Traditional chinese painting style transfer method and system based on laplacian pyramid

Abstract

This invention discloses a method and system for transferring the style of traditional Chinese painting based on the Laplacian pyramid, belonging to the field of computer vision technology. The method constructs a multi-scale style transfer model based on the Laplacian pyramid. The multi-scale style transfer model includes a Laplacian pyramid decomposition module, a basic style transfer network, a detail enhancement network, an edge information selection module, and a Laplacian pyramid reconstruction module. Preprocessed content image datasets and preprocessed style image datasets are input into the multi-scale style transfer model for training. Style loss functions and content loss functions are used to optimize the multi-scale style transfer model, resulting in a multi-scale style transfer generation model. The target content image is then input into the multi-scale style transfer generation model to generate a stylized image in the style of traditional Chinese painting. This invention can generate high-quality stylized images with the style of traditional Chinese painting.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a method and system for transferring the style of traditional Chinese painting based on the Laplace pyramid. Background Technology

[0002] With the continuous expansion of computer vision technology, style transfer, as a new image processing technique, is being used more and more widely. The task of style transfer is to generate a stylized image by combining the basic content structure of a content image with a style template of a style image. The stylized image retains the basic content structure of the content image, such as the shape and outline of objects, while maintaining an overall color and texture style similar to the style image.

[0003] While numerous style transfer models based on neural networks exist, most utilize Western paintings with rich colors and complex textures as style images for transfer. Unlike most Western paintings, traditional Chinese paintings are characterized by clean composition, abstract content, simple layout, and muted tones. When existing style transfer models are used to directly transfer the style of traditional Chinese paintings, the resulting stylized images often exhibit numerous messy and unnatural textures, negatively impacting the overall visual effect. Summary of the Invention

[0004] In view of the characteristics of traditional Chinese painting, this invention provides a method and system for transferring the style of traditional Chinese painting based on the Laplace Pyramid.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] In a first aspect, the present invention provides a method for transferring the style of traditional Chinese painting based on the Laplace Pyramid, comprising:

[0007] Obtain a content image dataset and a style image dataset, and preprocess the content image dataset and the style image dataset respectively; the style images are in the traditional Chinese painting style.

[0008] A multi-scale style transfer model based on the Laplacian pyramid is constructed; the multi-scale style transfer model includes a Laplacian pyramid decomposition module, a basic style transfer network, a detail enhancement network, an edge information selection module, and a Laplacian pyramid reconstruction module;

[0009] The preprocessed content image dataset and the preprocessed style image dataset are input into the multi-scale style transfer model for training, and the style loss function and content loss function are used to optimize the multi-scale style transfer model to obtain a multi-scale style transfer generation model.

[0010] The target content image is input into a multi-scale style transfer generation model to generate a stylized image in the style of traditional Chinese painting.

[0011] Secondly, the present invention also provides a Chinese traditional painting style transfer system based on the Laplace Pyramid, comprising:

[0012] A data preprocessing module is used to acquire a content image dataset and a style image dataset, and to preprocess the content image dataset and the style image dataset respectively; the style images are in the traditional Chinese painting style.

[0013] A multi-scale style transfer model construction module is used to construct a multi-scale style transfer model based on the Laplace pyramid; the multi-scale style transfer model includes a Laplace pyramid decomposition module, a basic style transfer network, a detail enhancement network, an edge information selection module, and a Laplace pyramid reconstruction module.

[0014] The multi-scale style transfer generation model generation module is used to input the preprocessed content image dataset and the preprocessed style image dataset into the multi-scale style transfer model for training, and to optimize the multi-scale style transfer model using style loss function and content loss function to obtain a multi-scale style transfer generation model.

[0015] The stylized image generation module is used to input the target content image into the multi-scale style transfer generation model to generate a stylized image in the style of traditional Chinese painting.

