Image generation method and device

Abstract

The embodiment of the invention provides an image generation method and device, and the method comprises the steps: inputting an ancient character rubbing image and a character description text of an ancient character into a pre-constructed character generator, and generating an initial ancient character image through the character generator; generating a background image of the ancient character carrier based on physical simulation; performing fusion processing on the initial ancient character image and the background image to obtain a fused ancient character image; converting the fused ancient character image into a three-dimensional ancient character image; inputting the three-dimensional ancient character image, the ancient character rubbing image and the character description text into a pre-constructed discriminator, and outputting an evaluation result of the three-dimensional ancient character image by the discriminator; and regenerating the three-dimensional ancient character image based on the evaluation result. According to the method, the high-quality three-dimensional ancient character image can be generated, an effective data sample is provided for model training of ancient character recognition, and the recognition accuracy is improved.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to an image generation method and apparatus. Background Technology

[0002] Ancient Chinese character recognition, especially for precious and complex ancient scripts like oracle bone script, has long faced multiple severe challenges. First, the original carriers of ancient scripts (e.g., tortoise shells and animal bones used for oracle bone inscriptions) have survived thousands of years, resulting in a limited number of intact specimens, most of which are damaged to varying degrees. This makes clear, complete physical samples for research extremely scarce, failing to meet the massive training data requirements of large-scale deep learning models. Second, the morphology (including stroke thickness, structure, and direction) of the same character can vary significantly across different periods, regions, and even engravers. Furthermore, the physical characteristics of the carriers themselves (e.g., tortoise shells and animal bones), such as uneven surfaces and existing cracks, also affect the character's morphology, increasing the difficulty of recognition. Finally, ancient scripts are carved on three-dimensional carriers; the depth and three-dimensionality of the engravings are crucial for accurate character recognition and understanding. Traditional two-dimensional rubbings or photographs cannot fully preserve this three-dimensional information, leading to the loss of many valuable features during recognition, affecting accuracy and robustness.

[0003] To solve the problem of insufficient data, data enhancement and synthetic data generation techniques are widely used in image recognition. Traditional image generation methods, such as using generative adversarial networks (GANs) to generate ancient character images, have many limitations, mainly including: first, the scarcity of physical samples makes it difficult for traditional GANs to obtain sufficient training, which can easily lead to mode collapse or low-quality images that lack diversity. Second, the structure of ancient characters, the correctness of strokes, and the physical properties of the carrier material all have important semantic information. Traditional GANs only focus on visual features (such as texture, color, and shape) when generating images, and lack the ability to understand the semantic content of images. The discriminator may not be able to determine whether the generated image meets the semantic characteristics of ancient characters (such as stroke structure and morphological rules), resulting in semantic inconsistencies in the generated data. The generated characters "look like" but have structural errors or do not conform to physical laws. For example, the generated oracle bone script images may have errors in stroke connection and direction, or the generation of cracks does not conform to the principles of mechanics, and the texture of the material is also far from the real oracle bone. In addition, traditional GANs are difficult to directly generate three-dimensional images, and cannot compensate for the lack of three-dimensional information. Even if some post-processing methods are used to try to add three-dimensional effects, they lack physical basis and are difficult to truly restore the natural curvature and crack depth of the original carrier. These limitations make traditional image generation methods ineffective in the application of ancient character data enhancement and auxiliary recognition. SUMMARY

[0004] Therefore, the embodiments of the present application aim to provide an image generation method and device.

[0005] To achieve the above purpose, the embodiments of the present application provide an image generation method, comprising:

[0006] inputting an ancient character rubbing image and a character description text of an ancient character into a pre-constructed character generator, and generating an initial ancient character image by the character generator;

[0007] generating a background image of an ancient character carrier based on physical simulation;

[0008] fusing the initial ancient character image and the background image to obtain a fused ancient character image;

[0009] converting the fused ancient character image into a three-dimensional ancient character image;

[0010] inputting the three-dimensional ancient character image, the ancient character rubbing image, and the character description text into a pre-constructed discriminator, and outputting an evaluation result of the three-dimensional ancient character image by the discriminator;

[0011] re-generate a three-dimensional ancient character image based on the evaluation result.

[0012] Optionally, the generating the background image of the ancient character carrier based on the physical simulation comprises:

[0013] constructing a carrier physical model according to a pre-constructed geometric feature dataset and a material texture feature dataset;

[0014] simulating based on the carrier physical model to obtain a geometric feature and a crack feature of the carrier;

[0015] extracting a texture feature from the obtained high-resolution ancient character image by using a Gabor filter;

[0016] simulating a crack propagation path on a carrier surface of the carrier physical model by using a Poisson process to obtain a crack propagation feature;

[0017] generating a background image according to the geometric feature, the crack feature, the texture feature and the crack propagation feature of the carrier.

[0018] Optionally, the simulating the carrier physical model to obtain a geometric feature of the carrier comprises:

[0019] establishing a thin shell wave equation describing a deformation behavior of the carrier physical model under an external force;

[0020] taking a preset curvature variation parameter as an input, and solving the thin shell wave equation by using a numerical method to obtain a displacement field and a new geometry of the carrier physical model in a three-dimensional space.

[0021] Optionally, the extracting a texture feature from the obtained high-resolution ancient character image by using a Gabor filter comprises:

[0022] obtaining a high-resolution ancient character image by using microscopic photography;

[0023] extracting a feature response of a carrier material from the high-resolution ancient character image by using a constructed Gabor filter set;

[0024] analyzing a statistical property of the material according to the feature response;

[0025] generating a texture image corresponding to the carrier material by using a parameterized texture synthesis method according to the statistical property.

[0026] Optionally, the simulating a crack propagation path on a carrier surface by using a Poisson process to obtain a crack propagation feature comprises:

[0027] simulating the Poisson process according to a material parameter of the carrier to generate a crack point;

[0028] According to the stress field obtained by solving the thin shell wave equation and the fracture toughness of the material, the propagation path of the carrier surface is determined.

