Method for generating digital duplication of cartoon character based on AI (Artificial Intelligence)

Abstract

The invention relates to the field of artificial intelligence, in particular to an AI-based cartoon character digital duplicate generation method. Comprising the steps of collecting and preprocessing an original cartoon image to obtain a cartoon image; inputting the cartoon image into a convolutional layer of a convolutional neural network for feature extraction to obtain a preliminary cartoon feature map, and processing the preliminary cartoon feature map through a feature mapping layer to obtain a cartoon feature map; and obtaining the personalized feature map, combining the personalized feature map with the cartoon feature map to generate a character feature map, inputting the character feature map into the generative network, optimizing the parameters of the generative network by adopting a composite difference index optimization algorithm, generating a cartoon character optimized image, and repairing the cartoon character optimized image to obtain a cartoon character digital duplication image. The technical problems that the prior art depends on a universal feature extraction mode, is difficult to adapt to the requirements of various different users, and cannot generate a cartoon role digital duplication image according with user preferences according to the personalized requirements of the users are solved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and particularly to a method for generating digital avatars of comic characters based on AI. Background Art

[0002] With the rapid development of deep learning and computer vision technologies, image generation technologies based on artificial intelligence have made remarkable progress in multiple fields. In particular, the widespread application of convolutional neural networks (CNNs) and generative adversarial networks (GANs) has greatly improved the accuracy and efficiency of image generation and image processing tasks. The artificial intelligence-based image generation technology can achieve automated image recognition, style transfer, and content generation by efficiently extracting features from input images, and is widely used in industries such as art creation, game development, and virtual reality. At the same time, image restoration technology is also gradually developing. By means of deep learning algorithms, it can effectively repair defects in images, improve image quality and detail performance, and promote the progress of digital art creation and personalized image generation technologies.

[0003] However, the existing methods for generating digital avatars of comic characters have the following technical problems: The existing technologies rely on general feature extraction methods, lack pertinence, are difficult to meet the needs of various different users, cannot generate digital avatar images of comic characters that meet the user's preferences according to the user's personalized needs, and lack personalized customization capabilities; The optimization algorithms of the existing generation networks ignore the subtle differences between the target comic character image and the ideal comic character image, resulting in low accuracy of the finally generated optimized comic character images and unable to meet the requirements of high-quality digital avatars of comic characters; The existing loss functions often fail to effectively process the details of images, resulting in the overall performance of the images not meeting the expected effect. Summary of the Invention

[0004] The present invention provides a method for generating digital avatars of comic characters based on AI to solve the technical problems that the existing technologies rely on general feature extraction methods, lack pertinence, are difficult to meet the needs of various different users, cannot generate digital avatar images of comic characters that meet the user's preferences according to the user's personalized needs, and lack personalized customization capabilities; The optimization algorithms of the existing generation networks ignore the subtle differences between the target comic character image and the ideal comic character image, resulting in low accuracy of the finally generated optimized comic character images and unable to meet the requirements of high-quality digital avatars of comic characters; The existing loss functions often fail to effectively process the details of images, resulting in the overall performance of the images not meeting the expected effect.

[0005] A method for generating digital avatars of comic characters based on AI according to the present invention specifically includes the following technical solutions:

[0006] An AI-based method for generating digital avatars of comic characters, comprising the following steps:

[0007] S1: Collect and preprocess the original comic images to obtain comic images; input the comic images into the convolutional layer of a convolutional neural network for feature extraction to obtain preliminary comic feature maps, and then process the preliminary comic feature maps through a feature mapping layer to obtain comic feature maps;

[0008] S2: Obtain personalized feature maps and combine them with the comic feature maps to generate character feature maps, input the character feature maps into a generation network, optimize the generation network parameters using a composite difference index optimization algorithm, generate optimized comic character images, and repair the optimized comic character images to obtain digital avatar images of comic characters.

[0009] Preferably, the S1 specifically includes:

[0010] Input the comic images into the convolutional layer and the feature mapping layer of a convolutional neural network to perform feature extraction on the comic images. The convolutional layer performs a convolution operation on the input comic images by applying a convolution kernel to obtain preliminary comic feature maps.

[0011] Preferably, the S1 specifically includes:

[0012] Based on the preliminary comic feature maps, introduce a feature mapping formula and process the preliminary comic feature maps through weighted sum and non-linear mapping.