[0016] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0017] This invention discloses a method and system for transferring the style of traditional Chinese painting based on the Laplace pyramid. Targeting the characteristics of traditional Chinese painting—clean composition, abstract content, simple layout, and subdued color tones—this invention utilizes a deep neural network model and the Laplace pyramid image decomposition and reconstruction method to achieve multi-scale transfer learning, ultimately generating high-quality stylized images with the style of traditional Chinese painting. The deep neural network model provided in this embodiment includes a basic style transfer network, a detail enhancement network, and an edge information selection module. The basic style transfer network performs initial style transfer at low resolution, generating a low-resolution stylized image from a low-resolution content image and a low-resolution style image. The detail enhancement network progressively enhances the details of the stylized image at high resolution, including local textures and content structure details, ultimately generating a high-resolution stylized image. The edge information module selects more important semantic edge information from the detail enhancement network for training, ensuring that the final stylized image retains key content structures. Therefore, the method or system provided by this invention can generate high-quality stylized images with the style of traditional Chinese painting. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating a method for transferring traditional Chinese painting styles based on the Laplace Pyramid, provided as an embodiment of the present invention;

[0020] Figure 2 A framework diagram of a multi-scale style transfer model based on the Laplace pyramid provided in an embodiment of the present invention;

[0021] Figure 3 A schematic diagram of the style attention module provided in an embodiment of the present invention.

[0022] Figure 4 This is a schematic diagram of the edge information selection module provided in an embodiment of the present invention.

[0023] Figure 5 A comparison chart of qualitative experimental results of the present invention and five existing methods provided for embodiments of the present invention;

[0024] Figure 6This is a schematic diagram of a Chinese traditional painting style transfer system based on the Laplace Pyramid, provided as an embodiment of the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0027] In recent years, deep learning-based style transfer methods have developed rapidly in the field of computer vision technology, with Gatys et al. ( <Gatys,L. A.,Ecker,A. S.,&Bethge,M. (2016). Image style transferusing convolutional neural networks. In Proceedings of the IEEE conference oncomputer vision and pattern recognition (pp. 2414-2423)> This paper proposes an inspiring image-based style transfer method that utilizes pre-trained VGG (Image Update Classification) data. <Simonyan,K.,&Zisserman,A.(2014). Very deep convolutional networks for large-scale image recognition.arXiv preprint arXiv:1409.1556> A network extracts correlations between features to iteratively update a white noise image, ultimately yielding a stylized image. Inspired by Gatys et al., Johnson et al. ( <Johnson,J.,Alahi,A.,&Fei-Fei,L. (2016,October). Perceptual losses for real-time styletransfer and super-resolution. In European conference on computer vision (pp.694-711). Springer,Cham> ) and Ulyanov et al. <Ulyanov,D.,Lebedev,V.,Vedaldi,A.,&Lempitsky,V. (2016). Texture networks: Feed-forward synthesis of textures andstylized images. arXiv preprint arXiv:1603.03417> More efficient model-update-based style transfer methods are proposed, which train a feedforward neural network using a style image and then generate arbitrary content images in real time. Both methods can accelerate the training optimization process.AdaIN( <Huang,X.,&Belongie,S. (2017). Arbitrary style transfer in real-time with adaptiveinstance normalization. In Proceedings of the IEEE international conferenceon computer vision (pp. 1501-1510)> WCT utilizes adaptive instantiation normalization to learn multiple styles during training, enabling the generation of stylized images of any style after training. <Li,Y.,Fang,C.,Yang,J.,Wang,Z.,Lu,X.,&Yang,M. H. (2017). Universal style transfer via featuretransforms. Advances in neural information processing systems,30> This method integrates content and stylistic features using whitening and colorization transformations, then decodes the integrated stylized features through a series of decoders, ultimately generating stylized images of any style. Specifically addressing the characteristics of traditional Chinese painting, Li et al. ( <Li,B.,Xiong,C.,Wu,T.,Zhou,Y.,Zhang,L.,&Chu,R. (2018,December). Neuralabstract style transfer for chinese traditional painting. In Asian Conferenceon Computer Vision (pp. 212-227). Springer,Cham> A novel extended Gaussian difference method is proposed to preserve the abstract characteristics of traditional Chinese painting during style transfer.

[0028] Unlike these methods, this invention utilizes a multi-scale learning approach. Based on Laplace pyramid decomposition and image reconstruction, it employs the provided basic style transfer network, detail enhancement network, and edge information selection module to achieve high-quality style transfer of traditional Chinese paintings. Both qualitative and quantitative comparisons show results superior to five recently proposed style transfer methods with better performance. A literature search revealed no publicly available reports identical to this invention.