[0029] Optionally, the Poisson process is simulated according to the material parameters of the carrier to generate the crack point, comprising:

[0030] A plurality of candidate points are randomly generated on the surface of the carrier, and the intensity value of each candidate point is calculated according to a preset intensity function;

[0031] According to the intensity value of the candidate point and the preset intensity threshold, the acceptance probability is calculated;

[0032] If the acceptance probability is greater than or equal to a random number generated in the interval [0, 1], the candidate point corresponding to the acceptance probability is taken as the crack point.

[0033] Optionally, the fused ancient character image is converted into a three-dimensional ancient character image, comprising:

[0034] A depth image of the fused ancient character image is generated using a depth estimation method;

[0035] A left eye view and a right eye view of the depth image are generated using a disparity mapping method;

[0036] The left eye view and the right eye view are input into a pre-constructed image occlusion completion model, and a left eye view and a right eye view pair after occlusion completion are output by the image occlusion completion model.

[0037] Optionally, the discriminator is implemented based on a multi-modal large model; the three-dimensional ancient character image, the ancient character rubbing image and the character description text of the ancient character are input into a pre-constructed discriminator, and an evaluation result of the three-dimensional ancient character image is output by the discriminator, comprising:

[0038] The discriminator processes the three-dimensional ancient character image, the ancient character rubbing image and the character description text to obtain a three-dimensional ancient character image feature vector, an ancient character rubbing image feature vector and a character description text feature vector;

[0039] According to the similarity between the three-dimensional ancient character image feature vector and the ancient character rubbing image feature vector, the image authenticity of the three-dimensional ancient character image compared with the ancient character rubbing image is judged;

[0040] According to the similarity between the three-dimensional ancient character image feature vector and the character description text feature vector, the semantic matching degree of the three-dimensional ancient character image and the character description text is judged;

[0041] According to the three-dimensional ancient character image feature vector, the physical rationality of the background of the carrier is judged.

[0042] Optionally, based on the evaluation result, a three-dimensional ancient character image is re-generated, comprising:

[0043] The character generator, the physical simulation, and the image occlusion completion model are taken as a generator to perform adversarial learning with the discriminator, the generator re-generates a three-dimensional ancient character image based on an evaluation result of the discriminator for the three-dimensional ancient character image until a set end condition is met, and a final three-dimensional ancient character image is obtained.

[0044] Embodiments of the present application also provide an image generation apparatus, comprising:

[0045] An initial image generation module is configured to input an ancient character rubbing image and a character description text of an ancient character into a pre-constructed character generator, and generate an initial ancient character image by the character generator;

[0046] A background image generation module is configured to generate a background image of an ancient character carrier based on a physical simulation;

[0047] A fusion module is configured to perform fusion processing on the initial ancient character image and the background image to obtain a fused ancient character image;

[0048] A conversion module is configured to convert the fused ancient character image into a three-dimensional ancient character image;

[0049] A discrimination module is configured to input the three-dimensional ancient character image, the ancient character rubbing image, and the character description text into a pre-constructed discriminator, and output an evaluation result of the three-dimensional ancient character image by the discriminator;

[0050] An adversarial learning module is configured to re-generate a three-dimensional ancient character image based on the evaluation result.

[0051] As can be seen from the above, the image generation method and apparatus provided by the embodiments of the present application input an ancient character rubbing image and a character description text of an ancient character into a pre-constructed character generator, generate an initial ancient character image by the character generator, generate a background image of an ancient character carrier based on a physical simulation, perform fusion processing on the initial ancient character image and the background image to obtain a fused ancient character image, convert the fused ancient character image into a three-dimensional ancient character image, input the three-dimensional ancient character image, the ancient character rubbing image, and the character description text into a pre-constructed discriminator, output an evaluation result of the three-dimensional ancient character image by the discriminator, and re-generate a three-dimensional ancient character image based on the evaluation result. The present application can generate a high-quality and diversified three-dimensional ancient character image, provide effective data samples for model training of ancient character recognition, and improve recognition accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0052] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the accompanying drawings needed to be used in the description of the embodiments or the prior art will be briefly introduced. Obviously, the accompanying drawings in the following description only need to explain the present application, and other drawings can be obtained by those of ordinary skill in the art without any creative work.

[0053] Figure 1 The method flowchart of the embodiments of the present application is shown in the figure.

[0054] Figure 2 The device structure block diagram of the embodiments of the present application is shown in the figure.

[0055] Figure 3 The electronic device structure block diagram of the embodiments of the present application is shown in the figure. DETAILED DESCRIPTION

[0056] In order to make the objects, technical solutions and advantages of the present disclosure clearer, the present disclosure will be further described in detail below with reference to specific embodiments and drawings.

[0057] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the embodiments of the present application should be understood as the general meaning understood by those skilled in the art to which the present disclosure belongs. The terms "first", "second" and the like used in the embodiments of the present application do not represent any order, number or importance, but are only used to distinguish different components. The terms "include" or "contain" and the like mean that the elements or objects before the terms cover the elements or objects listed after the terms and their equivalents, and do not exclude other elements or objects. The terms "connect" or "connect" and the like are not limited to physical or mechanical connection, but can include electrical connection, whether direct or indirect. The terms "up", "down", "left", "right" and the like only represent relative positional relationship, which may change accordingly when the absolute position of the described object changes.

[0058] As described in the background, due to the problems of lack of real sample, large difference in character form, lack of three-dimensional information and the like, ancient character recognition has always been a difficulty in the field of artificial intelligence. Limited by the difficulty of collecting real samples and the cost of labeling, the ancient character recognition model based on deep learning has low recognition accuracy when facing characters of different forms and different angles. In order to realize data enhancement, ancient character images are generated based on limited real samples, which still faces the problems of limited training samples, difficult to understand semantic information, poor three-dimensional information restoration effect and the like.

[0059] In view of this, embodiments of this application provide an image generation method that can generate a large number of high-quality and diverse ancient character image samples based on scarce physical samples. It utilizes physical simulation technology and a multimodal semantic visual encoder to generate background images with correct material physics and reasonable crack mechanics, thereby improving the physical authenticity and semantic accuracy of the ancient character images. By employing monocular depth estimation and stereo vision mapping technology, it generates three-dimensional ancient character images, which can provide richer visual information and improve the recognition accuracy of ancient character characters.

