Digital culture symbol automatic fusion method based on generative adversarial network

Abstract

The invention discloses a digital culture symbol automatic fusion method based on a generative adversarial network, and relates to the field of digital art, computer vision and deep learning, and the method comprises the following steps: S1, carrying out the preprocessing of a digital culture symbol image; s2, extracting feature vectors by using a convolutional neural network; s3, generating a symbol fusion image by using a generator; s4, evaluating the authenticity of the symbol fusion image through a discriminator; s5, continuously iteratively generating the weight of the adversarial network by using an improved RAdam algorithm; s6, optimizing the visual quality of the symbol fusion image; s7, calculating information entropy to score image quality; and S8, providing a user interaction visualization function. According to the method, the self-attention generative adversarial network and the deep learning technology are combined, the digital culture symbols are accurately and efficiently generated and fused, and the method is particularly suitable for automatic processing of complex backgrounds and diversified symbols and has a wide application prospect.

Description

Technical Field

[0001] This invention relates to the fields of generative adversarial networks, digital cultural symbol processing, and image processing, and particularly to an automatic fusion method for digital cultural symbols based on generative adversarial networks. Background Technology

[0002] With the advent of the digital age, the creation and dissemination of cultural symbols face unprecedented opportunities and challenges. In the field of modern design, digital cultural symbols are not only important carriers of cultural expression but also key elements of cross-cultural communication. However, current symbol design largely relies on manual creation, a process that is cumbersome and inefficient. Traditional digital cultural symbol design methods typically depend on the designer's experience and artistic perception, employing manual or template-based methods to complete symbol creation and integration. While these methods may meet needs in certain situations, their lack of flexibility and creative diversity often fails to meet the demands of rapidly evolving digital cultural dissemination.

[0003] The limitation of traditional symbol design lies in its inefficiency in the generation and fusion of symbol elements. Existing methods often rely on fixed templates and manually defined rules for symbol combination. This creative approach is not only time-consuming but also struggles to generate diverse symbols that meet specific needs. When dealing with complex symbol images, symbol fusion, and innovation, traditional methods have very limited expressiveness and adaptability. Especially for application scenarios involving large amounts of historical symbol data and symbol variations, manual design and traditional methods lack efficient processing tools.

[0004] Furthermore, traditional symbol design methods often neglect the diversity and flexibility of symbols during the generation of symbol images, making it difficult to adapt to complex visual elements and graphic design requirements. For example, in the process of symbol fusion and feature extraction, traditional methods cannot effectively capture the complex details in the image, resulting in generated symbol images that may lack visual appeal or accuracy in cultural connotation. Even when using traditional machine learning techniques for symbol design, they often fail to fully utilize the potential information in the image data, causing the model to be unable to capture the multi-dimensional relationships of symbol features, thus affecting the quality and application effectiveness of the generated symbol images.

[0005] Therefore, how to achieve the automatic generation and fusion of symbols based on modern deep learning technology has become an urgent problem to be solved in the field of digital cultural symbol creation. Summary of the Invention

[0006] This invention proposes an automatic fusion method for digital cultural symbols based on generative adversarial networks (GANs). By introducing an improved RAdam algorithm, it can generate and fuse symbol images more accurately. This method not only automatically extracts important features from symbol images but also uses deep learning models to diversify and optimize symbol image creation, significantly improving the efficiency and quality of symbol design. By introducing the improved RAdam algorithm, the model can adjust its generation strategy in real time to meet dynamically changing needs, generating symbol images that conform to specific cultural backgrounds and application requirements. This overcomes many limitations of traditional symbol design methods in symbol generation and fusion, providing an efficient, accurate, and innovative solution for the creation of digital cultural symbols.

[0007] An automatic fusion method for digital cultural symbols based on generative adversarial networks according to an embodiment of the present invention includes the following steps: S1. Input multiple digital cultural symbol image data, preprocess the input images, including denoising, normalization and contrast enhancement, and generate preprocessed image data; S2. Based on the preprocessed image data, use a convolutional neural network to extract features and output the feature vector of each digital cultural symbol. S3. Based on symbolic feature vectors, generate symbolic fusion images using a generator from a generative adversarial network; S4. Based on the symbol feature vector, the discriminator evaluates the authenticity of the symbol fusion image and adjusts the generator weights accordingly. S5. Use the improved RAdam algorithm to continuously iterate and generate adversarial network weights to update the symbolic fusion image; S6. Perform boundary smoothing, detail enhancement, and color adjustment on the symbol fusion image; S7. Evaluate the symbol fusion image by scoring the image quality based on structural similarity and information entropy, and generate the symbol fusion image. S8. Display the symbol fusion image through a visual interface, allowing users to edit and adjust the symbol fusion image, and record the symbol fusion image after real-time adjustment.