[0029] Example 1

[0030] Based on multi-scale learning, and considering the characteristics of traditional Chinese painting such as clean composition, abstract content, simple layout, and plain color scheme, this invention provides a method for transferring the style of traditional Chinese painting based on the Laplace Pyramid, specifically including the following steps:

[0031] Step 100: Obtain the content image dataset and the style image dataset, and preprocess the content image dataset and the style image dataset respectively.

[0032] In this embodiment of the invention, step 100 selects a suitable image dataset as the content image dataset, and simultaneously collects and selects a portion of traditional Chinese paintings as the style image dataset. Preprocessing operations such as augmentation and resizing are performed on both the content image dataset and the style image dataset to make them suitable for model training and testing. The resized images are uniformly 512 pixels in size. 512.

[0033] Step 200: Construct a multi-scale style transfer model based on the Laplace pyramid; the multi-scale style transfer model, such as... Figure 2 As shown, it includes a Laplacian pyramid decomposition module, a basic style transfer network, a detail enhancement network, an edge information selection module, and a Laplacian pyramid reconstruction module.

[0034] exist Figure 2 middle, , , These represent content images at high, medium, and low resolutions, respectively. A style of image representing low resolution; , These represent content residual images at high and medium resolution, respectively. The original edge feature map representing the content image at medium resolution; , These represent stylized residual images at high and medium resolution, respectively. , , These represent stylized images at high, medium, and low resolutions, respectively.

[0035] In this embodiment of the invention, the preprocessed content image dataset includes multiple preprocessed content images, and the preprocessed style image dataset includes multiple preprocessed style images. Both the preprocessed content images and the preprocessed style images are high-resolution images.

[0036] In the Laplace pyramid decomposition module,

[0037] (1) The preprocessed content image is decomposed using the Laplacian pyramid method to obtain a low-resolution content image, a medium-resolution content residual image, and a high-resolution content residual image.

[0038] (2) The preprocessed style image is decomposed using the Laplacian pyramid method to obtain a low-resolution style image, a medium-resolution style residual image, and a high-resolution style residual image.

[0039] Using the Laplacian pyramid method to decompose preprocessed high-resolution images (uniformly sized and preprocessed content and style images) yields low-resolution images (content and style images), medium-resolution residual images (content and style images), and high-resolution residual images (content and style images). Specifically, each time a high-resolution content image is processed, one low-resolution content image, one medium-resolution content residual image, and one high-resolution content residual image are obtained. Similarly, each time a high-resolution style image is processed, one low-resolution style image, one medium-resolution style residual image, and one high-resolution style residual image are obtained.

[0040] Low-resolution images are used in the base style transfer network to generate low-resolution stylized images. Medium / high-resolution residual images are used in the detail enhancement network to generate high-resolution stylized residual images.

[0041] In the basic style transfer network, a low-resolution stylized image is generated based on the low-resolution content image and the low-resolution style image.

[0042] The basic style transfer network includes a pre-trained VGG-19 network, two style attention modules, and a decoder network that is symmetric to the VGG-19 network.

[0043] Specifically:

[0044] The first step is to use a pre-trained VGG-19 network as an encoder, inputting low-resolution content images and low-resolution style images respectively. Content features and style features are extracted from the ReLU_4_1 and ReLU_5_1 layers of the VGG-19 network respectively, and used for style embedding.

[0045] The second step involves using a style attention module to integrate style features into content features, while preserving the content structure as much as possible and embedding rich style templates based on multi-scale features. Specifically, for example... Figure 3 As shown, after extracting content features and style features, the attention map between content features and style features is first calculated. Figure 3 middle , , These represent content features, style features, and stylistic features, respectively. The specific formulas are shown below:

[0046] ;

[0047] in The characteristics represent those after mean variance channel normalization. , , , All of these are implemented using convolutional layers. The resulting attention map is then passed through a convolutional layer and added to the original content feature elements to obtain the final stylized feature. This feature integrates the pattern and texture of the style image using the spatial distribution of the content image's semantics. The first style attention module embeds the content and style features from the ReLU_4_1 layer to obtain a stylized feature. The second style attention module then embeds the content and style features from the ReLU_5_1 layer to obtain another stylized feature. Finally, these two stylized features are summed element-wise to obtain the final stylized feature.