[0060] The technical solution of this application will be further described in detail below through specific embodiments.

[0061] like Figure 1 As shown, this application provides an image generation method, including:

[0062] S101: Input the ancient script rubbing image and the ancient script character description text into a pre-built character generator, and the character generator generates an initial ancient script image;

[0063] In this embodiment, images of ancient characters on physical objects, rubbings of ancient characters, character descriptions of ancient characters, and related literature are collected to construct an ancient character image dataset, a geometric morphology feature dataset of the original carriers of the ancient characters, and a material texture feature dataset. The geometric morphology feature dataset includes carrier contour coordinate data extracted from images of ancient characters on physical objects and rubbings, as well as key feature data such as crack direction, bifurcation, and distribution patterns. The material texture feature dataset includes high-resolution texture images of the carrier surface (e.g., acquired through microscopic imaging), key feature data such as color distribution and material roughness extracted from the high-resolution texture images, and local special texture features. Based on the constructed ancient character image dataset, a preset generator model is trained to obtain a trained character generator. This character generator can generate an initial ancient character image containing ancient characters based on the input ancient character rubbing image and the character description text. Optionally, the generator model can be a conditional GAN ​​model or a conditional variational autoencoder (VAE) model.

[0064] In some methods, ancient scripts include, but are not limited to, oracle bone script, bronze script, and cuneiform script, which are characters inscribed on carriers such as tortoise shells, animal bones, bronzes, and clay tablets. Character descriptions include the composition of ancient characters (e.g., "from [character name], from [phonetic character]"), meaning (e.g., "representing sacrificial activities"), and common morphological features (e.g., "wide at the top, with a vertical stroke at the bottom").

[0065] S102: Background image of ancient script carrier generated based on physical simulation;

[0066] In the embodiment, according to the constructed geometric feature data set and material texture feature data set, a background image of the ancient character carrier is generated by simulating natural bending shape, material texture, surface crack feature and environmental light and shadow of the original carrier based on physical simulation and multi-scale geometric deformation algorithm.

[0067] In some ways, the background image of the ancient character carrier is generated based on physical simulation, comprising:

[0068] A carrier physical model is constructed according to the pre-constructed geometric feature data set and material texture feature data set;

[0069] The carrier physical model is simulated to obtain the shape feature and crack feature of the carrier;

[0070] The Gabor filter is used to extract the texture feature from the obtained high-resolution ancient character image;

[0071] The Poisson process is used to simulate the crack propagation path on the carrier surface of the carrier physical model to obtain the crack propagation feature;

[0072] The background image is generated according to the shape feature, crack feature, texture feature and crack propagation feature of the carrier.

[0073] In the embodiment, in order to generate a background image with physically correct material and reasonable crack mechanics, a carrier physical model is constructed based on the constructed geometric feature data set and material texture feature data set, the carrier physical model is simulated, the natural bending shape and cracks of the ancient character carrier caused by stress are simulated, the Gabor filter is used to extract the microscopic texture of the specific material of the carrier, the Poisson process is used to simulate the crack propagation path on the carrier surface, and the shape and distribution of the crack are ensured to comply with the mechanical law. Finally, the shape feature, texture feature and crack propagation feature of the carrier are fused to form a realistic background image of the ancient character carrier.

[0074] In some ways, the initial bending degree, specific shape and surface crack of the ancient character carrier can be randomly simulated and generated. In this way, for the same ancient character, the position, direction and density generated by random generation multiple times are different, a large number of background image samples with different shapes and rich details can be generated, the diversity of data samples is improved, the ancient character recognition model is trained based on the diversified data samples, and the recognition accuracy can be improved.

[0075] In some embodiments, the thin shell theory is used to describe and analyze the deformation behavior of a shell structure with a thickness much smaller than the size of the other two dimensions under stress. Considering that the geometric features of the original carrier of the ancient character have certain similarity with the thin shell model, the carrier physical model is constructed based on the "thin shell wave equation" in the application, and the generated carrier shape has physical interpretability.

[0076] In some ways, the carrier physical model is simulated to obtain the morphological characteristics of the carrier, including:

[0077] A thin shell wave equation is established to describe the deformation behavior of the carrier physical model under the action of external force;

[0078] The preset curvature change parameter is taken as input, and the thin shell wave equation is solved by numerical method to obtain the displacement field of the carrier physical model in three-dimensional space and the new form.

[0079] In this embodiment, based on the geometric morphological feature data set and the material texture feature data set, the carrier model is modeled, the two-dimensional surface of the carrier model is meshed to obtain a meshed carrier model, i.e. the carrier physical model, and each mesh element is regarded as a small element of the shell; According to the data recorded in the data set, the material properties and geometric parameters of the meshed carrier model are defined, wherein the material properties include elastic modulus, Poisson's ratio, etc., and the geometric parameters include initial curvature, thickness, etc. A partial differential equation set is established to describe the deformation behavior of the carrier physical model under the action of external force, i.e. the thin shell wave equation, which is used to simulate the deformation of the carrier under the action of stress release in the archaeological excavation process, shrinkage stress in the natural weathering process, or local stress generated during writing. The thin shell wave equation considers the bending, stretching and shear deformation of the shell, and adopts the dimensionless von Kármán type thin shell wave equation set (von Karman equation), which has the following form:

[0080]

[0081] Wherein, ρ, h are the density and thickness of the shell (i.e. the carrier), The bending stiffness of the carrier is represented by E, v, which are the elastic modulus and Poisson's ratio, κ(w) is the Gaussian curvature tensor induced by the normal displacement w(x, y, t), σ(w) is the membrane stress tensor generated by the displacement w, p(x, y, t) is the external load for simulating the natural shrinkage or stress crack of the carrier, u(x, y, t) is the face displacement in the x direction (the movement of the shell material point along the shell itself or the neutral surface), v(x, y, t) is the face displacement in the y direction, x, y are spatial coordinates, and t is time.