[0008] Optionally, S1 specifically includes: S11. Receive multiple digital cultural symbol image data as input, perform size and format standardization processing on the input images, and output image data with uniform size and format; S12. Denoise the input image data by using Gaussian filtering or median filtering algorithms, and output the denoised image. S13. Normalize the denoised image data, adjust the pixel values ​​to a uniform range, and output the normalized image. S14. Enhance the contrast of the normalized image and output the preprocessed image data.

[0009] Optionally, S2 specifically includes: S21. Input the preprocessed image data, and use the first convolutional layer of the convolutional neural network to perform convolution operation on the input image to extract low-level features; S22. Pass the low-level features to the next convolutional layer for multi-layer feature extraction. Each convolutional operation extracts feature information at different scales to obtain a feature map. S23. Perform pooling operation on the feature maps extracted from each layer, and convert the pooled feature maps into one-dimensional feature vectors through a fully connected layer to generate the feature representation of each digital cultural symbol and output the symbol feature vector.

[0010] Optionally, S3 specifically includes: S31. Input the symbolic feature vector as the initial condition of the generator network, pass the feature vector to the generator model, and output the input feature data of the generator network. S32. Input the symbolic feature vector into the fully connected layer of the generator. The fully connected layer maps the symbolic feature vector to a high-dimensional feature space for feature expansion and outputs the expanded high-dimensional feature data. S33. Input the mapped feature data into multiple convolutional layers, process the feature data layer by layer through the convolutional layers, extract image features at different levels, and output the feature map after the convolution operation. S34. Input the feature map after the convolution operation into the deconvolution layer, restore the feature map to the size of the original image through the deconvolution operation, generate the symbol fusion image, and output the restored symbol image. S35. Input the generated symbol image into the activation function layer of the generator, perform nonlinear mapping processing on the image, and output the symbol fusion image.

[0011] Optionally, S4 specifically includes: S41. The input symbol fusion image is subjected to multiple convolution operations through the convolutional layer in the discriminator to finally obtain the high-level features of the symbol fusion image; S42. The discriminator compares the symbolic fusion image with the real image based on the high-level features of the symbolic fusion image, and calculates the authenticity of the symbolic fusion image. S43. Adjust the generator weights based on the realism of the symbol-fused image.

[0012] Optionally, the improved RAdam algorithm specifically includes: Initialize the learning rate, first moment estimate, second moment estimate, first and second moment decay rates, bias correction and sparse data correction factor; Calculate the sparsity of the input data, analyze the proportion of non-zero elements in the gradient, and calculate the sparsity value. Based on the sparsity value, the first and second moment estimates are corrected, and the learning rate is adjusted by calculating the sparse data correction factor. Based on the corrected first-order moment and second-order moment estimates, the gradient is calculated using the backpropagation algorithm, and the gradient of the loss function is propagated back to each layer of the network using the chain rule to calculate the gradient of each parameter. Add noise perturbation to the gradient and update the gradient; Based on the sparse data correction factor, the first-order moment estimate is updated by weighted averaging the current gradient with the first-order moment estimate from the previous step; the square of the current gradient is calculated and weighted averaging with the second-order moment estimate from the previous round to update the second-order moment estimate. By introducing bias correction, the first and second moment estimates are corrected. The update step size of each parameter is calculated using the corrected first and second moment estimates. The parameters of the generative adversarial network are updated based on the corrected learning rate and the dynamically adjusted step size. Based on moment estimation, multi-step learning rate adjustment is performed. By gradually increasing or decreasing the learning rate, the learning process at each step is dynamically adjusted to obtain the adjusted learning rate. Combined with gradient information, the weights of the adversarial generative network are updated. Based on the multi-step learning rate strategy, the learning rate of each training stage is further adjusted and iterated repeatedly until the weights of the generative adversarial network converge, and the generated symbolic fusion image is updated.

[0013] Optionally, S6 includes: S61. Based on the generated symbol image, perform boundary smoothing processing. By applying a Gaussian filter, remove noise and cluttered boundaries in the image to generate a smooth boundary image. S62. Based on the smooth boundary image, use Laplacian pyramid enhancement to optimize image details, highlight important features, and generate an enhanced image; S63. Based on the enhanced image, apply Lab space transformation to adjust the hue, saturation, and brightness of the image, and update the symbolic fusion image.