[0048] The third step involves reconstructing the stylized features using a decoder network symmetrical to VGG-19, outputting a low-resolution stylized image.

[0049] In the aforementioned detail enhancement network,

[0050] Upsampling is performed on the low-resolution content image and the low-resolution stylized image respectively to obtain the upsampled low-resolution content image and the upsampled low-resolution stylized image.

[0051] Based on the medium-resolution content residual image, the upsampled low-resolution content image, and the upsampled low-resolution stylized image, an edge information selection module and a convolutional activation function layer are used to output a medium-resolution stylized residual image.

[0052] Furthermore, upsampling is performed on both the low-resolution content image and the low-resolution stylized image to obtain upsampled low-resolution content images and upsampled low-resolution stylized images, specifically including:

[0053] (1) The low-resolution content image is upsampled using the bidirectional interpolation method to obtain the upsampled low-resolution content image.

[0054] (2) The low-resolution stylized image is upsampled using the bidirectional interpolation method to obtain the upsampled low-resolution stylized image.

[0055] The detail enhancement network includes several convolutional activation function layers, several residual blocks, and one convolutional layer; the convolutional activation function layer is that each convolutional layer is connected to a leaky ReLU layer as an activation function.

[0056] Furthermore, based on the medium-resolution content residual image, the upsampled low-resolution content image, and the upsampled low-resolution stylized image, an edge information selection module and a convolutional activation function layer are used to output a medium-resolution stylized residual image, specifically including:

[0057] a) Take the medium-resolution content residual image, the upsampled low-resolution content image, and the upsampled low-resolution stylized image as input, and output an image feature map after passing through several convolutional activation layers.

[0058] b. An edge information selection module is used to select edge information from more important content images and incorporate it into the detail enhancement network. This ensures that the final high-resolution stylized image retains key content structures, such as object outlines. The edge information selection module is implemented through a channel attention module, which adaptively integrates local features along the channel dimension from a global perspective. This approach enhances the semantic expression of features and leverages the interdependencies between channel maps to emphasize related feature maps. Specifically, for example... Figure 4 As shown, after the convolutional layer, the original edge feature map of the medium-resolution content image is used as the input to the edge information selection module. First, the channel attention map is calculated. Figure 4 middle, Edge images representing medium-resolution content images; The edge features representing the content image are shown in the following formula:

[0059] ;

[0060] in This represents the features of the edge map obtained after the convolutional layer. Then, the resulting channel attention map is multiplied by the original edge feature map using a matrix multiplication operation, and finally added element-wise to the original edge feature map to obtain the final output edge feature map, as shown in the following formula:

[0061] ;

[0062] in This represents a weight that can be learned from scratch. The final result is a selected edge feature. This feature is then added to the detail enhancement network and concatenated with the feature obtained in the first step.

[0063] c. The image feature map and the final edge feature map output by the edge information selection module are stitched together, and the stitched feature map is passed through several residual blocks, a convolutional activation layer and a convolutional layer in sequence to output a medium-resolution stylized residual image.

[0064] Similarly, in the aforementioned detail enhancement network,

[0065] Upsampling is performed on the medium-resolution content image and the medium-resolution stylized residual image to obtain the upsampled medium-resolution content image and the upsampled medium-resolution stylized residual image.

[0066] Based on the high-resolution content residual image, the upsampled medium-resolution content image, and the upsampled medium-resolution stylized image, a high-resolution stylized residual image is obtained by using convolutional activation layers, residual blocks, and convolutional layers.

[0067] In the Laplace pyramid reconstruction module,

[0068] The medium-resolution stylized image is obtained by summing the elements of the upsampled low-resolution stylized image and the medium-resolution stylized residual image.

[0069] Upsampling is performed on the medium-resolution stylized image to obtain the upsampled medium-resolution stylized image;

[0070] The high-resolution stylized image is obtained by summing the elements of the upsampled medium-resolution stylized image and the high-resolution stylized residual image.

[0071] At this point, the multi-scale style transfer model based on the Laplace pyramid has been completed.

[0072] Step 300: Input the preprocessed content image dataset and the preprocessed style image dataset into the multi-scale style transfer model for training, and use the style loss function and content loss function to optimize the multi-scale style transfer model to obtain a high-performance multi-scale style transfer generation model.