[0082] In order to drive the deformation of the carrier, a curvature-driven partial differential equation (PDE) dynamic deformation strategy is adopted, that is, The curvature energy functional is defined as:

[0083]

[0084] The Fréchet derivative is obtained, and the curvature-driven evolution equation is obtained:

[0085]

[0086] where a is a positive diffusion coefficient and k0 is a pre-defined or randomly generated initial curvature field. This equation drives the dynamic deformation of the carrier surface displacement w(x, y, t) in real time through the curvature difference Ak- k0), so that the final shape is physically interpretable and consistent with the curvature distribution of real carriers such as tortoise shells / bone.

[0087] According to formulas (1)-(3), by pre-defining or randomly generating an initial curvature field k0, or dynamically adjusting the curvature of each part of the carrier according to the simulated stress situation, the curvature variation parameter is obtained. The curvature variation parameter is taken as input, and the thin shell wave equation is solved by numerical method (such as finite element method or finite difference method) to simulate the displacement field, stress field and new shape of the carrier physical model in three-dimensional space. In this way, the natural bending, local depression or bulge, and even irregular overall distortion of the carrier caused by internal stress or external load can be truly simulated. For example, the arch shape of the tortoise carapace, the relatively flat but irregular edge feature of the plastron, and the bending of the bone due to drying shrinkage or uneven stress can be simulated.

[0088] In some embodiments, the carrier physical model can also be combined with fracture mechanics to simulate the generation and propagation of cracks. When the stress at a certain place of the shell exceeds the fracture strength of the material, a crack can be introduced, and the energy-based method or stress intensity factor criterion can be used to predict the propagation direction and path of the crack. This physically-based crack generation method is more realistic than simple random line drawing or texture superposition, and can better reflect the mechanical correlation between the crack and the overall deformation of the carrier. By introducing the thin shell wave equation and the curvature-driven PDE dynamic deformation, the limitations of traditional two-dimensional affine transformation in shape diversity expression can be overcome, and carrier shapes with rich geometric details and physical rationality can be generated, which not only improves the visual realism of the generated image, but also provides an accurate geometric basis for subsequent crack generation, material texture extraction, and three-dimensional reconstruction. This physically interpretable modeling method makes the generated ancient character images not just a pile of pixels, but also reflects their inherent properties as physical entities.

[0089] In some embodiments, after generating carrier shape features with physical interpretability, in order to further improve the realism of the synthesized image, a "crack-material joint prior" strategy is adopted to make the synthesized image visually satisfy the dual requirements of "material physical correctness" and "crack mechanical reasonableness". This strategy mainly includes two aspects: using multi-scale Gabor filtering to extract and generate material texture features of the carrier, and using Poisson process to simulate the propagation path of the crack.

[0090] where the Gabor filter is used to extract texture features from high-resolution ancient character images, including:

[0091] high-resolution ancient text image is obtained through microphotography;

[0092] a feature response of the carrier material is extracted from the high-resolution ancient text image using the constructed Gabor filter bank;

[0093] statistical properties of the material are analyzed according to the feature response;

[0094] a texture image corresponding to the carrier material is generated through a parametric texture synthesis method according to the statistical properties.

[0095] In this embodiment, the surface of the carrier has unique micro-textures, which are important visual features of the material of the carrier. The Gabor filter is a linear filter, and its impulse response is a harmonic defined by the product of a Gaussian function and a sine function, which has good frequency selection and direction selection characteristics, and can effectively extract the texture features in the image. In order to capture the texture features of the specific material of the carrier at different scales, a multi-scale Gabor filter bank is used. Specifically, by adjusting the scale parameter (i.e. the size of the Gaussian window) and the direction parameter of the Gabor filter, a series of filters that can respond to different spatial frequencies and different directional textures are constructed. A high-resolution ancient text image is obtained through microphotography, and the feature response of the carrier material is extracted from the high-resolution ancient text image using the constructed filter bank. The statistical properties of the material, such as the roughness, directionality, periodicity, etc. of the texture, are analyzed according to the feature response. In the generation stage, a texture image similar in appearance to the real carrier material is generated through a parametric texture synthesis method (e.g. a method based on Markov random field or a deep learning method) according to the statistical properties of the material. This Gabor filter-based texture feature extraction and material generation method can ensure that the generated material is close to the real carrier material in microstructure, thereby meeting the requirement of "material physics correctness".

[0096] In some ways, for a multi-scale Gabor filter bank, the Gabor filter kernel of the kth scale and the lth direction is:

[0097]

[0098] where x', y' are rotated coordinates, the center frequency f k = f0 / 2 k-1 , the variance b is a constant, and φ is a phase shift.

[0099] In some ways, the propagation path of the crack on the surface of the carrier is simulated using a Poisson process to obtain crack propagation features, including:

[0100] According to the material quality parameter of the carrier, a Poisson process is simulated to generate a crack point;

[0101] According to the stress field obtained by solving the thin shell wave equation and the fracture toughness of the material, an extension path on the surface of the carrier is determined.

[0102] The method for generating the crack point comprises:

[0103] A plurality of candidate points are randomly generated on the surface of the carrier, and intensity values of each candidate point are calculated according to a preset intensity function;

[0104] According to the intensity value of the candidate point and a preset intensity threshold, an acceptance probability is calculated;

[0105] If the acceptance probability is greater than or equal to a random number generated in the interval [0, 1], the candidate point corresponding to the acceptance probability is taken as the crack point.

[0106] In this embodiment, it is considered that the crack on the carrier is the result of the joint action of multiple factors such as long-term burial, environmental change (shrinkage and expansion caused by temperature and humidity change) or external stress, and the morphology and distribution of the crack have certain randomness but also follow certain mechanical laws. Poisson process is used to describe the random occurrence of events in continuous time or space. The present application regards the initiation and propagation of the crack as a sequence of random events, and by setting a suitable intensity function, the frequency and density of the crack appearing on the surface of the carrier can be controlled. Specifically, the crack initiation obeys a non-homogeneous Poisson process, and the intensity function is:

[0107]

[0108] Wherein, λ0 is the basic intensity of the material, d(x, y) is the distance between the current point (x, u) and the existing crack, d0 is the inhibition distance, β is the control attenuation, σ max is the local maximum principal stress, σ c is the fracture threshold, and each parameter is known for a specific point on the carrier.

