, a Chinese character generation method is proposed, and the method includes:

[0007] Obtain a sample random noise vector, and perform feature transformation on the sample random noise vector to obtain a sample noise feature map;

[0008] Perform hydraulic feature decomposition on the sample noise feature map according to at least two preset pressure decomposition parameters respectively to obtain sample noise sub-hydraulic features corresponding to each pressure decomposition parameter; where any two of the pressure decomposition parameters are different;

[0009] Performing cross-scale feature coupling on at least two sample noise sub-hydraulic features through a preset initial cross-scale coupler to obtain a sample multi-scale hydraulic feature;

[0010] Performing image generation on the multi-scale hydraulic features of the sample to obtain a sample random Chinese character image;

[0011] Performing multi-scale image discrimination on the sample random Chinese character images to obtain comprehensive evaluation data of image authenticity;

[0012] generating a pressure gradient signal according to the comprehensive image authenticity evaluation data, and updating the parameters of the initial cross-scale coupler according to the pressure gradient signal to obtain a target cross-scale coupler;

[0013] The target cross-scale coupler performs cross-scale feature coupling on at least two target noise sub-features corresponding to a preset target noise vector to obtain a target noise multi-scale feature, and performs image generation on the target noise multi-scale feature to obtain a target Chinese character.

[0014] Optionally, the at least two pressure decomposition parameters include a second pressure decomposition parameter of the first pressure decomposition parameter, and the first pressure decomposition parameter is smaller than the second pressure decomposition parameter;

[0015] The step of performing hydraulic feature decomposition on the sample noise feature map according to at least two preset pressure decomposition parameters to obtain a sample noise sub-hydraulic feature of each pressure decomposition parameter includes:

[0016] Performing a 1×1 convolution on the sample noise feature map to obtain a first sample convolution sub-feature;

[0017] performing nonlinear mapping on the first pressure decomposition parameter to obtain a first sample hydraulic mapping feature;

[0018] multiplying the first sample convolution sub-feature and the first sample hydraulic mapping feature element by element to obtain the sample noise sub-hydraulic feature corresponding to the first pressure decomposition parameter;

[0019] Performing a 1×1 convolution on the sample noise sub-hydraulic feature corresponding to the first pressure decomposition parameter to obtain a second sample convolution sub-feature;

[0020] performing nonlinear mapping on the second pressure decomposition parameter to obtain a second sample hydraulic mapping feature;

[0021] The second sample convolution sub-feature and the second sample hydraulic mapping feature are element-wise multiplied to obtain the sample noise sub-hydraulic feature corresponding to the second pressure decomposition parameter.

[0022] Optionally, performing nonlinear mapping on the first pressure decomposition parameter to obtain a first sample hydraulic mapping feature includes:

[0023] Multiplying a preset hydraulic decomposition weight matrix by the first pressure decomposition parameter to obtain a first sample hydraulic mapping vector;

[0024] Adding the first sample hydraulic mapping vector to a preset bias vector to obtain a first sample hydraulic bias vector;

[0025] Feature activation is performed on the first sample hydraulic pressure offset vector to obtain the first sample hydraulic pressure mapping feature.

[0026] Optionally, the image authenticity comprehensive assessment data includes a first scale-specific error vector corresponding to the first pressure decomposition parameter and a second scale-specific error vector corresponding to the second pressure decomposition parameter;

[0027] After performing multi-scale image discrimination on the sample random Chinese character image to obtain comprehensive image authenticity evaluation data, the method further includes:

[0028] Updating the first pressure decomposition parameter and the second pressure decomposition parameter specifically includes:

[0029] multiplying the first scale-specific error vector by a preset error projection matrix to obtain a first initial pressure adjustment amount;

[0030] Limiting the amplitude of the first initial pressure adjustment amount to obtain a first target pressure adjustment amount;

[0031] obtaining an updated first pressure decomposition parameter by adding the first target pressure adjustment amount to the first pressure decomposition parameter;

[0032] multiplying the second scale-specific error vector by a preset error projection matrix to obtain a second initial pressure adjustment amount;

[0033] limiting the amplitude of the second initial pressure adjustment amount to obtain a second target pressure adjustment amount;

[0034] The second target pressure adjustment amount is added to the second pressure decomposition parameter to obtain an updated second pressure decomposition parameter.

[0035] Optionally, performing cross-scale feature coupling on at least two sample noise sub-hydraulic features through a preset initial cross-scale coupler to obtain a sample multi-scale hydraulic feature includes:

[0036] performing average pooling on the sample noise sub-hydraulic features corresponding to the first pressure decomposition parameter to obtain a first sample pooled sub-hydraulic feature;

[0037] Calculating the difference between the second pressure decomposition parameter and the first pressure decomposition parameter to obtain a hierarchical pressure gradient, and converting the hierarchical pressure gradient through a preset pressure difference conversion network to obtain a first sample pressure difference conversion feature;

[0038] multiplying the first sample pooling sub-hydraulic feature by the first sample pressure difference conversion feature element-by-element to obtain a hydraulic information transmission feature;

[0039] performing feature fusion based on the hydraulic information transmission feature and the sample noise sub-hydraulic feature corresponding to the second pressure decomposition parameter to obtain an updated sample noise sub-hydraulic feature corresponding to the second pressure decomposition parameter;

[0040] Feature fusion is performed on the sample noise sub-hydraulic feature corresponding to the first pressure decomposition parameter and the updated sample noise sub-hydraulic feature corresponding to the second pressure decomposition parameter to obtain a sample multi-scale hydraulic feature.

[0041] Optionally, performing multi-scale image discrimination on the sample random Chinese character image to obtain comprehensive image authenticity evaluation data includes:

[0042] Performing multi-scale downsampling on the sample random Chinese character image to obtain at least two Chinese character images to be evaluated; wherein the at least two Chinese character images to be evaluated include a first Chinese character image to be evaluated at a first scale and a second Chinese character image to be evaluated at a second scale, wherein the first scale is smaller than the second scale;

[0043] Calculating the image proportions of the first Chinese character image to be evaluated and the second Chinese character image to be evaluated to obtain a first attention weight coefficient of the first Chinese character image to be evaluated and a second attention weight coefficient of the second Chinese character image to be evaluated;

[0044] Obtaining a first sub-discriminator corresponding to the first scale, and performing image discrimination on the first Chinese character image to be evaluated using the first sub-discriminator to obtain first image authenticity discrimination data;

[0045] performing multiplication processing on the first attention weight coefficient and the first image authenticity discrimination data to obtain first image authenticity sub-assessment data;

[0046] Obtaining a second sub-discriminator corresponding to the second scale, and performing image discrimination on the second Chinese character image to be evaluated using the second sub-discriminator to obtain second image authenticity discrimination data;

[0047] performing multiplication processing on the second attention weight coefficient and the second image authenticity discrimination data to obtain second image authenticity sub-assessment data;

[0048] Data fusion is performed based on the first image authenticity sub-assessment data and the second image authenticity sub-assessment data to obtain the image authenticity comprehensive assessment data.