[0073] During model training, style loss and content loss functions are used to optimize the overall model. First, the content image, style image, and generated stylized image are input into a pre-trained VGG-19 network, where content features are obtained at different layers. Stylistic Features and stylistic features , Represents different layers of VGG-19.

[0074] The first style loss function is calculated by summing the mean difference and variance difference between style features and stylized features, and the specific formula is as follows: (1)

[0075] (1);

[0076] in and The mean and variance of the feature map are represented.

[0077] The second style loss function calculates the relaxed land movement distance between the style features and the stylized features, which is used as the second style loss value. The specific formula is as follows:

[0078] (2);

[0079] in, This is the cost matrix, which is obtained by calculating the cosine distance between style features and stylized features. The calculation formula is as follows:

[0080] (3);

[0081] The first content loss function calculates the Euclidean distance between content features and stylistic features, using it as the first content loss value. The calculation formula is as follows:

[0082] (4);

[0083] The second content loss function is the self-similarity loss function, which can maintain the similarity between content features and style features. The specific formula is as follows:

[0084] (5);

[0085] in The height of representative features The width of the representative feature. and The cosine distance between content features and the cosine distance between stylized features can be obtained by formula (3).

[0086] Determine if the loss value converges

[0087] The overall content loss function of the model is expressed as:

[0088]

[0089] The overall style loss function is expressed as:

[0090]

[0091] in , , , The values ​​are set to 1, 15, 50, and 80 respectively. The Adam method is used to optimize the update model, reducing the content loss value. and style loss value Convergence has been achieved. If the content loss value ultimately reaches 9.62... The style loss was 0.01, ultimately reaching 33.37. If the value is 0.01, it means the loss value has converged, and we proceed to the next step; if neither the content loss value nor the style loss value has reached the above value, we return to the previous steps and repeat.

[0092] Determine whether the generated stylized image meets the expected effect.

[0093] The average PSNR, SSIM, and LPIPS values ​​are calculated for 200 generated stylized images and their respective style images. If all three results are higher than the average PSNR, SSIM, and LPIPS values ​​calculated for 200 stylized images of the same style generated by the other five style transfer methods and the style image, it indicates that the generated stylized images have met the expected results, and the model optimization is complete. Otherwise, return to the previous steps and repeat.

[0094] Step 400: Input the target content image into the multi-scale style transfer generation model to generate a stylized image in the style of traditional Chinese painting.

[0095] During use, different style images can be used to train and generate different multi-scale style transfer generation models. Then, any content image can be used to generate any stylized image of a certain style.

[0096] To verify the effectiveness of the present invention, this example trains models for different style transfers based on images of different styles, and then generates high-quality stylized images.

[0097] Through qualitative comparison, such as Figure 5 As shown in Table 1, comparing the stylized images generated by this invention with those generated by other mainstream style transfer models, it can be seen that the visual quality of this invention is significantly higher than that of the other models. Through quantitative comparison, as shown in Table 1, comparing the metrics between the stylized images generated by this invention and those generated by other mainstream style transfer models, it can be seen that this invention significantly outperforms other models in terms of image quality and similarity to the style images. This invention demonstrates excellent performance in style transfer for traditional Chinese painting.

[0098] Table 1 Quantitative Comparison Table

[0099]

[0100] To address the issue that most current style transfer models only apply style transfer to Western oil paintings with rich colors and complex textures at a single scale, this paper proposes a multi-scale approach to achieve style transfer tailored to the characteristics of traditional Chinese painting. The Laplacian pyramid is used for multi-scale training, allowing for selective use of information from different scales of content and style images. At low resolution, the basic style transfer network can roughly transfer the color distribution and texture of the style image, discarding some unimportant details while retaining the overall basic content structure, thus saving significant training time. At high resolution, the detail enhancement network gradually adds details to the stylized image, including local style details, while an edge information selection module selectively retains more important semantic information from content edges. The overall model improves the quality of the generated images while maintaining efficiency. This algorithm outperforms existing algorithms both in terms of intuitive results and objective performance.

[0101] Example 2

[0102] In order to implement the method corresponding to Embodiment 1 above and achieve the corresponding functions and technical effects, a Chinese traditional painting style transfer system based on the Laplace Pyramid is provided below.