[0109] ​A plurality of candidate points are randomly generated on the entire carrier surface, the intensity value of each candidate point is calculated according to the intensity function shown in formula (5), the maximum intensity value λ max is selected as the intensity threshold value of the reference from the intensity values of all candidate points on the entire carrier surface, and the acceptance probability of any candidate point (x, y) is equal to the actual intensity value λ (x, y) of the point divided by the intensity threshold value λ max, that is, p = λ / λ max; the acceptance probability of the candidate point is compared with a random number generated in the interval [0, 1], if the acceptance probability is greater than or equal to the random number, then the candidate point is accepted as a real crack point, otherwise it is rejected. That is, the greater the intensity value, the more concentrated the stress or the more fragile the material, and the higher the probability that the corresponding candidate point is accepted and finally becomes a real crack point. In this way, a crack network that is physically consistent in distribution, natural in shape and reasonable in location can be simulated and generated.

[0110] In some ways, the probability of the number of crack points N(A) in the region A is:

[0111]

[0112] Where k is a non-negative integer, representing the specific number of crack initiation points appearing in region A, and formula (6) describes the probability of exactly k crack points appearing in region A.

[0113] In some ways, the propagation path of the crack can be guided in combination with the stress distribution of the carrier (for example, the stress field calculated by the thin shell wave equation) and the fracture toughness of the material. For example, cracks tend to initiate in stress concentration or material weakness areas and propagate along the direction of maximum principal stress or maximum energy release rate. By simulating the Poisson process, a series of crack points (i.e. crack initiation points) and propagation directions can be generated, thereby constructing a crack network that is natural in shape and reasonable in distribution, and the crack propagation path follows the direction of maximum principal stress:

[0114] dl = v(σ)dt + εdW t (7)

[0115]

[0116] Where the parameters v0, m, ε are preset hyperparameters, v0 determines the growth rate of the crack, m determines the acceleration size of the crack under the action of stress, and ε determines the bending degree and randomness of the crack path; dl represents the small length increment of the crack tip in a small time step dt, σ is the local maximum principal stress at the current position of the crack tip, which is the main physical quantity driving the crack propagation, and W tis a standard Wiener process (or Brownian motion), dWt represents an infinitesimal increment of the standard Wiener process, which can be understood as a completely random "walk" or "perturbation" that occurs in an infinitesimal time step dt.

[0117] The crack generation method based on the Poisson process is more dynamic and realistic than a simple random line or a fixed pattern crack template, and can better reflect the mechanical mechanism of crack formation and meet the requirement of "reasonable crack mechanics".

[0118] In some ways, the morphological features, crack features, texture features and crack propagation features of the extracted carrier are fused to generate a background image containing carrier detail information. When fusing, the interaction between the crack and the material needs to be considered. For example, the depth of the crack will affect the apparent color and luster of the material, and the material at the edge of the crack may have a slight curl or peeling. The geometric details of the crack can be realistically represented on the surface of the material through techniques such as displacement mapping or normal mapping. At the same time, the color and texture of the crack should be different from the surrounding material to enhance the visual contrast. The "Gabor-Poisson" joint sampling strategy is adopted in this embodiment, and the fusion mode of the material and the crack can generate a carrier surface that is highly realistic in vision, conforms to the physical properties of the material and follows the mechanical laws of the crack, laying a foundation for subsequent character superposition and stereoscopic processing.

[0119] S103: Fuse the initial ancient character image and the background image to obtain a fused ancient character image;

[0120] In this embodiment, the initial ancient character image containing ancient character characters and the background image with carrier detail information obtained through physical simulation are fused to obtain a two-dimensional fused ancient character image containing both ancient character characters and carrier details.

[0121] S104: Convert the fused ancient character image into a three-dimensional ancient character image;

[0122] In this embodiment, the two-dimensional fused ancient character image is converted into a three-dimensional ancient character image, and the method comprises:

[0123] A depth estimation method is used to generate a depth image of the fused ancient character image;

[0124] A disparity mapping method is used to generate a left eye view and a right eye view of the depth image;

[0125] The left eye view and the right eye view are input into a pre-constructed image occlusion completion model, and the left and right eye view pairs after occlusion completion are output by the image occlusion completion model.

[0126] In this embodiment, a depth estimation method based on deep learning is used to generate a depth image of the fused ancient script image. For example, the fused ancient script image is input into the MiDaS (Monocular Depth Estimation in the Wild) monocular depth estimation model. The model extracts three-dimensional geometric information such as the undulation of the carrier surface and the depth of character inscriptions from the two-dimensional fused ancient script image, predicts the depth value of each pixel, and outputs a depth image. This depth image is a grayscale image, and the brightness value of the pixel represents the depth information of the pixel.

[0127] Based on depth images, pixel-level disparity mapping technology is used to generate corresponding left and right eye views, thus initially constructing stereo vision. Disparity refers to the difference in horizontal pixel coordinates of the same spatial point in the left and right eye views, and it is inversely proportional to the depth of that point. Based on the principles of camera imaging geometry, a mathematical model between depth and disparity can be established. For each pixel in the input image, based on its corresponding depth value in the depth map, the pixel distance (i.e., disparity) that the pixel should move horizontally when generating the right or left eye view can be calculated. By horizontally translating each pixel in the input image according to the calculated disparity value and filling it onto a new canvas, the right or left eye view can be initially generated. It is important to note that the direction of translation is different for the left and right eyes; for example, for the right eye view, distant points are translated to the left, and nearby points are translated to the right. In this way, a pair of left and right eye views with disparity can be generated from a single input image, thus initially constructing stereo vision.

[0128] During the process of generating left and right eye views through parallax mapping, due to the change in viewing angle, the generated left and right eye views may contain some occluded areas. That is, some areas visible in the original image from one viewpoint may be obscured by foreground objects from another viewpoint, thus forming "holes" or "occluded areas" in the new view. For example, when viewing a raised oracle bone carrier from the left, part of its right edge is visible, but in the generated right eye view (equivalent to viewing from further to the right), this part of the area may be obscured by the carrier itself. The occluded area will seriously affect the viewing effect of the stereoscopic image and the accuracy of the generated image.