[0049] Optionally, the image proportion calculation is performed based on the first Chinese character image to be evaluated and the second Chinese character image to be evaluated to obtain a first attention weight coefficient of the first Chinese character image to be evaluated and a second attention weight coefficient of the second Chinese character image to be evaluated, respectively, including:

[0050] Performing global average pooling on the first Chinese character image to be evaluated to obtain a first image pooling vector;

[0051] Performing a vector inner product calculation on the preset Chinese character direction vector and the first image pooling vector to obtain a first initial attention score;

[0052] Performing exponential mapping on the first initial attention score to obtain a first target attention score;

[0053] Performing global average pooling on the second Chinese character image to be evaluated to obtain a second image pooling vector;

[0054] Performing a vector inner product calculation on the preset Chinese character direction vector and the second image pooling vector to obtain a second initial attention score;

[0055] Performing exponential mapping on the second initial attention score to obtain a second target attention score;

[0056] Adding the first target attention score and the second target attention score to obtain a total attention score;

[0057] Calculate the ratio of the first target attention score to the total attention score to obtain the first attention weight coefficient;

[0058] The second target attention score is ratioed to the total attention score to obtain the second attention weight coefficient.

[0059] To achieve the above-mentioned purpose, a second aspect of the embodiments of the present application provides a Chinese character generation device, the device comprising:

[0060] A noise acquisition module is used to acquire a sample random noise vector and perform feature conversion on the sample random noise vector to obtain a sample noise feature map;

[0061] a feature decomposition module, configured to perform hydraulic feature decomposition on the sample noise feature map according to at least two preset pressure decomposition parameters, to obtain a sub-hydraulic feature of the sample noise for each of the pressure decomposition parameters; wherein any two of the pressure decomposition parameters are different;

[0062] a cross-scale coupling module, configured to perform cross-scale feature coupling on at least two sample noise sub-hydraulic features through a preset initial cross-scale coupler to obtain a sample multi-scale hydraulic feature;

[0063] An image generation module is used to generate an image of the multi-scale hydraulic features of the sample to obtain a sample random Chinese character image;

[0064] An image discrimination module is used to perform multi-scale image discrimination on the sample random Chinese character images to obtain comprehensive evaluation data of image authenticity;

[0065] a parameter updating module, configured to generate a pressure gradient signal according to the comprehensive image authenticity evaluation data, and update the parameters of the initial cross-scale coupler according to the pressure gradient signal to obtain a target cross-scale coupler;

[0066] The Chinese character generation module is used to perform cross-scale feature coupling on at least two target noise sub-features corresponding to a preset target noise vector through the target cross-scale coupler to obtain a target noise multi-scale feature, and perform image generation on the target noise multi-scale feature to obtain a target Chinese character.

[0067] To achieve the above-mentioned purpose, the third aspect of an embodiment of the present application proposes an electronic device, which includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, it implements the Chinese character generation method described in the first aspect.

[0068] To achieve the above-mentioned purpose, the fourth aspect of an embodiment of the present application proposes a storage medium, which is a computer-readable storage medium and stores a computer program. When the computer program is executed by a processor, it implements the Chinese character generation method described in the first aspect above.

[0069] The Chinese character generation method and device, electronic device, and storage medium proposed in this application first obtain a sample random noise vector, which randomly points to a character; then perform feature conversion on the sample random noise vector to obtain a sample noise feature map, which is used to simulate the feature map corresponding to a Chinese character; further, perform hydraulic feature decomposition on the sample noise feature map to obtain sample noise sub-hydraulic features corresponding to multiple pressure decomposition parameters. This step mainly simulates the characteristics of different pressure levels in the hydraulic system to achieve hydraulic feature decomposition, which can fully decompose the structural information of the Chinese character structure at different levels; further, perform cross-scale feature coupling on at least two sample noise sub-hydraulic features through an initial cross-scale coupler to obtain sample multi-scale hydraulic features. This step draws on the principle of pressure transmission in fluid mechanics, and can improve the representation ability of the sample multi-scale hydraulic features on the structural information of Chinese characters; further, the sample multi-scale hydraulic features are image generated to obtain a sample random Chinese character image, and the sample random Chinese character image is subjected to multi-scale image discrimination to obtain image authenticity comprehensive evaluation data, and then a pressure gradient signal is generated based on the image authenticity comprehensive evaluation data, and then the parameters of the initial cross-scale coupler are updated to obtain a target cross-scale coupler; in this way, the initial cross-scale coupler is dynamically updated by the pressure gradient signal to obtain a target cross-scale coupler, which can improve the adaptability to Chinese characters of different structures; finally, in actual use, the target Chinese characters are generated by the trained target cross-scale coupler. In summary, the present application can adaptively generate Chinese characters of different structures, improve the accuracy of Chinese character generation, and especially reduce the situation where Chinese characters appear broken or sticky.

[0070] Additional aspects and advantages of the present application will be given in part in the description below, and in part will become obvious from the description below, or will be learned through practice of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 This is a flow chart of the Chinese character generation method provided in an embodiment of the present application;

[0072] Figure 2 yes Figure 1 Flowchart of step 104 in FIG.

[0073] Figure 3 yes Figure 1 Flowchart of step 103 in FIG.

[0074] Figure 4 yes Figure 1 Flowchart of step 105 in FIG.

[0075] Figure 5 yes Figure 4 Flowchart of step 402 in FIG.

[0076] Figure 6 is a flow chart of a Chinese character generation method provided by another embodiment of the present application;

[0077] Figure 7 This is a block diagram of the module structure of the Chinese character generation device provided in an embodiment of the present application;

[0078] Figure 8 This is a schematic diagram of the hardware structure of the electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0079] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0080] It should be noted that although the device schematics illustrate functional module divisions and the flowcharts illustrate logical sequences, in certain circumstances, the steps shown or described may be performed in a sequence that differs from the module divisions in the device or the sequence in the flowcharts. The terms "first," "second," and so on, in the specification, claims, and drawings, are used to distinguish similar items and are not necessarily used to describe a specific sequence or precedence.

[0081] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this application pertains. The terms used herein are for the purpose of describing the embodiments of this application only and are not intended to limit this application.

[0082] First, let’s analyze some of the terms used in this application:

[0083] Artificial intelligence (AI) is a new technical discipline that studies and develops theories, methods, technologies, and application systems for simulating, extending, and expanding human intelligence. A branch of computer science, AI seeks to understand the essence of intelligence and produce new intelligent machines that can respond in a manner similar to human intelligence. Research in this field includes robotics, speech recognition, image recognition, natural language processing, and expert systems. AI can simulate the information processes of human consciousness and thinking. It also encompasses the theories, methods, technologies, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, to perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results.