[0014] Optionally, S7 includes: S71. Based on the symbolic fusion image and the real image, calculate the similarity between the symbolic fusion image and the real image according to the structural similarity index, and output the similarity score; S72. Based on the fused image, the details and information content of the image are evaluated by calculating the information entropy, and the entropy value of the image is output. S73. Based on the scoring results of structural similarity and information entropy, comprehensively evaluate the quality of the generated image and update the symbolic fusion image.

[0015] Optionally, S8 includes: S81 provides a visual user interface to display symbol fusion images for users to view in real time; S82. Supports users to edit symbol fusion images through a visual user interface, adjusting the hue, saturation, brightness, and contrast of the symbol fusion images; S83. Record the edited and adjusted symbolic fusion image and feed it back to the generative adversarial network (GAN) for further iterative updates to the GAN weights.

[0016] The beneficial effects of this invention are: (1) This invention, by combining generative adversarial networks (GANs) and an improved RAdam algorithm, successfully overcomes the shortcomings of traditional digital cultural symbol fusion methods in handling complex cultural symbols. Traditional methods typically rely on rule-based symbol synthesis techniques, which often perform poorly when dealing with cultural symbols in diverse and complex contexts. By introducing GANs, this invention can effectively generate more realistic and culturally distinctive digital cultural symbols through training the generator and discriminator to play against each other, especially in scenarios with complex backgrounds and high requirements for symbol detail, significantly improving the quality and accuracy of symbol fusion. The improved RAdam algorithm optimizes the gradient update problem in traditional optimization algorithms, solves the problem of poor adaptability to sparse and imbalanced data during training, improves the stability and convergence speed of model training, and further enhances the performance of symbol fusion. This method provides an efficient and accurate solution for the automatic fusion of digital cultural symbols, significantly improving the fusion effect in large-scale cultural symbol synthesis and digital art creation, and promoting the technological development of cultural innovation and artistic creation.

[0017] (2) This invention achieves precise fusion and re-creation across different cultural symbol datasets by combining generative adversarial networks (GANs) and an improved RAdam algorithm. Digital cultural symbols often possess complex structures and multi-dimensional features. Through GANs, the model can capture subtle connections and changes between symbols, thereby achieving a more natural and harmonious symbol fusion. Furthermore, the improved RAdam algorithm effectively accelerates the training process, ensuring efficient training and optimization of the model on massive symbol datasets, further enhancing the model's generalization ability and enabling it to adapt to the fusion needs of diverse cultural symbols. Through this method, digital cultural symbols can be accurately and personalizedly fused and expressed in various creative scenarios. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1This is an overall flowchart of a method for automatic fusion of digital cultural symbols based on generative adversarial networks proposed in this invention; Figure 2 This is a flowchart of the improved RAdam algorithm for the automatic fusion method of digital cultural symbols in generative adversarial networks proposed in this invention. Detailed Implementation

[0019] The invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figure 1-2 An automatic fusion method for digital cultural symbols based on generative adversarial networks includes the following steps: S1. Input multiple digital cultural symbol image data, preprocess the input images, including denoising, normalization and contrast enhancement, and generate preprocessed image data; S2. Based on the preprocessed image data, use a convolutional neural network to extract features and output the feature vector of each digital cultural symbol. S3. Based on symbolic feature vectors, generate symbolic fusion images using a generator from a generative adversarial network; S4. Based on the symbol feature vector, the discriminator evaluates the authenticity of the symbol fusion image and adjusts the generator weights accordingly. S5. Use the improved RAdam algorithm to continuously iterate and generate adversarial network weights to update the symbolic fusion image; S6. Perform boundary smoothing, detail enhancement, and color adjustment on the symbol fusion image; S7. Evaluate the symbol fusion image by scoring the image quality based on structural similarity and information entropy, and generate the symbol fusion image. S8. Display the symbol fusion image through a visual interface, allowing users to edit and adjust the symbol fusion image, and record the symbol fusion image after real-time adjustment.

[0021] In this embodiment, S1 specifically includes: S11. Receive multiple digital cultural symbol image data as input, perform size and format standardization processing on the input images, and output image data with uniform size and format; S12. Denoise the input image data by using Gaussian filtering or median filtering algorithms, and output the denoised image. S13. Normalize the denoised image data, adjust the pixel values ​​to a uniform range, and output the normalized image. S14. Enhance the contrast of the normalized image and output the preprocessed image data.