[0129] To address this issue, this embodiment employs an image completion technique based on generative adversarial networks (GANs). Left and right eye views with occluded regions are input into a trained image occlusion completion model. This model is built upon a GAN, for example, using a Pix2Pix or similar structure, and is trained with an incomplete image as input and a complete image as the target. The generator of the image occlusion completion model predicts and fills in the content of the occluded region, ensuring consistency with the texture, structure, and semantic information of the surrounding area. The discriminator judges whether the completed image is realistic and natural. Through adversarial learning, the image occlusion completion model learns how to reasonably infer and generate the content of the occluded portion based on contextual information, outputting a pair of completed left and right eye views, significantly improving the realism and completeness of the stereoscopic effect.

[0130] This embodiment constructs a "closed-loop" processing flow from two-dimensional images to three-dimensional stereo images through three steps: monocular depth estimation, pixel-level parallax mapping, and occlusion completion. It can effectively recover three-dimensional information from two-dimensional fused ancient text images and generate high-quality three-dimensional ancient text images without the need for expensive binocular cameras or complex three-dimensional scanning equipment, significantly reducing data acquisition costs and providing a new approach for the three-dimensional digitization of ancient texts.

[0131] S105: Input the three-dimensional ancient character image, the ancient character rubbing image, and the character description text of the ancient character into the pre-built discriminator, and the discriminator outputs the evaluation result of the three-dimensional ancient character image;

[0132] In this embodiment, after generating a three-dimensional ancient character image, the three-dimensional ancient character image, the ancient character rubbing image, and the character description text of the ancient character are input into a pre-constructed discriminator. The discriminator outputs the evaluation result of the three-dimensional ancient character image. Through adversarial training, the realism and semantic consistency of the three-dimensional ancient character image are optimized to generate a highly realistic three-dimensional ancient character image.

[0133] In some implementations, the discriminator selects a pre-trained multimodal large model with strong image understanding capabilities (e.g., Contrastive Language-Image Pretraining, CLIP model). The multimodal large model can not only extract the visual features of the image, but also combine the semantic information of the image and text modalities to understand the semantic content of the image (e.g., determine whether the generated image conforms to the stroke rules and morphological features of ancient characters), which significantly improves the authenticity and semantic consistency of the generated data, and is especially suitable for tasks that require high semantic consistency, such as ancient character recognition.

[0134] In some embodiments, the three-dimensional ancient script image, the ancient script rubbing image, and the character description text of the ancient script are input into a pre-constructed discriminator, which outputs the evaluation result of the three-dimensional ancient script image, including:

[0135] The discriminator processes the three-dimensional ancient character image, the ancient character rubbing image, and the character description text to obtain the feature vectors of the three-dimensional ancient character image, the ancient character rubbing image, and the character description text.

[0136] Based on the similarity between the feature vectors of the three-dimensional ancient characters image and the feature vectors of the ancient characters rubbing image, the image authenticity of the three-dimensional ancient characters image is judged compared with that of the ancient characters rubbing image.

[0137] Based on the similarity between the feature vectors of the three-dimensional ancient characters image and the feature vectors of the character description text, the semantic matching degree between the three-dimensional ancient characters image and the character description text is determined;

[0138] Based on the feature vectors of the three-dimensional ancient script image, the physical plausibility of the background of the carrier is determined.

[0139] In this embodiment, a discriminator based on a multimodal large model simultaneously processes the input 3D ancient character image, ancient character rubbing image, and ancient character character description text, obtaining their feature vectors in the shared embedding space. Based on the similarity between the feature vectors, the semantic consistency between them is determined, thereby evaluating the quality of the generated image from multiple dimensions. Specifically, this includes: judging the image authenticity of the generated 3D ancient character image compared to the ancient character rubbing image based on the similarity between the feature vectors of the 3D ancient character image and the feature vectors of the character description text (e.g., cosine similarity); judging the semantic matching degree between the 3D ancient character image and the character description text based on the similarity between the feature vectors of the 3D ancient character image and the feature vectors of the character description text (e.g., whether the generated character structure conforms to the text description); and determining the physical plausibility of the background of the carrier based on the feature vectors of the 3D ancient character image (e.g., whether the cracks conform to mechanical laws, whether the material is realistic, which can be indirectly constrained by the understanding of the overall semantics of the image by the multimodal large model, or combined with a dedicated physical plausibility evaluation module).

[0140] In some approaches, if the generator produces a character that looks like ancient script but whose structure doesn't match the character description text, or if the physical features of the background (such as crack morphology or material texture) don't match the character description text (e.g., how the cracks should be distributed, what texture the material should have), the discriminator will give a lower semantic matching score. This mechanism forces the generator to focus not only on pixel-level visual realism during optimization but also on the high-level semantic accuracy of the image content, thus effectively solving the problem of "looking like it but semantically wrong."

[0141] S106: Based on the evaluation results, regenerate the three-dimensional ancient script image.

[0142] In this embodiment, a character generator, physical simulation, and image occlusion completion model are used as generators to engage in adversarial learning with a discriminator. The generator regenerates the 3D ancient character image based on the discriminator's evaluation of the 3D ancient character image until a set termination condition is met, resulting in the final 3D ancient character image. Specifically, based on the discriminator's evaluation of the generated 3D ancient character image, the character generator regenerates the initial ancient character image, and the physical simulation regenerates the background image. The regenerated initial ancient character image and the background image are then merged to regenerate the merged ancient character image. Further 3D transformation and occlusion completion are then performed to regenerate the 3D ancient character image. The discriminator evaluates the regenerated 3D ancient character image. Through adversarial learning between the generator and the discriminator, the quality of the 3D ancient character image is improved.