[0084] Natural language processing (NLP): NLP uses computers to process, understand, and apply human languages ​​(such as Chinese and English). A branch of artificial intelligence, NLP is an interdisciplinary field between computer science and linguistics, often referred to as computational linguistics. Natural language processing encompasses grammatical analysis, semantic analysis, and discourse comprehension. Natural language processing is commonly used in technical fields such as machine translation, handwritten and printed character recognition, speech recognition and text-to-speech conversion, information and image processing, information extraction and filtering, text classification and clustering, public opinion analysis, and opinion mining. It encompasses data mining, machine learning, knowledge acquisition, knowledge engineering, artificial intelligence research related to language processing, and linguistics research related to language computing.

[0085] Generative Adversarial Network (GAN): This consists of a generative model G and a discriminative model D. The goal of a GAN is to learn the distribution of training data. To learn this distribution, an input noise variable is first defined and then mapped to the data space. Here, G is a generative model composed of a multilayer perceptron network with parameters. Furthermore, a discriminative model is defined to determine whether the input data comes from the generative model or the training data. The output of D is the probability that x is the training data. Finally, D is trained to accurately determine the data source, and G is trained to generate data that conforms to the distribution of the training data as closely as possible.

[0086] Existing Chinese character generation techniques typically utilize generative adversarial networks (GANs), such as the stroke-decomposition-based CharacterGAN and improved methods incorporating residual networks. While these GANs can improve the naturalness and diversity of generated fonts, they still have significant limitations when handling high-resolution output and complex structures. In the field of image generation, multi-scale GANs achieve coarse-to-fine feature synthesis through a cascaded network structure, while wavelet-based methods enhance detail preservation through frequency domain decomposition. In particular, the successful application of the composite multi-scale basic scale entropy algorithm in hydraulic system fault diagnosis demonstrates its effectiveness for multi-level feature extraction. This algorithm separates complex signals into distinct frequency components through adaptive modal decomposition, making it well-suited for handling the hierarchical structural features of Chinese characters. Furthermore, the attention mechanism significantly enhances the GAN discriminator's sensitivity to local features through dynamic weight allocation, enabling the model to focus on key areas such as stroke intersections and transitions. Hybrid loss function designs, such as combining an adversarial loss with a reconstruction loss, further optimize generation quality, balancing visual realism and structural accuracy.

[0087] There are three main limitations of existing technologies:

[0088] (1) Insufficient cross-scale feature coupling: Traditional multi-scale GANs adopt a fixed-weight fusion strategy, making it difficult to adaptively adjust the dependency relationships between features of different scales (resolutions), resulting in structural distortion when the generated characters are scaled. Especially when dealing with complex Chinese characters containing multiple levels of details, it is difficult to balance the high-frequency details and low-frequency structures;

[0089] (2) Degradation of local details: Feature decomposition methods based on a single entropy value cannot accurately capture the topological constraints between Chinese character strokes, and are prone to breakage or adhesion during the generation of complex structures (such as characters like "鬱" and "龘"). This ignores the unique structural rules of Chinese characters, such as the integrity of radicals and the spatial relationships between strokes;

[0090] (3) Lack of dynamic adaptability: Existing methods lack a pressure adaptive mechanism similar to that of a hydraulic system, resulting in the inability to dynamically adjust the feature extraction intensity according to the character complexity during the training process. Using the same processing strategy for all Chinese characters, it is impossible to optimize the parameter configuration for simple and complex glyphs separately, leading to unreasonable resource allocation.

[0091] Based on this, the embodiments of this application propose a Chinese character generation method, a Chinese character generation device, an electronic device, and a computer-readable storage medium. By combining with physical principles and starting from the composition structure of Chinese characters, pressure decomposition parameters are introduced by the user to achieve Chinese character generation, which can adaptively generate Chinese characters of different structures, improve the accuracy of Chinese character generation, and especially reduce the occurrence of breakage or adhesion of Chinese characters.

[0092] The Chinese character generation method provided by the embodiments of this application can be applied to terminals and server sides, and can also be software running on the server side. The server side can be configured as an independent physical server, or as a server cluster or distributed system composed of multiple physical servers, or as a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms; the software can be an application that implements the Chinese character generation method, etc., but is not limited to the above forms.

[0093] The present application can be used in many general or special computer system environments or configurations. For example: server computers, multi-processor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics devices, network PCs, distributed computing environments including any of the above systems or devices, and the like. The present application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, and the like that perform specific tasks or implement specific abstract data types. The present application can also be practiced in distributed computing environments in which tasks are performed by remote processing devices connected via a communication network. In a distributed computing environment, program modules can be located in local and remote computer storage media, including storage devices.

[0094] The embodiments of the present application provide a Chinese character generation method, a Chinese character generation device, an electronic device, and a computer-readable storage medium, which are specifically illustrated by the following embodiments. First, the Chinese character generation method in the embodiments of the present application is described.

[0095] It should be noted that in each specific embodiment of the present application, when it comes to processing data related to the user's identity or characteristics such as the user's image data (such as ID images, etc.), the user's permission or consent must be obtained first, and the collection, use and processing of this data will comply with relevant laws, regulations and standards.

[0096] Reference Figure 1 , Figure 1 This is an optional flowchart of the Chinese character generation method provided in an embodiment of the present application, which may include but is not limited to steps 101 to 107.

[0097] Step 101: Obtain a sample random noise vector and perform feature conversion on the sample random noise vector to obtain a sample noise feature map;

[0098] Step 102: performing hydraulic feature decomposition on the sample noise feature map according to at least two preset pressure decomposition parameters to obtain a sub-hydraulic feature of the sample noise corresponding to each pressure decomposition parameter;

[0099] Step 103: performing cross-scale feature coupling on at least two sample noise sub-hydraulic features through a preset initial cross-scale coupler to obtain a sample multi-scale hydraulic feature;

[0100] Step 104, performing image generation on the sample multi-scale hydraulic features to obtain a sample random Chinese character image;

[0101] Step 105: Perform multi-scale image discrimination on the sample random Chinese character image to obtain comprehensive evaluation data of image authenticity;

[0102] Step 106: Generate a pressure gradient signal according to the comprehensive evaluation data of image authenticity, and update the parameters of the initial cross-scale coupler according to the pressure gradient signal to obtain a target cross-scale coupler;

[0103] Step 107: Perform cross-scale feature coupling on at least two target noise sub-features corresponding to a preset target noise vector through the target cross-scale coupler to obtain a target noise multi-scale feature, and perform image generation on the target noise multi-scale feature to obtain a target Chinese character.