[0022] This implementation achieves efficient preprocessing in image processing by performing size and format unification, denoising, normalization, and contrast enhancement on multiple digital cultural symbol image data. Multiple input images are received, and size and format unification ensures consistent image size and format, thereby improving the consistency and efficiency of subsequent processing. Gaussian filtering or median filtering algorithms are used to denoise the input images, effectively removing noise and ensuring image clarity. The denoised images are then normalized, adjusting pixel values ​​to a uniform range, resulting in better performance and stability during subsequent processing. Contrast enhancement technology is applied to the normalized images, making details more prominent and enhancing the image's visualization. This preprocessing improves image quality, providing high-quality input data for subsequent image analysis and applications.

[0023] In this embodiment, S2 specifically includes: S21. Input the preprocessed image data, and use the first convolutional layer of the convolutional neural network to perform convolution operation on the input image to extract low-level features; S22. Pass the low-level features to the next convolutional layer for multi-layer feature extraction. Each convolutional operation extracts feature information at different scales to obtain a feature map. S23. Perform pooling operation on the feature maps extracted from each layer, and convert the pooled feature maps into one-dimensional feature vectors through a fully connected layer to generate the feature representation of each digital cultural symbol and output the symbol feature vector.

[0024] This implementation uses a convolutional neural network to perform multi-level feature extraction on preprocessed image data, achieving advanced feature recognition in digital cultural symbol images. The preprocessed image data is input and convolved through the first convolutional layer of the convolutional neural network to extract low-level features, laying the foundation for subsequent feature extraction. These low-level features are then passed to the next convolutional layer for multi-level feature extraction. Each convolutional operation extracts feature information at different scales, forming multi-level feature maps that comprehensively describe the image details. Pooling is performed on the feature maps extracted from each layer to reduce the feature dimensionality. A fully connected layer then converts the pooled feature maps into one-dimensional feature vectors, generating the feature representation for each digital cultural symbol. The output symbol's feature vector provides accurate image feature data for subsequent symbol recognition and processing tasks.

[0025] In this embodiment, S3 specifically includes: S31. Input the symbolic feature vector as the initial condition of the generator network, pass the feature vector to the generator model, and output the input feature data of the generator network. S32. Input the symbolic feature vector into the fully connected layer of the generator. The fully connected layer maps the symbolic feature vector to a high-dimensional feature space for feature expansion and outputs the expanded high-dimensional feature data. S33. Input the mapped feature data into multiple convolutional layers, process the feature data layer by layer through the convolutional layers, extract image features at different levels, and output the feature map after the convolution operation. S34. Input the feature map after the convolution operation into the deconvolution layer, restore the feature map to the size of the original image through the deconvolution operation, generate the symbol fusion image, and output the restored symbol image. S35. Input the generated symbol image into the activation function layer of the generator, perform nonlinear mapping processing on the image, and output the symbol fusion image.

[0026] This implementation uses a generator model in a generative adversarial network (GAN) to process symbolic feature vectors, achieving automatic generation and fusion of digital cultural symbols. The input symbolic feature vector serves as the initial condition for the generator network, passing it to the generator model and outputting the generator network's input feature data. The symbolic feature vector is then fed into the generator's fully connected layer, which maps it to a high-dimensional feature space for feature expansion, outputting expanded high-dimensional feature data. This mapped feature data is then fed into multiple convolutional layers, which process the feature data layer by layer, extracting image features at different levels and outputting a feature map after the convolution operation. The convolutional feature map is then fed into a deconvolutional layer, restoring the feature map to the original image size and generating a fused symbolic image, which is then output as the restored symbolic image. Finally, the generated symbolic image is fed into the generator's activation function layer for nonlinear mapping, outputting the final fused symbolic image. This process effectively generates and fuses symbolic images, providing accurate and high-quality digital cultural symbolic images for subsequent applications.

[0027] In this embodiment, S4 specifically includes: S41. The input symbol fusion image is subjected to multiple convolution operations through the convolutional layer in the discriminator to finally obtain the high-level features of the symbol fusion image; S42. The discriminator compares the symbolic fusion image with the real image based on the high-level features of the symbolic fusion image, and calculates the authenticity of the symbolic fusion image. S43. Adjust the generator weights based on the realism of the symbol-fused image.