[0013] Preferably, the S1 specifically includes:

[0014] In the implementation process of the feature mapping formula, use the weight matrix of the feature mapping to weight the preliminary comic feature maps, adjust the importance of different features; suppress the influence of extreme values through logarithmic transformation; and process the preliminary comic feature maps using an exponential function and a non-linear activation function to obtain comic feature maps.

[0015] Preferably, the S1 specifically includes:

[0016] The feature mapping formula is as follows:

[0017]

[0018] Where G is the comic feature map; σ is the non-linear activation function; F represents the preliminary comic feature map output by the convolution operation; W is the weight matrix of the feature mapping; b′ is the bias term in the feature mapping process; γ is the adjustment parameter; log(1 + F·W + b′) represents the logarithmic transformation of the preliminary comic feature map weighted by the weight matrix; exp represents the exponential operation.

[0019] Preferably, the S2 specifically includes:

[0020] The formula for personalized modeling is as follows:

[0021]

[0022] Among them, P is the generated character feature map; ξ is the weighting coefficient; G is the comic feature map; G 2 represents the calculation of the square of the elements in the comic feature map; G' is the personalized feature map; G' 3 represents the calculation of the cube of the elements in the personalized feature map; δ is the exponential adjustment parameter.

[0023] Preferably, the S2 specifically includes:

[0024] In the implementation process of the composite difference index optimization algorithm, by minimizing the composite loss function, the generation network parameters are optimized. The optimization target formula of the composite difference index optimization algorithm is as follows:

[0025]

[0026] Among them, θ* represents the optimized generation network parameters; G θ (P) represents the result generated by the generation network according to the character feature map P under the generation network parameters θ, that is, the target comic character image; T is the ideal comic character image; is the adjustment hyperparameter; represents the optimal solution in the minimization problem, indicating that the generation network parameters θ are adjusted to minimize the composite loss function; is the exponential decay term.

[0027] Preferably, the S2 specifically includes:

[0028] In the implementation process of the composite difference index optimization algorithm, the generation network uses the optimized generation network parameters to generate a comic character optimized image based on the character feature map.

[0029] Preferably, the S2 specifically includes:

[0030] After the comic character optimized image is generated, a weighted repair algorithm based on error is used to repair the generated comic character optimized image. The weighted repair algorithm based on error compresses the error through a square root function, maintains the naturalness and original style of the comic character optimized image during the repair process, and finally generates a comic character digital avatar image through personalized adjustment.

[0031] The beneficial effects of the technical solution of the present invention are:

[0032] 1. By introducing personalized modeling, the present invention converts the user's preferences (such as gender, age, clothing style, etc.) into personalized feature vectors, and fuses them with the comic feature maps extracted by the convolutional neural network, effectively solving the problem of insufficient personalized needs, being able to generate comic character images that meet the specific requirements of users, greatly improving the personalization degree of image generation, and making the finally generated optimized comic character images more vivid and in line with the user's expectations.

[0033] 2. The present invention optimizes the generation network parameters through the composite difference index optimization algorithm, solving the problem of too large a difference between the target comic character image and the ideal comic character image in the prior art; the composite difference index optimization algorithm finely regulates the image generation process by combining the difference measurement based on the target comic character image and the ideal comic character image and the exponential decay term, reducing the error between the target comic character image and the ideal comic character image, and making the finally generated optimized comic character images more accurate and in line with the feature requirements of the ideal comic character.

[0034] 3. The present invention uses the error-based weighted repair algorithm to repair the generated optimized comic character images, effectively improving the overall quality and detail performance of the images; through the repair of the optimized comic character images, the possible defects are corrected, and at the same time, the natural sense and original style of the comic characters are maintained, making the finally generated digital avatar images of comic characters not only more artistic, but also have higher user satisfaction, meeting the personalized needs and expectations of users. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a flowchart of a method for generating a digital avatar of a comic character based on AI according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0036] In order to further elaborate on the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0037] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs.

[0038] The following specifically describes the specific solution of a method for generating a digital avatar of a comic character based on AI provided by the present invention in conjunction with the drawings.