[0104] Steps 101 to 107 illustrated in the embodiments of the present application: First, obtain a sample random noise vector, which randomly points to a character; then perform feature transformation on the sample random noise vector to obtain a sample noise feature map, which is used to simulate the feature map corresponding to a Chinese character; further, perform hydraulic feature decomposition on the sample noise feature map to obtain sample noise sub-hydraulic features corresponding to multiple pressure decomposition parameters. This step mainly simulates the characteristics of different pressure levels in the hydraulic system to achieve hydraulic feature decomposition, and can fully decompose the structural information of the Chinese character structure at different levels; further, perform cross-scale feature coupling on at least two sample noise sub-hydraulic features through the initial cross-scale coupler to obtain a sample multi-scale hydraulic feature. This step draws on the principle of pressure transmission in fluid mechanics and can improve the representation ability of the sample multi-scale hydraulic feature for the Chinese character structure information; further, perform image generation on the sample multi-scale hydraulic feature to obtain a sample random Chinese character image, and perform multi-scale image discrimination on the sample random Chinese character image to obtain comprehensive evaluation data of image authenticity. Then, generate a pressure gradient signal based on the comprehensive evaluation data of image authenticity, and then update the parameters of the initial cross-scale coupler to obtain a target cross-scale coupler; in this way, dynamically updating the initial cross-scale coupler through the pressure gradient signal to obtain a target cross-scale coupler can improve the adaptability to Chinese characters with different structures; finally, in actual use, generate a target Chinese character through the trained target cross-scale coupler. In summary, the present application can adaptively generate Chinese characters with different structures, improve the accuracy of Chinese character generation, and especially can reduce the occurrence of breakage or adhesion of Chinese characters.

[0105] In an example, in the advertising design scenario in the field of fintech, when an insurance agent expects to design a copywriting advertisement picture (such as a picture including characters like "favorable", "purchase", "discount", etc.) for the marketing of insurance products (such as auto insurance, health insurance, etc.), the above Chinese character generation method can be called to achieve Chinese character generation, so as to obtain a copywriting advertisement picture with a specific marketing style, which is convenient for insurance agents to market insurance products.

[0106] In another example, in a poster generation scenario in the fintech field, a user might want to create a promotional poster for a new energy vehicle. The aforementioned Chinese character generation method can be used to generate a poster with text for the vehicle (e.g., a poster containing words like "long range of 900 kilometers" and "fully intelligent cockpit" in a specific font style), making it easier for sales staff to promote the new energy vehicle.

[0107] In another example, in the healthcare field, a doctor might want to remind a patient to take medication on time daily, such as "take the medicine after dinner every day." By incorporating the aforementioned Chinese character generation technology, a standardized and clear image of the medical instruction "take the medicine after dinner every day" can be generated. This image can then be attached to the medication, making it easier for the patient to understand the information.

[0108] It should be noted that the above-mentioned Chinese character generation method can also be applied to scenarios such as medical material generation in the medical and health field. The specific generation process is similar to the medical labeling scenario and will not be repeated here.

[0109] In step 101 of some embodiments, a sample random noise vector is obtained, and feature conversion is performed on the sample random noise vector to obtain a sample noise feature map. 128 (a vector in the 128-dimensional real space, ∈ indicates belonging) is converted into a sample noise feature map F0∈R through a fully connected layer 8×8×512 (8×8 size, 512-channel feature map).

[0110] In step 102 of some embodiments, the sample noise feature map is subjected to hydraulic feature decomposition according to at least two preset pressure decomposition parameters to obtain the sample noise sub-hydraulic feature corresponding to each pressure decomposition parameter. Wherein, any two pressure decomposition parameters are different. For example, the sample noise feature map F0 is decomposed into four sample noise sub-hydraulic features by the CMBSE decomposition module. (k from 1 to 4 F k The sample noise sub-hydraulic characteristics can also be called intrinsic mode functions (IMFs). Each IMF corresponds to a pressure decomposition parameter P k ∈[0.2,0.8] (the value range is 0.2 to 0.8).

[0111] In one embodiment, the at least two pressure decomposition parameters include a second pressure decomposition parameter of a first pressure decomposition parameter, and the first pressure decomposition parameter is smaller than the second pressure decomposition parameter. Figure 2 , step 104 may include:

[0112] Step 201: Perform 1×1 convolution on the sample noise feature map to obtain a first sample convolution sub-feature;

[0113] Step 202 , performing nonlinear mapping on the first pressure decomposition parameter to obtain a first sample hydraulic mapping feature;

[0114] Step 203: multiply the first sample convolution sub-feature and the first sample hydraulic mapping feature element-by-element to obtain a sample noise sub-hydraulic feature corresponding to the first pressure decomposition parameter;

[0115] Step 204 , performing a 1×1 convolution on the sample noise sub-hydraulic feature corresponding to the first pressure decomposition parameter to obtain a second sample convolution sub-feature;

[0116] Step 205 , performing nonlinear mapping on the second pressure decomposition parameter to obtain a second sample hydraulic mapping feature;

[0117] Step 206 : Multiply the second sample convolution sub-feature and the second sample hydraulic mapping feature element-by-element to obtain a sample noise sub-hydraulic feature corresponding to the second pressure decomposition parameter.

[0118] In step 201, 1×1 convolution is mainly used for channel-dimensional feature transformation.

[0119] In step 202 , nonlinear mapping is mainly used to implement mapping from pressure to feature space.

[0120] In one embodiment, step 202 may include: multiplying a preset hydraulic decomposition weight matrix by a first pressure decomposition parameter to obtain a first sample hydraulic mapping vector; adding the first sample hydraulic mapping vector to a preset bias vector to obtain a first sample hydraulic bias vector; and performing feature activation on the first sample hydraulic bias vector to obtain a first sample hydraulic mapping feature.

[0121] Specifically, this application draws inspiration from the field of hydraulic system fault diagnosis and customizes a feature decomposition strategy for the Chinese character generation task. This application receives a sample noise feature map F∈R H×W×C (characteristic tensor of height H, width W, and number of channels C), by simulating the characteristics of different pressure decomposition parameters in the hydraulic system, the characteristics are decomposed into multiple intrinsic mode functions. Specifically, the calculation process of the kth IMF is:

[0122] F k =Conv (1×1)(F) ⊙σ(W p P k +b p ),

[0123] Among them, Conv 1×1 is a 1×1 convolution kernel, used for channel dimension feature transformation; P k is the k-th level pressure decomposition parameter, and its initial value is set in an arithmetic sequence (P1 = 0.2, P4 = 0.8); Wp and b p is the learnable hydraulic decomposition weight matrix and bias vector, which realize the nonlinear mapping of pressure to feature space; σ is the Sigmoid activation function, which constrains the pressure value to the range of [0,1]; ⊙ represents the Hadamard product (element-wise multiplication).

[0124] The benefit of the embodiment of the above steps 201 to 206 is that, by customizing the IMF decomposition strategy for Chinese character generation, the unique topological structure of Chinese characters is effectively maintained, the adverse effects of complex structures on Chinese character generation are reduced, and the quality of Chinese character generation is improved.

[0125] In step 103 of some embodiments, a preset initial cross-scale coupler is used to perform cross-scale feature coupling on at least two sample noise sub-hydraulic features to obtain a sample multi-scale hydraulic feature. The initial cross-scale coupler is a neural network that simulates the key component of the hydraulic system's pressure transmission mechanism and is responsible for the flow and fusion of information between IMFs at different scales. Drawing on the principles of pressure transmission in fluid mechanics, the initial cross-scale coupler controls the flow of feature information between different resolutions through adaptive coupling coefficients.