[0143] In this embodiment, a character generator, a physical simulation module, and an image occlusion completion model are collectively used as a generator, which engages in adversarial learning with a discriminator based on a multimodal large model. The generator aims to produce images that are both realistic and conform to semantic and physical laws, capable of "deceiving" the discriminator. The discriminator aims to better distinguish between real and generated ancient character images and to determine the semantic consistency between the image and the text describing the characters. Through the adversarial learning mechanism between the generator and the discriminator, and through backpropagation to continuously optimize the parameters of each module, the entire system can iteratively improve the quality, physical realism, and semantic accuracy of the generated 3D ancient character images.

[0144] In some approaches, the termination condition can be that the adversarial learning process between the generator and discriminator converges, or a preset number of iterations is reached, or a preset performance metric is achieved, at which point a final, highly realistic, and high-quality 3D ancient character image is output. Based on this, a data sample set can be constructed using the generated high-quality 3D ancient character images. This data sample set can then be used to train an ancient character recognition model, thus addressing the problem of insufficient training data for ancient character recognition. Simultaneously, high-quality data samples can effectively improve the accuracy and generalization ability of the ancient character recognition model. Furthermore, high-quality 3D ancient character images also provide researchers with richer visual information, helping to understand the inscription features and carrier morphology of oracle bone inscriptions from a 3D perspective.

[0145] The technical effects of this application are illustrated below with comparative test data.

[0146] The performance of the method in this application is compared with that of traditional generative adversarial networks (GANs) on the task of generating oracle bone inscriptions. As shown in Table 1, compared with traditional GANs, the three key indicators of this application—Fréchet Inception Distance (FID), Structural Similarity Index (SSIM), and Peak Signal-to-Noise Ratio (PSNR)—are significantly improved. This indicates that the method in this application can generate oracle bone inscription images with higher quality, more accurate structure, and richer details. It provides an effective technical means to solve the pain points of scarce physical samples, large morphological differences, and lack of three-dimensional information in the field of ancient script research.

[0147] Table 1 Quality evaluation indicators of images generated by this application and traditional generative adversarial networks.

[0148]

[0149] The image generation method provided in this application, based on a small dataset of ancient script images and data, uses a character generator to generate an initial ancient script image containing ancient characters; through innovative physical simulation and multi-scale geometric deformation algorithms, it simulates the natural morphological features of the carrier to generate a background image of the ancient script carrier; the initial ancient script image and the background image are then fused to obtain a fused ancient script image; combining depth estimation and disparity mapping algorithms, the two-dimensional fused ancient script image is converted into a three-dimensional ancient script image with stereoscopic visual effects, enabling the generator to reconstruct the shape of ancient characters on the surface of a three-dimensional carrier; a pre-trained multimodal large model with powerful image understanding capabilities is used as a discriminator, enabling the model to... This method generates ancient character synthetic data from a small number of descriptive texts and real ancient character images, ultimately producing high-quality, diverse 3D ancient character images. It overcomes the shortcomings of traditional synthetic data generation methods, which struggle to utilize 3D spatial features and semantic information. This provides effective data samples for training models for ancient character recognition, addresses the pain point of lacking high-fidelity synthetic data for ancient character recognition, and improves the accuracy of ancient character recognition. Simultaneously, it provides richer visual information for ancient character research, contributing to improved recognition accuracy and understanding of character inscription techniques. This method is scalable and applicable to small-sample learning scenarios, promoting the intelligent development of ancient character research.

[0150] It should be noted that the method in this embodiment can be executed by a single device, such as a computer or server. The method can also be applied in a distributed scenario, where multiple devices cooperate to complete the task. In such a distributed scenario, one of these devices may execute only one or more steps of the method in this embodiment, and the multiple devices will interact with each other to complete the method described.

[0151] It should be noted that the above description describes specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims may be performed in a different order than that shown in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0152] like Figure 2 As shown in the illustration, this application also provides an image generation apparatus, comprising:

[0153] The initial image generation module is used to input the ancient script rubbing image and the character description text of the ancient script into a pre-built character generator, which generates the initial ancient script image.

[0154] The background image generation module is used to generate background images of ancient script carriers based on physical simulation.

[0155] The fusion module is used to fuse the initial ancient character image and the background image to obtain a fused ancient character image;

[0156] The conversion module is used to convert fused ancient script images into three-dimensional ancient script images;

[0157] The discrimination module is used to input three-dimensional ancient character images, ancient character rubbing images, and character description text into a pre-built discriminator, which outputs the evaluation results of the three-dimensional ancient character images.

[0158] The adversarial learning module is used to regenerate 3D images of ancient texts based on the evaluation results.

[0159] For ease of description, the above devices are described in terms of function, divided into various modules. Of course, in implementing the embodiments of this application, the functions of each module can be implemented in one or more software and / or hardware.

[0160] The apparatus described above is used to implement the corresponding methods in the foregoing embodiments and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0161] Figure 3 This embodiment illustrates a more specific hardware structure of an electronic device, which may include a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, memory 1020, input / output interface 1030, and communication interface 1040 are interconnected internally via the bus 1050.

[0162] The processor 1010 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, 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 this specification.

[0163] The memory 1020 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1020 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program code is stored in the memory 1020 and is called and executed by the processor 1010.

[0164] The input / output interface 1030 is used to connect input / output modules to realize information input and output. Input / output modules can be configured as components within the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touchscreens, microphones, various sensors, etc., while output devices may include displays, speakers, vibrators, indicator lights, etc.

[0165] The communication interface 1040 is used to connect a communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0166] Bus 1050 includes a pathway for transmitting information between various components of the device, such as processor 1010, memory 1020, input / output interface 1030, and communication interface 1040.

[0167] It should be noted that although the above-described device only shows the processor 1010, memory 1020, input / output interface 1030, communication interface 1040, and bus 1050, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the embodiments of this specification, and not necessarily all the components shown in the figures.

[0168] The electronic devices described above are used to implement the corresponding methods in the foregoing embodiments and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0169] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0170] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of this disclosure (including the claims) is limited to these examples; within the framework of this disclosure, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the embodiments of this application as described above, which are not provided in the details for the sake of brevity.