[0039] Refer to the appendixFigure 1 , which shows a flowchart of a method for generating a digital avatar of a comic character based on AI provided by an embodiment of the present invention. The method includes the following steps:

[0040] 51: Collect and preprocess the original comic images to obtain comic images; input the comic images into the convolutional layer of a convolutional neural network for feature extraction to obtain a preliminary comic feature map, and then process the preliminary comic feature map through a feature mapping layer to obtain a comic feature map;

[0041] First, it is necessary to collect original comic images to form an original comic image dataset. The sources of the original comic images include public comic image libraries, authorized comic works, and personalized data provided by users, etc. The construction of the original comic image dataset should cover diverse character features, such as: gender, clothing, expression, background, etc., to ensure that the convolutional neural network (CNN) can learn comprehensive and rich feature information.

[0042] Perform preprocessing operations on the original comic images in the original comic image dataset. The preprocessing operations include unified processing of image sizes and data augmentation techniques. After the preprocessing operations, comic images are obtained, and a comic image dataset is formed by the comic images. The unified processing of image sizes adjusts all original comic images to 224x224 pixels to meet the input requirements of the convolutional neural network (CNN) model. The convolutional neural network (CNN) model is widely used in image recognition tasks and can effectively extract spatial features in images. The data augmentation techniques increase the diversity of the original comic image dataset through methods such as rotation, scaling, and flipping of images, preventing the convolutional neural network model from overfitting during training, thereby improving the robustness and generalization ability of the convolutional neural network model. The above preprocessing operations are well-known technical means to those skilled in the art and will not be elaborated here.

[0043] Furthermore, input the comic images into the convolutional layer and feature mapping layer of a convolutional neural network (CNN) to perform feature extraction on the comic images. The convolutional layer performs convolutional operations on the input comic images by applying multiple convolutional kernels to capture comic image features, such as edges, textures, and shapes, to obtain a preliminary comic feature map; then input the preliminary comic feature map into the feature mapping layer for weighted and non-linear processing to further enhance and refine the comic image features, forming a comic feature map. The comic feature map contains deep information of the comic images and can be used for subsequent personalized modeling and image generation tasks.

[0044] The implementation formula of the convolutional operation is as follows:

[0045]

[0046] Among them, F represents the preliminary comic feature map output by the convolution operation, which is a high-dimensional representation of the comic image features; n is the number of convolution kernels, which is set according to the specific implementation scenario; i is the index variable of the convolution kernel; K i is the i-th convolution kernel, which is used to extract specific types of comic image features, such as edges, textures, etc.; I i is the input comic image, which is the pixel matrix of the comic image; is the element-wise convolution operation, that is, element-wise multiplication; α i is the amplitude adjustment parameter, which is used to enhance the amplitude of the convolution output and enhance the comic image features, and is set according to the expert experience method; b i is the bias term of the convolution operation, which is used to adjust the preliminary comic feature map of the output; β i is the response adjustment parameter, which is used to adjust the response degree of the convolution output, so that the convolutional neural network model can better adapt to the changes of different comic image features when processing complex comic image information, and is set according to expert experience; the convolution kernel K i and the bias term b of the convolution operation i are obtained by training using a comic dataset in the convolutional neural network through the backpropagation algorithm. The training of the convolutional neural network belongs to the prior art and will not be elaborated here.

[0047] After the convolution operation, the preliminary comic feature map is obtained. By introducing the feature mapping formula, the preliminary comic feature map is further processed through weighted sum and non-linear mapping to improve the expression ability of the preliminary comic feature map. First, the preliminary comic feature map is weighted using the weight matrix of the feature mapping to adjust the importance of different features; then, the influence of extreme values is suppressed through logarithmic transformation to maintain stability; further, the response and expression of the features are enhanced using the exponential function; finally, the preliminary comic feature map is further processed using a non-linear activation function to make the final output more accurate and rich, providing a more reliable feature representation for subsequent personalized modeling and character generation.

[0048] The feature mapping formula is as follows:

[0049]

[0050] Among them, G is the comic feature map, representing the final output of the comic feature representation; σ is a non-linear activation function used to introduce non-linearity; F represents the preliminary comic feature map output by the convolution operation; W is the weight matrix of the feature mapping, used to weight the preliminary comic feature map and determine the relative importance of different features, which is obtained by training with a comic dataset in a convolutional neural network through the backpropagation algorithm; b′ is the bias term in the feature mapping process, used to adjust the comic feature map obtained by the feature mapping, which is obtained by training with a comic dataset in a convolutional neural network through the backpropagation algorithm; γ is a regulation parameter used to control the intensity of the feature mapping, which is set according to the expert experience method; log(1 + F·W + b′) represents the logarithmic transformation of the preliminary comic feature map weighted by the weight matrix, suppressing the instability caused by extreme values; exp represents the exponential operation.