[0028] This implementation uses a discriminator to process the symbol-fused image, achieving adversarial learning between the generator and the discriminator, thereby improving the quality and realism of the symbol image. The discriminator performs multiple convolution operations on the input symbol-fused image through convolutional layers, obtaining high-level features that provide a foundation for subsequent image realism evaluation. Based on these high-level features, the discriminator compares the symbol-fused image with the real image, calculating the realism of the symbol-fused image. Through this comparison, the discriminator can determine the similarity between the generated symbol image and the real image, evaluating the image quality. Based on the realism of the symbol-fused image, the generator's weights are adjusted, optimizing its generation capabilities to produce more realistic and high-quality symbol images. This process improves the generation effect of digital cultural symbols through the interaction between the generator and the discriminator.

[0029] In this embodiment, the improved RAdam algorithm specifically includes: Initialize the learning rate, first moment estimate, second moment estimate, first and second moment decay rates, bias correction and sparse data correction factor; Calculate the sparsity of the input data, analyze the proportion of non-zero elements in the gradient, and calculate the sparsity value. Based on the sparsity value, the first and second moment estimates are corrected, and the learning rate is adjusted by calculating the sparse data correction factor. Based on the corrected first-order moment and second-order moment estimates, the gradient is calculated using the backpropagation algorithm, and the gradient of the loss function is propagated back to each layer of the network using the chain rule to calculate the gradient of each parameter. Add noise perturbation to the gradient and update the gradient; Based on the sparse data correction factor, the first-order moment estimate is updated by weighted averaging the current gradient with the first-order moment estimate from the previous step; the square of the current gradient is calculated and weighted averaging with the second-order moment estimate from the previous round to update the second-order moment estimate. By introducing bias correction, the first and second moment estimates are corrected. The update step size of each parameter is calculated using the corrected first and second moment estimates. The parameters of the generative adversarial network are updated based on the corrected learning rate and the dynamically adjusted step size. Based on moment estimation, multi-step learning rate adjustment is performed. By gradually increasing or decreasing the learning rate, the learning process at each step is dynamically adjusted to obtain the adjusted learning rate. Combined with gradient information, the weights of the adversarial generative network are updated. Based on the multi-step learning rate strategy, the learning rate of each training stage is further adjusted and iterated repeatedly until the weights of the generative adversarial network converge, and the generated symbolic fusion image is updated.

[0030] This implementation optimizes the training process of generative adversarial networks (GANs) under sparse data conditions by introducing an improved RAdam algorithm. Initialization is performed on the learning rate, first-order moment estimates, second-order moment estimates, first- and second-order moment decay rates, bias correction, and sparse data correction factor to prepare initial parameters for training. The sparsity of the input data is calculated, the proportion of non-zero elements in the gradient is analyzed, and the sparsity value is calculated. Based on the sparsity value, the first- and second-order moment estimates are corrected, and the learning rate is adjusted using the sparse data correction factor to ensure the stability of model training under sparse data conditions. Based on the corrected moment estimates, the gradient is calculated using the backpropagation algorithm, and the gradient of the loss function is propagated back to each layer of the network using the chain rule to calculate the gradient of each parameter. To improve the robustness of model training, noise perturbation is added to the gradient, and the gradient is updated. During the update process, based on the sparse data correction factor, the current gradient is weighted and averaged with the previous first-order moment estimate to update the first-order moment estimate; simultaneously, the square of the current gradient is calculated and weighted and averaged with the previous second-order moment estimate to update the second-order moment estimate. By introducing bias correction, the first and second moment estimates are further refined. These refined moment estimates are then used to calculate the update step size for each parameter. The parameters of the generative adversarial network (GAN) are updated based on the refined learning rate and the dynamically adjusted step size. A multi-step learning rate adjustment strategy is employed, gradually increasing or decreasing the learning rate to dynamically adjust the learning process at each step, resulting in an adjusted learning rate. Combined with gradient information, the parameters of each GAN are updated, yielding the updated network weights. Based on the multi-step learning rate strategy, the learning rate for each training stage is further adjusted. Through iterative optimization, the GAN weights converge, successfully updating and generating the symbolic fusion image. This process ensures that, even under complex training conditions, the GAN can efficiently and accurately generate and optimize symbolic images.

[0031] In this embodiment, S6 specifically includes: S61. Based on the generated symbol image, perform boundary smoothing processing. By applying a Gaussian filter, remove noise and cluttered boundaries in the image to generate a smooth boundary image. S62. Based on the smooth boundary image, use Laplacian pyramid enhancement to optimize image details, highlight important features, and generate an enhanced image; S63. Based on the enhanced image, apply Lab space transformation to adjust the hue, saturation, and brightness of the image, and update the symbolic fusion image.