[0171] Additionally, to simplify the description and discussion, and to avoid obscuring the embodiments of this application, the well-known power / ground connections to integrated circuit (IC) chips and other components may or may not be shown in the provided drawings. Furthermore, the apparatus may be shown in block diagram form to avoid obscuring the embodiments of this application, and this also takes into account the fact that the details of the implementation of these block diagram apparatuses are highly dependent on the platform on which the embodiments of this application will be implemented (i.e., these details should be fully understood by those skilled in the art). While specific details (e.g., circuits) have been set forth to describe exemplary embodiments of this disclosure, it will be apparent to those skilled in the art that the embodiments of this application can be implemented without these specific details or with variations thereof. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0172] Although this disclosure has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed.

[0173] The embodiments of this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of this application should be included within the protection scope of this disclosure.

Claims

1. An image generation method characterized by, The method comprises the following steps: inputting an ancient character rubbing image and a character description text of the ancient character into a pre-constructed character generator to generate an initial ancient character image by the character generator; generating a background image of an ancient character carrier based on physical simulation; fusing the initial ancient character image and the background image to obtain a fused ancient character image; converting the fused ancient character image into a three-dimensional ancient character image; inputting the three-dimensional ancient character image, the ancient character rubbing image and the character description text into a pre-constructed discriminator to output an evaluation result of the three-dimensional ancient character image by the discriminator; re-generating a three-dimensional ancient character image based on the evaluation result.

2. The method of claim 1, wherein, The method for generating the background image of the ancient character carrier based on physical simulation comprises the following steps: constructing a carrier physical model according to a pre-constructed geometric feature data set and a material texture feature data set; simulating based on the carrier physical model to obtain morphological features and crack features of the carrier; extracting texture features from a high-resolution ancient character image obtained by using a Gabor filter; simulating crack propagation paths on the carrier surface of the carrier physical model by using a Poisson process to obtain crack propagation features; generating a background image according to the morphological features, the crack features, the texture features and the crack propagation features of the carrier.

3. The method of claim 2, wherein, The method for simulating the carrier physical model to obtain morphological features of the carrier comprises the following steps: establishing a thin shell wave equation describing the deformation behavior of the carrier physical model under external force; inputting a pre-set curvature change parameter as an input, and solving the thin shell wave equation by a numerical method to obtain a displacement field and a new form of the carrier physical model in a three-dimensional space.

4. The method of claim 2, wherein, The method for extracting texture features from a high-resolution ancient character image obtained by using a Gabor filter comprises the following steps: obtaining a high-resolution ancient character image by microscopic photography; extracting feature responses of the carrier material from the high-resolution ancient character image by using a constructed Gabor filter set; analyzing the statistical properties of the material according to the feature responses; generating a texture image corresponding to the carrier material by using a parameterized texture synthesis method according to the statistical properties.

5. The method of claim 3, wherein, The method for simulating crack propagation paths on the carrier surface by using a Poisson process to obtain crack propagation features comprises the following steps: simulating a Poisson process according to the material parameters of the carrier to generate crack points; determining the propagation path on the carrier surface according to the stress field obtained by solving the thin shell wave equation and the fracture toughness of the material.

6. The method of claim 5, wherein, The method for simulating a Poisson process according to the material parameters of the carrier to generate crack points comprises the following steps: randomly generating a plurality of candidate points on the carrier surface, and calculating the intensity value of each candidate point according to a pre-set intensity function; calculating an acceptance probability according to the intensity value of the candidate point and a pre-set intensity threshold; if the acceptance probability is greater than or equal to a random number generated in the interval [0, 1], the candidate point corresponding to the acceptance probability is taken as a crack point.

7. The method of claim 1, wherein, The method for converting the fused ancient character image into a three-dimensional ancient character image comprises the following steps: generating a depth image of the fused ancient character image by using a depth estimation method; generating a left eye view and a right eye view of the depth image by using a disparity mapping method; input the left-eye view and the right-eye view into a pre-constructed image occlusion completion model, and output the left-eye view and the right-eye view after occlusion completion by the image occlusion completion model.

8. The method of claim 1, wherein, The discriminator is implemented based on a multi-modal large model; the three-dimensional ancient character image, the ancient character rubbing image and the character description text of the ancient character are input into a pre-constructed discriminator, and an evaluation result of the three-dimensional ancient character image is output by the discriminator, including: The discriminator processes the three-dimensional ancient character image, the ancient character rubbing image and the character description text to obtain a three-dimensional ancient character image feature vector, an ancient character rubbing image feature vector and a character description text feature vector; According to the similarity between the three-dimensional ancient character image feature vector and the ancient character rubbing image feature vector, the image authenticity of the three-dimensional ancient character image compared with the ancient character rubbing image is judged; According to the similarity between the three-dimensional ancient character image feature vector and the character description text feature vector, the semantic matching degree of the three-dimensional ancient character image and the character description text is judged; According to the three-dimensional ancient character image feature vector, the physical rationality of the background of the carrier is judged.

9. The method of claim 7, wherein, Based on the evaluation result, a three-dimensional ancient character image is regenerated, including: The character generator, the physical simulation and the image occlusion completion model are used as a generator to perform adversarial learning with the discriminator; the generator regenerates a three-dimensional ancient character image based on the evaluation result of the three-dimensional ancient character image by the discriminator until a set end condition is met, and finally obtains a three-dimensional ancient character image.

10. An image generation apparatus characterized by comprising: including: An initial image generation module is configured to input an ancient character rubbing image and a character description text of an ancient character into a pre-constructed character generator, and generate an initial ancient character image by the character generator; A background image generation module is configured to generate a background image of an ancient character carrier based on physical simulation; A fusion module is configured to fuse the initial ancient character image and the background image to obtain a fused ancient character image; A conversion module is configured to convert the fused ancient character image into a three-dimensional ancient character image; A discrimination module is configured to input the three-dimensional ancient character image, the ancient character rubbing image and the character description text into a pre-constructed discriminator, and output an evaluation result of the three-dimensional ancient character image by the discriminator; An adversarial learning module is configured to regenerate a three-dimensional ancient character image based on the evaluation result.

Images