[0051] S2: Obtain the personalized feature map and combine it with the comic feature map to generate the character feature map. Input the character feature map into the generation network, and use the composite difference index optimization algorithm to optimize the generation network parameters to generate the optimized comic character image, and repair the optimized comic character image to obtain the digital avatar image of the comic character.

[0052] In the personalized modeling stage, according to the user's personalized feature vector, generate the features of the target comic character. The user's personalized feature vector is generated based on the user's input preferences, such as the user's gender, age, clothing style, and facial expression, etc. The user input will be converted into numerical values and combined into a multi-dimensional vector. For example, gender is represented by 0 for male and 1 for female, and age is represented by 0, 1, 2, etc. for different age groups. Finally, all the user's preferences form the personalized feature vector, which is combined with the comic feature map extracted by the convolutional neural network to generate a comic character that meets the personalized needs.

[0053] To combine the personalized feature vector with the comic feature map extracted by the convolutional neural network, it is necessary to expand the dimension of the personalized feature vector to obtain a personalized feature map with the same dimension as the comic feature map, and then perform weighted and non-linear combination on the personalized feature map and the comic feature map to generate the character feature map.

[0054] The formula for personalized modeling is as follows:

[0055]

[0056] Among them, P is the generated character feature map; ξ is the weighting coefficient, which determines the weights based on the comic feature map G and the personalized feature map G′, and is set according to the expert experience method; G is the comic feature map; G 2 represents the calculation of the square of the elements in the comic feature map; G′ is the personalized feature map; G′ 3It represents the calculation of the cube of the elements in the personalized feature map, which is used to enhance the non - linear influence of the personalized feature map; δ is an exponential adjustment parameter used to control the intensity of the non - linear combination and is set according to the expert experience method.

[0057] The generated character feature map is input into the generation network to generate the target comic character image. To ensure the accuracy of the target comic character image, a composite difference index optimization algorithm is introduced to generate an optimized comic character image. The composite difference index optimization algorithm makes the target comic character image output by the generation network more precisely close to the target feature map by adjusting the generation network parameters. The specific implementation process is as follows:

[0058] The composite difference index optimization algorithm realizes optimization by minimizing a composite loss function. The composite loss function combines two parts: on the one hand, it is based on the difference measure between the target comic character image and the ideal comic character image; on the other hand, an exponential decay term is introduced to smooth the error and reduce the impact of large differences.

[0059] Furthermore, by setting adjustment hyperparameters, the weights of the two parts of the composite loss function can be finely adjusted, making the finally generated optimized comic character image more in line with the user's expectations.

[0060] The optimization objective formula of the composite difference index optimization algorithm is as follows:

[0061]

[0062] Among them, θ* represents the optimized generation network parameters; G θ (P) represents the result generated by the generation network according to the character feature map P under the generation network parameters θ, that is, the target comic character image; T is the ideal comic character image, which is set according to the specific implementation scenario; is an adjustment hyperparameter used to control the influence degree of the composite loss function and is set according to the expert experience method; represents the optimal solution in the minimization problem, which means adjusting the generation network parameters θ to minimize the composite loss function and optimize the generation network parameters; is the exponential decay term.

[0063] The composite difference index optimization algorithm makes the generated image more accurate by reducing the influence of large errors, improving the stability of the generation network and the image quality. The generation network uses the optimized generation network parameters to generate an optimized comic character image I gen .

[0064] After the optimized image of the comic character is generated, the generated optimized image of the comic character is repaired to improve the overall quality and detail performance of the image. A weighted repair algorithm based on error is adopted, and the error is compressed by a square root function to maintain the naturalness and original style of the optimized image of the comic character during the repair process. After the repair process, not only the defects that may occur during the generation process are corrected, but also the naturalness and original style of the character are maintained. At the same time, through personalized adjustment, the quality and detail performance of the image are further improved, and finally a digital avatar image of the comic character that meets the user's needs is generated.

[0065] The formula for image repair is as follows:

[0066]

[0067] Where, I final is the image obtained after repairing the optimized image of the comic character, that is, the digital avatar image of the comic character; I gen is the optimized image of the comic character, which is generated by the generation network using the optimized generation network parameters based on the character feature map; λ is the repair intensity coefficient, which is used to control the correction degree during the repair process and is set according to the expert experience method.

[0068] After the image repair is completed, the digital avatar image of the comic character is obtained and provided to the user.