[0103] like Figure 6 As shown, the present invention also provides a Chinese traditional painting style transfer system based on the Laplace Pyramid, comprising:

[0104] Data preprocessing module 1 is used to acquire a content image dataset and a style image dataset, and to preprocess the content image dataset and the style image dataset respectively; the style images are in the traditional Chinese painting style.

[0105] Multi-scale style transfer model construction module 2 is used to construct a multi-scale style transfer model based on the Laplace pyramid; the multi-scale style transfer model includes a Laplace pyramid decomposition module, a basic style transfer network, a detail enhancement network, an edge information selection module, and a Laplace pyramid reconstruction module.

[0106] The multi-scale style transfer generation model generation module 3 is used to input the preprocessed content image dataset and the preprocessed style image dataset into the multi-scale style transfer model for training, and to optimize the multi-scale style transfer model using the style loss function and the content loss function to obtain a multi-scale style transfer generation model.

[0107] Stylized image generation module 4 is used to input the target content image into the multi-scale style transfer generation model to generate a stylized image in the style of traditional Chinese painting.

[0108] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0109] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for traditional Chinese painting style transfer based on Laplacian pyramid, characterized in that, include: Obtain a content image dataset and a style image dataset, and preprocess the content image dataset and the style image dataset respectively; The style of the image is in the traditional Chinese painting style; A multi-scale style transfer model based on the Laplacian pyramid is constructed; the multi-scale style transfer model includes a Laplacian pyramid decomposition module, a basic style transfer network, a detail enhancement network, an edge information selection module, and a Laplacian pyramid reconstruction module; In the Laplacian pyramid decomposition module, the Laplacian pyramid method is used to decompose the preprocessed content image to obtain a low-resolution content image, a medium-resolution content residual image, and a high-resolution content residual image. The preprocessed style image is decomposed using the Laplacian pyramid method to obtain a low-resolution style image, a medium-resolution style residual image, and a high-resolution style residual image. The basic style transfer network includes a pre-trained VGG-19 network, two style attention modules, and a decoder network symmetrical to the VGG-19 network; in the basic style transfer network, a low-resolution stylized image is generated based on a low-resolution content image and a low-resolution style image. The detail enhancement network includes several convolutional activation function layers, several residual blocks, and a convolutional layer. The convolutional activation function layer includes a convolutional layer and an activation function layer connected to the output of the convolutional layer. In the detail enhancement network, the low-resolution content image and the low-resolution stylized image are upsampled to obtain an upsampled low-resolution content image and an upsampled low-resolution stylized image. Based on the medium-resolution content residual image, the upsampled low-resolution content image, and the upsampled low-resolution stylized image, an edge information selection module and a convolutional activation function layer are used to output a medium-resolution stylized residual image. Specifically, this includes: taking the medium-resolution content residual image, the upsampled low-resolution content image, and the upsampled low-resolution stylized image as input, passing them through several convolutional activation layers to output an image feature map; concatenating the image feature map and the final edge feature map output by the edge information selection module, and then passing the concatenated feature map through several residual blocks, a convolutional activation layer, and a convolutional layer in sequence to output a medium-resolution stylized residual image. In the edge information selection module, a channel attention map is calculated based on the original edge feature map of the medium-resolution content image; a matrix multiplication operation is performed on the channel attention map and the original edge feature map to obtain the target feature map; and the target feature map is added to the elements of the original edge feature map to obtain the final edge feature map. In the Laplacian pyramid reconstruction module, the upsampled low-resolution stylized image and the medium-resolution stylized residual image are summed to obtain the medium-resolution stylized image; the medium-resolution stylized image is upsampled to obtain the upsampled medium-resolution stylized image; the upsampled medium-resolution stylized image and the high-resolution stylized residual image are summed to obtain the high-resolution stylized image. The preprocessed content image dataset and the preprocessed style image dataset are input into the multi-scale style transfer model for training, and the style loss function and content loss function are used to optimize the multi-scale style transfer model to obtain a multi-scale style transfer generation model. The target content image is input into a multi-scale style transfer generation model to generate a stylized image in the style of traditional Chinese painting. 2.The traditional Chinese painting style transfer method based on Laplacian pyramid according to claim 1, wherein, The preprocessed content image dataset includes multiple preprocessed content images, and the preprocessed style image dataset includes multiple preprocessed style images. 3.The traditional Chinese painting style transfer method based on Laplacian pyramid of claim 1, wherein, The process of upsampling the low-resolution content image and the low-resolution stylized image to obtain the upsampled low-resolution content image and the upsampled low-resolution stylized image specifically includes: A bidirectional interpolation method is used to upsample the low-resolution content image to obtain an upsampled low-resolution content image. The low-resolution stylized image is upsampled using a bidirectional interpolation method to obtain the upsampled low-resolution stylized image.