[0126] In one embodiment, referring to Figure 3 , step 103 may include:

[0127] Step 301 , performing average pooling on the sample noise sub-hydraulic features corresponding to the first pressure decomposition parameter to obtain a first sample pooled sub-hydraulic feature;

[0128] Step 302 , performing a difference calculation on the second pressure decomposition parameter and the first pressure decomposition parameter to obtain a hierarchical pressure gradient, and converting the hierarchical pressure gradient through a preset pressure difference conversion network to obtain a first sample pressure difference conversion feature;

[0129] Step 303 , performing element-wise multiplication of the first sample pooled sub-hydraulic feature and the first sample pressure difference conversion feature to obtain a hydraulic information transmission feature;

[0130] Step 304 , performing feature fusion based on the hydraulic information transmission feature and the sample noise sub-hydraulic feature corresponding to the second pressure decomposition parameter to obtain an updated sample noise sub-hydraulic feature corresponding to the second pressure decomposition parameter;

[0131] Step 305 : performing feature fusion on the sample noise sub-hydraulic features corresponding to the first pressure decomposition parameter and the updated sample noise sub-hydraulic features corresponding to the second pressure decomposition parameter to obtain a sample multi-scale hydraulic feature.

[0132] Specifically, information transmission between adjacent IMFs is achieved through the following process:

[0133]

[0134] Among them, ← represents assignment, AvgPool is a 2×2 average pooling operation, reducing the resolution to one-quarter of the original; f θ It is a pressure difference conversion network composed of 3 layers of multi-layer perceptron (MLP), with parameters θ; ΔP k→k+1 =P k+1 –P k It is the hierarchical pressure gradient that controls the intensity of information flow.

[0135] The benefit of the embodiment of steps 301 to 305 above is that the pressure transmission principle of the hydraulic system is introduced into the GAN architecture. By corresponding the pressure decomposition parameters to different feature scales, an innovative fusion of physical mechanism and deep learning is achieved, which significantly improves the generation quality of complex Chinese characters.

[0136] In step 104 of some embodiments, an image generator may be used to generate an image of the sample multi-scale hydraulic features to obtain a sample random Chinese character image. For example, the image generator may perform a three-layer transposed convolution on the sample multi-scale hydraulic features to obtain a sample random Chinese character image, and the image resolution of the image may be 64×64 or 128×128 pixels.

[0137] In step 105 of some embodiments, multi-scale image discrimination is performed on the sample random Chinese character image to obtain comprehensive evaluation data of image authenticity. The discriminator adopts a multi-scale pyramid structure and downsamples the input image to (s from 1 to 3 I s Set, representing the original image, 1 / 2 resolution, and 1 / 4 resolution respectively), and the weight α of each scale is calculated through the attention mechanism s (scalar weight coefficient) and weighted fusion discrimination results, and finally output the comprehensive evaluation data of image authenticity, which includes the authenticity probability and scale-specific error signal. The scale-specific error signal is a high-dimensional error vector.

[0138] In one embodiment, referring to Figure 4 , step 105 may include:

[0139] Step 401: Perform multi-scale downsampling on the sample random Chinese character image to obtain at least two Chinese character images to be evaluated;

[0140] Step 402: Calculate the image proportions based on the first Chinese character image to be evaluated and the second Chinese character image to be evaluated, and obtain a first attention weight coefficient for the first Chinese character image to be evaluated and a second attention weight coefficient for the second Chinese character image to be evaluated;

[0141] Step 403: Obtain a first sub-discriminator corresponding to the first scale, and use the first sub-discriminator to perform image discrimination on the first Chinese character image to be evaluated to obtain first image authenticity discrimination data;

[0142] Step 404: multiplying the first attention weight coefficient and the first image authenticity discrimination data to obtain first image authenticity sub-assessment data;

[0143] Step 405: Obtain a second sub-discriminator corresponding to the second scale, and perform image discrimination on the second Chinese character image to be evaluated using the second sub-discriminator to obtain authenticity discrimination data of the second image;

[0144] Step 406 , performing multiplication processing on the second attention weight coefficient and the second image authenticity discrimination data to obtain second image authenticity sub-assessment data;

[0145] Step 407 : performing data fusion based on the first image authenticity sub-assessment data and the second image authenticity sub-assessment data to obtain comprehensive image authenticity assessment data.

[0146] In step 401, the at least two Chinese character images to be evaluated may include a first Chinese character image to be evaluated at a first scale and a second Chinese character image to be evaluated at a second scale, where the first scale is smaller than the second scale. The at least two Chinese character images to be evaluated may also include the first Chinese character image to be evaluated at the first scale, the second Chinese character image to be evaluated at the second scale, and a third Chinese character image to be evaluated at a third scale, where the first scale is smaller than the second scale, and the second scale is smaller than the third scale.

[0147] In step 402, the attention weight coefficients corresponding to the first Chinese character image to be evaluated and the second Chinese character image to be evaluated can be calculated in combination with the attention mechanism.

[0148] In one embodiment, referring to Figure 5 , step 402 may include:

[0149] Step 501: performing global average pooling on a first Chinese character image to be evaluated to obtain a first image pooling vector;

[0150] Step 502: performing a vector inner product calculation on the preset Chinese character direction vector and the first image pooling vector to obtain a first initial attention score;

[0151] Step 503: Perform exponential mapping on the first initial attention score to obtain a first target attention score;

[0152] Step 504: performing global average pooling on the second Chinese character image to be evaluated to obtain a second image pooling vector;

[0153] Step 505: performing a vector inner product calculation on the preset Chinese character direction vector and the second image pooling vector to obtain a second initial attention score;

[0154] Step 506, performing exponential mapping on the second initial attention score to obtain a second target attention score;

[0155] Step 507, adding the first target attention score and the second target attention score to obtain a total attention score;

[0156] Step 508: Calculate the ratio of the first target attention score to the total attention score to obtain a first attention weight coefficient;

[0157] Step 509: Calculate the ratio of the second target attention score to the total attention score to obtain a second attention weight coefficient.

[0158] Calculation of the attention weight coefficient of the s-th scale:

[0159]

[0160] Among them, GAP represents the global average pooling operation (Global Average Pooling); v is the learnable Chinese character direction vector; v T Represents the transpose of v, used for vector inner product calculation; I s I represents the input image at the s-th scale, corresponding to the original resolution (s=1), 1 / 2 resolution (s=2) or 1 / 4 resolution (s=3); j Represents the input image at the scale corresponding to the sum variable j.

[0161] The benefit of the embodiment of the above steps 501 to 509 is that it can dynamically generate attention weight coefficients for images at different scales, which is conducive to automatically focusing on the most discriminative resolution level according to the characteristics of different Chinese characters during the image recognition process, thereby improving the recognition accuracy.