[0032] This implementation improves the visual quality of the fused symbol image by performing boundary smoothing, detail enhancement, and color adjustment on the generated symbol image. For the generated symbol image, boundary smoothing is performed by applying a Gaussian filter to remove noise and cluttered boundaries, generating a smooth boundary image. Based on the smoothed boundary image, Laplacian pyramid enhancement technology is used to optimize image details and highlight important features, generating an enhanced image. Through multi-scale feature enhancement, image quality, detail representation, and visual effects are improved, making the core content of the image clearer and more expressive. Based on the enhanced image, Lab color space conversion technology is applied to adjust the hue, saturation, and brightness of the image, thereby updating the fused symbol image and making it more delicate and rich in visual effects. Through this process, the quality of the generated symbol image is significantly improved, providing higher-quality image output for subsequent applications.

[0033] In this embodiment, S7 specifically includes: S71. Based on the symbolic fusion image and the real image, calculate the similarity between the symbolic fusion image and the real image according to the structural similarity index, and output the similarity score; S72. Based on the fused image, the details and information content of the image are evaluated by calculating the information entropy, and the entropy value of the image is output. S73. Based on the scoring results of structural similarity and information entropy, comprehensively evaluate the quality of the generated image and update the symbolic fusion image.

[0034] This implementation method assesses the quality of the generated symbolic fusion image by calculating structural similarity and information entropy to ensure that the image's detail and information content meet requirements. Based on the symbolic fusion image and the real image, the similarity between them is calculated using a structural similarity index, and a similarity score is output, reflecting the degree of structural similarity. Based on the symbolic fusion image, information entropy is calculated to evaluate the image's detail richness and information content, and the image's entropy value is output. Calculating information entropy helps determine the quality and complexity of the generated image, ensuring that the generated symbolic image is richer and more meaningful in terms of detail and information expression. Based on the evaluation results of structural similarity score and information entropy, the overall quality of the generated image is comprehensively evaluated, and the symbolic fusion image is then updated. Through the comprehensive evaluation of structural similarity score and information entropy, the quality of the generated image can be precisely controlled, achieving a high level in multiple dimensions, thereby providing higher-quality image materials for subsequent applications. This process provides a scientific basis for image quality control and optimization, ensuring that the generated image meets the expected standards.

[0035] In this embodiment, S8 specifically includes: S81 provides a visual user interface to display symbol fusion images for users to view in real time; S82. Supports users to edit symbol fusion images through a visual user interface, adjusting the hue, saturation, brightness, and contrast of the symbol fusion images; S83. Record the edited and adjusted symbolic fusion image and feed it back to the generative adversarial network (GAN) for further iterative updates to the GAN weights.

[0036] This implementation provides a visual user interface that allows users to view and edit symbolic fusion images in real time, thereby enhancing image personalization and optimization capabilities. The visual user interface displays the symbolic fusion image, allowing users to view and analyze its content in real time. This feature not only improves the interactivity and operability between the user and the system but also provides a mechanism for real-time feedback and optimization, ensuring image quality and meeting user expectations. Users can edit the symbolic fusion image through the interface, adjusting hue, saturation, brightness, and contrast to flexibly change the visual effect to meet specific needs. The user-edited symbolic fusion image is recorded and fed back to the generative adversarial network (GAN) for iterative updates to the network weights, thus optimizing the image generation process. By iteratively updating the GAN weights based on user feedback, the image generation process better adapts to user needs, improves image quality, and continuously optimizes generation capabilities, achieving higher quality and customized image generation. This process enables interaction between the user and the GAN, ensuring that the generated image can be finely adjusted and continuously optimized according to user requirements.

[0037] Example 1: To verify the feasibility of this invention in practical applications, it was applied to the automatic cultural symbol generation and fusion system of a large-scale digital cultural symbol generation and fusion platform (hereinafter referred to as "Platform A"). In traditional digital cultural symbol processing systems, the system typically relies on methods based on simple image processing or traditional generation algorithms to generate and fuse symbols. This method is not only inefficient but also has poor accuracy when handling complex backgrounds and diverse symbols, making it difficult to adapt to the diversity and complexity of different cultural symbols. To solve these problems, Platform A decided to adopt the automatic digital cultural symbol fusion method based on generative adversarial networks proposed in this invention.

[0038] During implementation, Platform A first preprocesses the multimodal symbol image data, including image denoising, normalization, and contrast enhancement, to ensure the accuracy and integrity of the data. By introducing a generative adversarial network (GAN) and an improved RAdam algorithm, Platform A successfully achieves efficient generation and fusion of cultural symbols in complex backgrounds. The GAN continuously optimizes its generation capabilities through adversarial training in generation and discrimination, while the improved RAdam algorithm optimizes gradient updates during the generation process, improving the model's training efficiency and stability, thereby ensuring the generation of high-quality symbol images.