4. The method of claim 1, wherein the method is based on a Laplacian pyramid. In the aforementioned detail enhancement network, Upsampling is performed on the medium-resolution content image and the medium-resolution stylized residual image to obtain the upsampled medium-resolution content image and the upsampled medium-resolution stylized residual image. Based on the high-resolution content residual image, the upsampled medium-resolution content image, and the upsampled medium-resolution stylized image, a high-resolution stylized residual image is obtained by using convolutional activation layers, residual blocks, and convolutional layers. 5.A Laplacian pyramid based traditional Chinese painting style transfer system, characterized in that, include: A data preprocessing module is used to acquire a content image dataset and a style image dataset, and to preprocess the content image dataset and the style image dataset respectively; The style of the image is in the traditional Chinese painting style; A multi-scale style transfer model construction module is used to construct a multi-scale style transfer model based on the Laplace pyramid; the multi-scale style transfer model includes a Laplace pyramid decomposition module, a basic style transfer network, a detail enhancement network, an edge information selection module, and a Laplace pyramid reconstruction module. In the Laplacian pyramid decomposition module, the Laplacian pyramid method is used to decompose the preprocessed content image to obtain a low-resolution content image, a medium-resolution content residual image, and a high-resolution content residual image. The preprocessed style image is decomposed using the Laplacian pyramid method to obtain a low-resolution style image, a medium-resolution style residual image, and a high-resolution style residual image. The basic style transfer network includes a pre-trained VGG-19 network, two style attention modules, and a decoder network symmetrical to the VGG-19 network; in the basic style transfer network, a low-resolution stylized image is generated based on a low-resolution content image and a low-resolution style image. The detail enhancement network includes several convolutional activation function layers, several residual blocks, and a convolutional layer. The convolutional activation function layer includes a convolutional layer and an activation function layer connected to the output of the convolutional layer. In the detail enhancement network, the low-resolution content image and the low-resolution stylized image are upsampled to obtain an upsampled low-resolution content image and an upsampled low-resolution stylized image. Based on the medium-resolution content residual image, the upsampled low-resolution content image, and the upsampled low-resolution stylized image, an edge information selection module and a convolutional activation function layer are used to output a medium-resolution stylized residual image. Specifically, this includes: taking the medium-resolution content residual image, the upsampled low-resolution content image, and the upsampled low-resolution stylized image as input, passing them through several convolutional activation layers to output an image feature map; concatenating the image feature map and the final edge feature map output by the edge information selection module, and then passing the concatenated feature map through several residual blocks, a convolutional activation layer, and a convolutional layer in sequence to output a medium-resolution stylized residual image. In the edge information selection module, a channel attention map is calculated based on the original edge feature map of the medium-resolution content image; a matrix multiplication operation is performed on the channel attention map and the original edge feature map to obtain the target feature map; and the target feature map is added to the elements of the original edge feature map to obtain the final edge feature map. In the Laplacian pyramid reconstruction module, the upsampled low-resolution stylized image and the medium-resolution stylized residual image are summed to obtain the medium-resolution stylized image; the medium-resolution stylized image is upsampled to obtain the upsampled medium-resolution stylized image; the upsampled medium-resolution stylized image and the high-resolution stylized residual image are summed to obtain the high-resolution stylized image. The multi-scale style transfer generation model generation module is used to input the preprocessed content image dataset and the preprocessed style image dataset into the multi-scale style transfer model for training, and to optimize the multi-scale style transfer model using style loss function and content loss function to obtain a multi-scale style transfer generation model. The stylized image generation module is used to input the target content image into the multi-scale style transfer generation model to generate a stylized image in the style of traditional Chinese painting.

Images