[0162] In one example, the final discrimination result is obtained by the sub-discriminators D s Weighted to get:

[0163]

[0164] Among them, D s represents the sub-discriminator of the s-th scale, which is dedicated to evaluating image features at a specific resolution.

[0165] The benefit of the embodiment of the above steps 401 to 407 is that, during the image recognition process, the most discriminative resolution level can be automatically focused on according to the characteristics of different Chinese characters, thereby improving the recognition accuracy.

[0166] In one embodiment, the comprehensive image authenticity assessment data includes a first scale-specific error vector corresponding to the first pressure decomposition parameter and a second scale-specific error vector corresponding to the second pressure decomposition parameter. Figure 6 After step 105, the Chinese character generation method may further include: updating the first pressure decomposition parameter and the second pressure decomposition parameter, specifically including:

[0167] Step 601: multiplying a first scale-specific error vector by a preset error projection matrix to obtain a first initial pressure adjustment value;

[0168] Step 602: limiting the amplitude of the first initial pressure adjustment amount to obtain a first target pressure adjustment amount;

[0169] Step 603: Add the first target pressure adjustment amount to the first pressure decomposition parameter to obtain an updated first pressure decomposition parameter;

[0170] Step 604: multiplying the second scale-specific error vector by a preset error projection matrix to obtain a second initial pressure adjustment value;

[0171] Step 605: limiting the second initial pressure adjustment amount to obtain a second target pressure adjustment amount;

[0172] Step 606 : Add the second target pressure adjustment amount to the second pressure decomposition parameter to obtain an updated second pressure decomposition parameter.

[0173] Specifically, a dynamic adjustment mechanism is set up, according to the k-th specific error vector e fed back by the discriminator k (High-dimensional error vector) Real-time update of the k-th pressure decomposition parameter:

[0174]

[0175] Where η is the learning rate (default value is 0.01); W e is the error projection matrix, which maps the high-dimensional error signal to the pressure adjustment amount; tanh is the hyperbolic tangent function, which limits the amplitude of the pressure adjustment.

[0176] The benefit of the embodiment of the above steps 601 to 606 is that, through the closed-loop feedback design, the model can continuously optimize the pressure distribution during the training process, perform adaptive processing on Chinese characters of different complexities, and improve the quality of Chinese character generation.

[0177] In step 106 of some embodiments, a pressure gradient signal is generated based on the comprehensive image authenticity assessment data, and the parameters of the initial cross-scale coupler are updated based on the pressure gradient signal to obtain a target cross-scale coupler. Specifically, a loss calculation is performed based on the comprehensive image authenticity assessment data and preset real data to obtain a total loss function. Then, a gradient calculation is performed on the parameters of the initial cross-scale coupler based on the total loss function to obtain a pressure gradient signal.

[0178] In order to adapt to the unique structural characteristics of Chinese characters, this embodiment introduces stroke perception constraints in the entropy calculation stage:

[0179]

[0180] in, is the gradient magnitude map of IMF (indicating the rate of change of features); (i, j) represents the two-dimensional spatial coordinates of the feature map, i.e., the pixel position of the i-th row and j-th column on the image; M stroke is the pre-computed stroke direction template (extracted by the Sobel operator); ‖·‖1 represents the L1 norm, which measures the absolute difference between the predicted gradient and the template; ∑ i,j (·) represents the summation operation over all spatial locations (i, j) in the feature map. This structure-aware constraint ensures that the generated features maintain the connectivity and topological relationships of Chinese character strokes, effectively addressing the stroke breakage and deformation issues common in traditional methods. It is used to constrain gradient updates.

[0181] In one example, this embodiment combines three loss functions to achieve a comprehensive optimization goal: adversarial loss L adv (to ensure the authenticity of generated Chinese characters), multi-scale reconstruction loss L rec (to ensure structural consistency at different resolutions) and pressure level consistency loss L pressure (Ensure that the pressure decomposition parameters are reasonably distributed).

[0182] The total loss function is:

[0183] L=λ1·L adv +λ2·L rec +λ3·L pressure ,

[0184] Where λ1, λ2, and λ3 are weight coefficients, determined by tuning the validation set. This multi-objective optimization strategy enables the model to balance global structure and local details, accurately preserving the structural characteristics of Chinese characters while ensuring overall authenticity.

[0185] This application also implements a variety of training enhancement techniques, including progressive growing, feature regularization, and batch adaptive weighting. Progressive growing first pre-trains the model at a low resolution (16×16), and then gradually increases to the target resolution (64×64 or 128×128), allowing the network to learn the overall structure first and then optimize local details. Feature regularization enhances the robustness of the model by adding small random perturbations (Gaussian noise, which follows a normal distribution with a mean of 0 and a standard deviation of 0.1) to the pressure parameters, simulating pressure fluctuations in real hydraulic systems. Batch adaptive weighting dynamically adjusts the loss weight according to the complexity distribution of each batch of samples (the complexity score calculated by the number of strokes and the topological structure) to ensure that both simple and complex characters are fully optimized.

[0186] It should be noted that the image generator and image discriminator can also be trained according to the total loss function.

[0187] In step 107 of some embodiments, a preset target noise vector is first feature-converted to obtain a target noise feature map; the target noise feature map is hydraulically decomposed according to at least two preset pressure decomposition parameters to obtain a target noise sub-hydraulic feature corresponding to each pressure decomposition parameter; cross-scale feature coupling is performed on at least two target noise sub-hydraulic features through a target cross-scale coupler to obtain a target multi-scale hydraulic feature; and image generation is performed on the target multi-scale hydraulic feature to obtain a target Chinese character.

[0188] Based on the above embodiments, the present application can at least achieve the following beneficial effects: 1. The pressure transmission principle of the hydraulic system is introduced into the GAN architecture, and the pressure levels correspond to different feature scales, which realizes the innovative integration of physical mechanism and deep learning, and significantly improves the generation quality of complex Chinese characters; 2. An IMF decomposition strategy is customized for Chinese character generation, and stroke connectivity constraints are introduced in entropy calculation, which effectively maintains the unique topological structure of Chinese characters and solves the common structural errors in traditional methods; a pressure-sensitive dynamic weight adjustment mechanism is designed, so that the network can automatically balance the feature contributions of different scales according to the complexity of the characters, realizing adaptive processing, which is particularly suitable for the high-precision generation of complex Chinese characters.