[0039] Platform A successfully overcomes the limitations of traditional methods in processing multimodal cultural symbols by combining generative adversarial networks (GANs) and an improved RAdam algorithm. First, Platform A uses a GAN to generate symbol images and a discriminator to continuously evaluate image quality, ensuring the generated symbol images better meet expectations. Second, the improved RAdam algorithm optimizes the model's training process, improving convergence speed and training stability, and ensuring efficient execution of the symbol fusion process.

[0040] During implementation, the technical team of Platform A observed that, compared with traditional symbol generation and fusion methods based on rules or simple deep learning, the automatic fusion method of digital cultural symbols based on generative adversarial networks and an improved RAdam algorithm in this invention significantly improves the accuracy and efficiency of symbol generation and fusion. Traditional methods often cannot handle the generation and fusion of complex backgrounds and diverse symbols, while the method of this invention can optimize the generation strategy in real time, ensuring efficient execution of symbol generation and fusion. By introducing generative adversarial networks, Platform A can accurately generate and fuse different cultural symbols, reducing the risk of misgeneration and omission, and improving the stability and robustness of the system.

[0041] To further verify the effectiveness of this method, Platform A compared the symbol images generated using the method of this invention with those generated using traditional methods. The comparison table is as follows: Table 1 Comparison of Digital Cultural Symbol Generation and Integration Methods on Platform A As shown in Table 1, with the application of the method of this invention, the symbol generation accuracy of platform A is improved from 78% to 92% compared to the traditional method, the generation efficiency is significantly improved, and the generation time is shortened from 150 minutes to 60 minutes. The system response time is also reduced by 58.3%, while the error generation rate is significantly reduced to only 6%. In addition, the method of this invention significantly improves processing power, increasing the number of symbols processed per second by 200%, and improving multi-task parallel processing capability by 23%.

[0042] Through the method of this invention, Platform A can more efficiently optimize the generation and fusion of digital cultural symbols, improve the efficiency of symbol generation and fusion, reduce resource waste, and enhance the accuracy of task execution. This method not only improves the system's automation level and reduces human intervention, but also significantly enhances the stability and robustness of the digital cultural symbol generation and fusion system, providing efficient technical support for large-scale cultural symbol generation and fusion.

[0043] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A digital cultural symbol automatic fusion method based on a generative adversarial network, characterized in that, The method comprises the following steps: S1, inputting a plurality of digital cultural symbol image data, preprocessing the input image, including denoising, normalization and contrast enhancement, and generating preprocessed image data; S2, based on the preprocessed image data, using a convolutional neural network to extract features, and outputting a feature vector of each digital cultural symbol; S3, based on the symbol feature vector, using a generator of a generative adversarial network to generate a symbol fusion image; S4, based on the symbol feature vector, evaluating the authenticity of the symbol fusion image through a discriminator, and adjusting the weight of the generator; S5, continuously iterating the weight of the generative adversarial network using an improved RAdam algorithm, and updating the symbol fusion image; S6, performing boundary smoothing, detail enhancement and color adjustment processing on the symbol fusion image; S7, evaluating the symbol fusion image, scoring the image quality through structural similarity and information entropy, and generating a symbol fusion image; S8, displaying the symbol fusion image through a visualization interface, supporting user editing and adjustment of the symbol fusion image, and recording the real-time adjusted symbol fusion image. 2.The method of claim 1, wherein, The S1 specifically comprises: S11, receiving a plurality of digital cultural symbol image data as input, and performing size and format unification processing on the input image, and outputting image data of uniform size and format; S12, denoising the input image data, using a Gaussian filter or a median filter algorithm, and outputting denoised image data; S13, normalizing the denoised image data, adjusting the pixel value to a uniform range, and outputting normalized image data; S14, enhancing the contrast of the normalized image, and outputting preprocessed image data. 3.The method of claim 1, wherein, The S2 specifically comprises: S21, inputting the preprocessed image data, using the first convolutional layer of the convolutional neural network to perform convolution operation on the input image, and extracting low-level features; S22, passing the low-level features to the next layer of convolutional layer for multi-layer feature extraction, each layer of convolutional operation extracts feature information of different scales, and obtains a feature map; S23, performing pooling operation on each layer of extracted feature map, converting the pooled feature map to a one-dimensional feature vector through a fully connected layer, generating a feature representation of each digital cultural symbol, and outputting a symbol feature vector. 4.The method of claim 1, wherein, The S3 specifically comprises: S31, inputting the symbol feature vector as the initial condition of the generator network, passing the feature vector to the generator model, and outputting the input feature data of the generator network; S32, inputting the symbol feature vector into the fully connected layer of the generator, mapping the symbol feature vector to a high-dimensional feature space through the fully connected layer for feature expansion, and outputting the expanded high-dimensional feature data; S33, inputting the mapped feature data into a plurality of convolutional layers, processing the feature data layer by layer through the convolutional layers, extracting image features of different levels, and outputting the feature map after convolution operation; S34, inputting the feature map after convolution operation into a deconvolutional layer, restoring the feature map to the size of the original image through deconvolution operation, generating a symbol fusion image, and outputting the restored symbol image; S35, inputting the generated symbol image into an activation function layer of the generator, performing nonlinear mapping processing on the image, and outputting a symbol fusion image.