[0069] In summary, a method for generating a digital avatar of a comic character based on AI is completed.

[0070] The sequence of the invention embodiments is only for description and does not represent the superiority or inferiority of the embodiments. The processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0071] Each embodiment in this specification is described in a progressive manner. The same or similar parts between each embodiment can be referred to each other, and the key points of each embodiment are the differences from other embodiments.

[0072] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of each embodiment of the present invention, and should all be included in the protection scope of the present invention.

Claims

1. A method for generating a digital avatar of a comic character based on AI, characterized in that: The following steps are involved: S1: Collect and preprocess the original comic image to obtain a comic image; input the comic image into the convolution layer of the convolutional neural network for feature extraction to obtain a preliminary comic feature map, and then process the preliminary comic feature map through the feature mapping layer to obtain a comic feature map; 52: Obtain the personalized feature map and combine it with the comic feature map to generate a character feature map, input the character feature map into the generation network, use the composite difference index optimization algorithm to optimize the generation network parameters, generate the comic character optimized image, and repair the comic character optimized image to obtain the comic character digital clone image.

2. The method for generating a digital avatar of a comic character based on AI according to claim 1, characterized in that: The S1 specifically includes: The comic image is input into the convolution layer and feature mapping layer of the convolutional neural network to extract features of the comic image. The convolution layer performs a convolution operation on the input comic image by applying a convolution kernel to obtain a preliminary comic feature map.

3. The method for generating a digital avatar of a comic character based on AI according to claim 2, characterized in that: The S1 specifically includes: Based on the preliminary comic feature map, a feature mapping formula is introduced, and the preliminary comic feature map is processed through weighted and nonlinear mapping.

4. The method for generating a digital avatar of a comic character based on AI according to claim 3, characterized in that: The S1 specifically includes: In the process of implementing the feature mapping formula, the weight matrix of the feature mapping is used to weight the preliminary comic feature map to adjust the importance of different features; the influence of extreme values ​​is suppressed by logarithmic transformation; and the preliminary comic feature map is processed using exponential function and nonlinear activation function to obtain the comic feature map.

5. The method for generating a digital avatar of a comic character based on AI according to claim 4, characterized in that: The S1 specifically includes: The feature mapping formula is as follows: Among them, G is the comic feature map; σ is the nonlinear activation function; F represents the preliminary comic feature map output by the convolution operation; W is the weight matrix of the feature mapping; b′ is the bias term in the feature mapping process; γ is the adjustment parameter; log(1+F·W+b′) represents the logarithmic transformation of the preliminary comic feature map after weighting by the weight matrix; exp represents the exponential operation.

6. The method for generating a digital avatar of a comic character based on AI according to claim 1, characterized in that: The S2 specifically includes: The formula for personalized modeling is as follows: Among them, P is the generated character feature map; ξ is the weighting coefficient; G is the comic feature map; G 2 represents the quadratic calculation of the elements in the comic feature map; G′ is the personalized feature map; G′ 3 It indicates that the elements in the personalized feature map are raised to the third power; δ is the exponential adjustment parameter.

7. The method for generating a digital avatar of a comic character based on AI according to claim 1, characterized in that: The S2 specifically includes: In the implementation process of the composite difference index optimization algorithm, the generation network parameters are optimized by minimizing the composite loss function. The optimization objective formula of the composite difference index optimization algorithm is as follows: Among them, θ * represents the optimized generation network parameters; G θ (P) represents the result generated by the generative network under the generative network parameters θ according to the character feature map P, i.e., the target comic character image; T is the ideal comic character image; is to adjust the hyperparameters; represents the optimal solution in the minimization problem, which means minimizing the composite loss function by adjusting the generation network parameters θ; is an exponential decay term.

8. The method for generating a digital avatar of a comic character based on AI according to claim 7, characterized in that: The S2 specifically includes: In the implementation of the composite difference index optimization algorithm, the generative network uses the optimized generative network parameters to generate optimized images of comic characters based on the character feature map.

9. The method for generating a digital avatar of a comic character based on AI according to claim 8, characterized in that: The S2 specifically includes: After the comic character optimized image is generated, an error-based weighted restoration algorithm is used to restore the generated comic character optimized image. The error-based weighted restoration algorithm compresses the error through a square root function, maintains the naturalness and original style of the comic character optimized image during the restoration process, and finally generates a comic character digital clone image through personalized adjustment.

Images