[0189] See also Figure 7 The embodiment of the present application also provides a Chinese character generation device, which can implement the above-mentioned Chinese character generation method. Figure 7The module structure block diagram of the Chinese character generation device provided in the embodiment of the present application includes: a noise acquisition module 701, which is used to obtain a sample random noise vector and perform feature conversion on the sample random noise vector to obtain a sample noise feature map; a feature decomposition module 702, which is used to perform hydraulic feature decomposition on the sample noise feature map according to at least two preset pressure decomposition parameters to obtain a sample noise sub-hydraulic feature of each pressure decomposition parameter; wherein any two pressure decomposition parameters are different; a cross-scale coupling module 703, which is used to perform cross-scale feature coupling on at least two sample noise sub-hydraulic features through a preset initial cross-scale coupler to obtain a sample multi-scale hydraulic feature; an image generation module 704, which is used to generate a sample noise sub-hydraulic feature. The image generation module 705 is used to perform multi-scale image discrimination on the sample random Chinese character image to obtain comprehensive evaluation data of image authenticity; the parameter updating module 706 is used to generate a pressure gradient signal according to the comprehensive evaluation data of image authenticity, and update the parameters of the initial cross-scale coupler according to the pressure gradient signal to obtain a target cross-scale coupler; the Chinese character generation module 707 is used to perform cross-scale feature coupling on at least two target noise sub-features corresponding to a preset target noise vector through the target cross-scale coupler to obtain target noise multi-scale features, and perform image generation on the target noise multi-scale features to obtain target Chinese characters.

[0190] In one embodiment, the Chinese character generation device further includes a pressure distribution updating module, configured to update the first pressure decomposition parameter and the second pressure decomposition parameter.

[0191] It should be noted that the specific implementation of the Chinese character generation device is basically the same as the specific embodiment of the above-mentioned Chinese character generation method, and will not be repeated here.

[0192] The present application also provides an electronic device comprising: a memory, a processor, a program stored in the memory and executable on the processor, and a data bus for enabling communication between the processor and the memory. When the program is executed by the processor, the aforementioned Chinese character generation method is implemented. The electronic device can be any intelligent terminal, including a tablet computer and an in-vehicle computer.

[0193] See also Figure 8 , Figure 8 The hardware structure of an electronic device according to another embodiment is shown. The electronic device includes:

[0194] The processor 801 can be implemented as a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of the present application.

[0195] The memory 802 can be implemented in the form of a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 802 can store an operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 802 and is called by the processor 801 to execute the Chinese character generation method of the embodiments of this application.

[0196] Input / output interface 803, used to implement information input and output;

[0197] Communication interface 804, used to implement communication interaction between this device and other devices, which can be achieved through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WiFi, Bluetooth, etc.);

[0198] Bus 805 , which transmits information between various components of the device (e.g., processor 801 , memory 802 , input / output interface 803 , and communication interface 804 );

[0199] The processor 801 , the memory 802 , the input / output interface 803 and the communication interface 804 are connected to each other in communication within the device via a bus 805 .

[0200] An embodiment of the present application also provides a storage medium, which is a computer-readable storage medium used for computer-readable storage. The storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the above-mentioned Chinese character generation method.

[0201] The memory, as a non-transient computer-readable storage medium, can be used to store non-transient software programs and non-transient computer executable programs. In addition, the memory may include a high-speed random access memory and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some embodiments, the memory may optionally include a memory remotely arranged relative to the processor, and these remote memories may be connected to the processor via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0202] The embodiments described in the embodiments of this application are intended to more clearly illustrate the technical solutions of the embodiments of this application and do not constitute a limitation on the technical solutions provided by the embodiments of this application. Those skilled in the art will appreciate that with the evolution of technology and the emergence of new application scenarios, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

[0203] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of the present application, and may include more or fewer steps than shown in the figures, or a combination of certain steps, or different steps.

[0204] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, i.e., they may be located in one place or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of this embodiment.

[0205] Those skilled in the art will appreciate that all or some of the steps in the methods, systems, and functional modules / units in the devices disclosed above may be implemented as software, firmware, hardware, or appropriate combinations thereof.

[0206] The terms "first", "second", "third", "fourth", etc. (if any) in the specification of the present application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequential order. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0207] It should be understood that in this application, "at least one (item)" means one or more, and "plurality" means two or more. "And / or" is used to describe the association relationship of associated objects, indicating that three relationships may exist. For example, "A and / or B" can mean: only A exists, only B exists, and A and B exist at the same time, where A and B can be singular or plural. The character " / " generally indicates that the previous and next associated objects are in an "or" relationship. "At least one of the following items" or similar expressions refers to any combination of these items, including any combination of single items or plural items. For example, at least one of a, b or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, c can be single or multiple.

[0208] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0209] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0210] In addition, the functional units in the various embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0211] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes multiple instructions for enabling an electronic device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store programs, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0212] The preferred embodiments of the present invention are described above with reference to the accompanying drawings, but are not intended to limit the scope of the present invention. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and essence of the present invention should be within the scope of the present invention.

Claims

1. A Chinese character generation method, characterized in that: The method comprises: Obtaining a sample random noise vector, and performing feature conversion on the sample random noise vector to obtain a sample noise feature map; Performing hydraulic feature decomposition on the sample noise feature map according to at least two preset pressure decomposition parameters to obtain a sub-hydraulic feature of the sample noise corresponding to each pressure decomposition parameter; wherein any two of the pressure decomposition parameters are different; Performing cross-scale feature coupling on at least two sample noise sub-hydraulic features through a preset initial cross-scale coupler to obtain a sample multi-scale hydraulic feature; Performing image generation on the multi-scale hydraulic features of the sample to obtain a sample random Chinese character image; Performing multi-scale image discrimination on the sample random Chinese character images to obtain comprehensive evaluation data of image authenticity; generating a pressure gradient signal according to the comprehensive image authenticity evaluation data, and updating the parameters of the initial cross-scale coupler according to the pressure gradient signal to obtain a target cross-scale coupler; The target cross-scale coupler performs cross-scale feature coupling on at least two target noise sub-features corresponding to a preset target noise vector to obtain a target noise multi-scale feature, and performs image generation on the target noise multi-scale feature to obtain a target Chinese character.

2. The method according to claim 1, characterized in that The at least two pressure decomposition parameters include a second pressure decomposition parameter of the first pressure decomposition parameter, the first pressure decomposition parameter being smaller than the second pressure decomposition parameter; The step of performing hydraulic feature decomposition on the sample noise feature map according to at least two preset pressure decomposition parameters to obtain a sample noise sub-hydraulic feature of each pressure decomposition parameter includes: Performing a 1×1 convolution on the sample noise feature map to obtain a first sample convolution sub-feature; performing nonlinear mapping on the first pressure decomposition parameter to obtain a first sample hydraulic mapping feature; multiplying the first sample convolution sub-feature and the first sample hydraulic mapping feature element by element to obtain the sample noise sub-hydraulic feature corresponding to the first pressure decomposition parameter; Performing a 1×1 convolution on the sample noise sub-hydraulic feature corresponding to the first pressure decomposition parameter to obtain a second sample convolution sub-feature; performing nonlinear mapping on the second pressure decomposition parameter to obtain a second sample hydraulic mapping feature; The second sample convolution sub-feature and the second sample hydraulic mapping feature are element-wise multiplied to obtain the sample noise sub-hydraulic feature corresponding to the second pressure decomposition parameter.