5. The method of claim 1, wherein the method is based on a generative adversarial network. The S4 specifically includes: S41, multiple convolution operations are performed on the input symbol fusion image through the convolution layer in the discriminator, and finally the high-level features of the symbol fusion image are obtained; S42, the discriminator calculates the authenticity of the symbol fusion image by comparing the differences between the symbol fusion image and the real image according to the high-level features of the symbol fusion image; S43, the generator weight is adjusted according to the authenticity of the symbol fusion image.

6. The method of claim 1, wherein the method is based on a generative adversarial network. The improved RAdam algorithm specifically includes: Initialize the learning rate, first moment estimation, second moment estimation, first and second moment decay rate, bias correction and sparse data correction factor; Calculate the sparsity of the input data, analyze the proportion of non-zero elements in the gradient, and calculate the sparsity value; According to the sparsity value, correct the first moment and second moment estimation, adjust the learning rate by calculating the sparse data correction factor; Based on the corrected first moment estimation and second moment estimation, the gradient is calculated using the back propagation algorithm, and through the chain rule, the gradient of the loss function will be propagated back to each layer of the network to calculate the gradient of each parameter; Add noise disturbance to the gradient and update the gradient; According to the sparse data correction factor, the current gradient is weighted and averaged with the first moment estimation of the previous step to update the first moment estimation; the square of the current gradient is calculated and weighted and averaged with the second moment estimation of the previous round to update the second moment estimation; By introducing bias correction, the first moment and second moment estimation are corrected, and the update step of each parameter is calculated through the corrected first moment estimation and second moment estimation, and the parameters of the generative adversarial network are updated according to the corrected learning rate and the dynamically adjusted step; According to the moment estimation, the learning rate is adjusted in multiple steps, the learning process of each step is dynamically adjusted by gradually increasing or decreasing the learning rate, the adjusted learning rate is obtained, and the adversarial generative network weight is updated combined with the gradient information; According to the multi-step learning rate strategy, the learning rate of each training stage is further adjusted, and the generative adversarial network weight is iteratively optimized until it converges, and the generated symbol fusion image is updated.

7. The method of claim 1, wherein the method is based on a generative adversarial network. The S6 includes the following steps: S61, according to the generated symbol image, boundary smoothing processing is performed, noise and chaotic boundaries in the image are removed by applying a Gaussian filter, and a smooth boundary image is generated; S62, based on the smooth boundary image, Laplacian pyramid enhancement is used to optimize image details and highlight important features to generate an enhanced image; S63, based on the enhanced image, Lab space conversion is applied to adjust the hue, saturation and brightness of the image, and the symbol fusion image is updated. 8.The method of claim 1, wherein, The S7 specifically includes: S71, based on the symbol fusion image and the real image, the similarity between the symbol fusion image and the real image is calculated according to the structural similarity index, and the similarity score is output; S72, based on the fusion image, the image details and information amount are evaluated by calculating the information entropy, and the entropy value of the image is output; S73, based on the scoring results of structural similarity and information entropy, the quality of the generated image is comprehensively evaluated, and the symbol fusion image is updated. 9.The method of claim 1, wherein, The S8 specifically includes: S81, a visual user interaction interface is provided to display the symbol fusion image for real-time viewing by users; S82, support the user to edit the symbol fusion image through the visual user interaction interface, adjust the hue, saturation, brightness and contrast of the symbol fusion image; S83, record the edited and adjusted symbol fusion image, feed back to the generative adversarial network, and further update the weight of the generative adversarial network.

Images