3. The method according to claim 2, characterized in that The performing nonlinear mapping on the first pressure decomposition parameter to obtain a first sample hydraulic mapping feature includes: Multiplying a preset hydraulic decomposition weight matrix by the first pressure decomposition parameter to obtain a first sample hydraulic mapping vector; Adding the first sample hydraulic mapping vector to a preset bias vector to obtain a first sample hydraulic bias vector; Feature activation is performed on the first sample hydraulic pressure offset vector to obtain the first sample hydraulic pressure mapping feature.

4. The method according to claim 2, characterized in that The image authenticity comprehensive assessment data includes a first scale-specific error vector corresponding to the first pressure decomposition parameter and a second scale-specific error vector corresponding to the second pressure decomposition parameter; After performing multi-scale image discrimination on the sample random Chinese character image to obtain comprehensive image authenticity evaluation data, the method further includes: Updating the first pressure decomposition parameter and the second pressure decomposition parameter specifically includes: multiplying the first scale-specific error vector by a preset error projection matrix to obtain a first initial pressure adjustment amount; Limiting the amplitude of the first initial pressure adjustment amount to obtain a first target pressure adjustment amount; obtaining an updated first pressure decomposition parameter by adding the first target pressure adjustment amount to the first pressure decomposition parameter; multiplying the second scale-specific error vector by a preset error projection matrix to obtain a second initial pressure adjustment amount; limiting the amplitude of the second initial pressure adjustment amount to obtain a second target pressure adjustment amount; The second target pressure adjustment amount is added to the second pressure decomposition parameter to obtain an updated second pressure decomposition parameter.

5. The method according to claim 2, characterized in that The cross-scale feature coupling of at least two sample noise sub-hydraulic features by a preset initial cross-scale coupler to obtain a sample multi-scale hydraulic feature includes: performing average pooling on the sample noise sub-hydraulic features corresponding to the first pressure decomposition parameter to obtain a first sample pooled sub-hydraulic feature; Calculating the difference between the second pressure decomposition parameter and the first pressure decomposition parameter to obtain a hierarchical pressure gradient, and converting the hierarchical pressure gradient through a preset pressure difference conversion network to obtain a first sample pressure difference conversion feature; multiplying the first sample pooling sub-hydraulic feature by the first sample pressure difference conversion feature element-by-element to obtain a hydraulic information transmission feature; performing feature fusion based on the hydraulic information transmission feature and the sample noise sub-hydraulic feature corresponding to the second pressure decomposition parameter to obtain an updated sample noise sub-hydraulic feature corresponding to the second pressure decomposition parameter; Feature fusion is performed on the sample noise sub-hydraulic feature corresponding to the first pressure decomposition parameter and the updated sample noise sub-hydraulic feature corresponding to the second pressure decomposition parameter to obtain a sample multi-scale hydraulic feature.

6. The method according to any one of claims 1 to 5, characterized in that The multi-scale image discrimination is performed on the sample random Chinese character image to obtain comprehensive image authenticity evaluation data, including: Performing multi-scale downsampling on the sample random Chinese character image to obtain at least two Chinese character images to be evaluated; wherein the at least two Chinese character images to be evaluated include a first Chinese character image to be evaluated at a first scale and a second Chinese character image to be evaluated at a second scale, wherein the first scale is smaller than the second scale; Calculating the image proportions of the first Chinese character image to be evaluated and the second Chinese character image to be evaluated to obtain a first attention weight coefficient of the first Chinese character image to be evaluated and a second attention weight coefficient of the second Chinese character image to be evaluated; Obtaining a first sub-discriminator corresponding to the first scale, and performing image discrimination on the first Chinese character image to be evaluated using the first sub-discriminator to obtain first image authenticity discrimination data; performing multiplication processing on the first attention weight coefficient and the first image authenticity discrimination data to obtain first image authenticity sub-assessment data; Obtaining a second sub-discriminator corresponding to the second scale, and performing image discrimination on the second Chinese character image to be evaluated using the second sub-discriminator to obtain second image authenticity discrimination data; performing multiplication processing on the second attention weight coefficient and the second image authenticity discrimination data to obtain second image authenticity sub-assessment data; Data fusion is performed based on the first image authenticity sub-assessment data and the second image authenticity sub-assessment data to obtain the image authenticity comprehensive assessment data.

7. The method according to claim 6, characterized in that The image proportion calculation is performed based on the first Chinese character image to be evaluated and the second Chinese character image to be evaluated, to obtain a first attention weight coefficient of the first Chinese character image to be evaluated and a second attention weight coefficient of the second Chinese character image to be evaluated, respectively, including: Performing global average pooling on the first Chinese character image to be evaluated to obtain a first image pooling vector; Performing a vector inner product calculation on the preset Chinese character direction vector and the first image pooling vector to obtain a first initial attention score; Performing exponential mapping on the first initial attention score to obtain a first target attention score; Performing global average pooling on the second Chinese character image to be evaluated to obtain a second image pooling vector; Performing a vector inner product calculation on the preset Chinese character direction vector and the second image pooling vector to obtain a second initial attention score; Performing exponential mapping on the second initial attention score to obtain a second target attention score; Adding the first target attention score and the second target attention score to obtain a total attention score; Calculate the ratio of the first target attention score to the total attention score to obtain the first attention weight coefficient; The second target attention score is ratioed to the total attention score to obtain the second attention weight coefficient.

8. A Chinese character generation device, characterized in that: The device comprises: A noise acquisition module is used to acquire a sample random noise vector and perform feature conversion on the sample random noise vector to obtain a sample noise feature map; a feature decomposition module, configured to perform hydraulic feature decomposition on the sample noise feature map according to at least two preset pressure decomposition parameters, to obtain a sub-hydraulic feature of the sample noise for each of the pressure decomposition parameters; wherein any two of the pressure decomposition parameters are different; a cross-scale coupling module, configured to perform cross-scale feature coupling on at least two sample noise sub-hydraulic features through a preset initial cross-scale coupler to obtain a sample multi-scale hydraulic feature; An image generation module is used to generate an image of the multi-scale hydraulic features of the sample to obtain a sample random Chinese character image; An image discrimination module is used to perform multi-scale image discrimination on the sample random Chinese character images to obtain comprehensive evaluation data of image authenticity; a parameter updating module, configured to generate a pressure gradient signal according to the comprehensive image authenticity evaluation data, and update the parameters of the initial cross-scale coupler according to the pressure gradient signal to obtain a target cross-scale coupler; The Chinese character generation module is used to perform cross-scale feature coupling on at least two target noise sub-features corresponding to a preset target noise vector through the target cross-scale coupler to obtain a target noise multi-scale feature, and perform image generation on the target noise multi-scale feature to obtain a target Chinese character.

9. An electronic device, characterized in that: The electronic device includes a memory and a processor, the memory stores a computer program, and the processor implements the Chinese character generation method according to any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the Chinese character generation method according to any one of claims 1 to 7 is implemented.

Images