Improved style transfer algorithm in decorative design of ceramic painting

Abstract

The art of decorating Chinese ceramic products is a precious cultural heritage. However, the illustration techniques used in ceramic decoration face challenges in terms of inheritance and protection due to limited transmission methods and low public awareness. In the era of artificial intelligence, traditional ceramic painting techniques are undergoing transformation. To address the lack of systematic guidance in ceramic painting pattern design and the uncertainty of desired design effects, an improved style transfer algorithm is proposed for decorating ceramic products. The algorithm utilizes the watershed algorithm to segment and extract pixels, ensuring shape consistency during the migration process. Additionally, local affine transformation preserves the integrity of points, lines, and surfaces. The designed improved style transfer algorithm is applied in practice, facilitating the design of ceramic cultural and creative products. This advancement simplifies the expression of decorative patterns in ceramic painting, promotes intelligent development in ceramic decoration design, and enables newcomers and the general public to engage in art creation, ultimately safeguarding and inheriting the techniques of ceramic decoration.

Keywords:

• ceramic decoration
• style transfer algorithm
• cultural heritage
• inheritance and protection

Public interest statement

Our research aims to promote the preservation and inheritance of ceramic painting and decoration art to meet the public’s demand for cultural diversity and creative expression. By improving style transfer algorithms, we provide an intelligent solution for ceramic cultural and creative product design. This facilitates the transmission of ceramic painting and decoration techniques and enables designers and the general public to participate more effectively in art creation. Our study strives to drive the modern transformation of ceramic art, protect the cultural treasure of Chinese civilization, and offer increased opportunities for artistic creation and enjoyment, catering to the interests and expectations of the public.

1. Introduction

Ceramics serve as a significant medium for preserving Chinese culture, and the decorative illustrations on ceramic products epitomize the distinctive style of Chinese ceramic art (Liu, 2022). Painting decoration is a prevalent form of expression in ceramic art design, allowing for the manifestation of artistic essence, cultural significance, and increased value of the artwork (Ci, 2020). Traditional Chinese ceramic paintings encompass various decorative patterns, such as blue and white, famille rose, ancient color, and underglaze red. These patterns can be employed individually or combined to create a distinctive and unique decorative style.

Figure 1. Ceramic pattern style.

As shown in Figure 1, the decorative patterns of traditional ceramic paintings have a variety of textures, and excellent works show exquisite artistic characteristics in their ornamentation and color (Ren, 2022). As outstanding representatives of art works, ceramics often have their own differences in composition, color, technique, etc. Artists often use the content of ceramic painting to express their attitude towards life and the pursuit of art (Qin, 1980). Therefore, ceramic painting art is a treasure of traditional Chinese culture, a combination of design art and spiritual aesthetics, and its design form and aesthetics are in turn important design elements of ceramic art (Yu et al., 2022).

However, in recent years, the modern ceramic art design has witnessed a shift towards diversified, comprehensive, and cosmopolitan development. This has led to an increased focus on the connotation expressed by ceramic painting decoration and a desire to promote the beauty of ceramic painting decoration that is closely aligned with everyday life. Traditional design methods are no longer sufficient to meet the multifaceted requirements of ceramic painting art. As humanity has entered a new era of science and technology, the advancement of the art of decorating Chinese ceramic products has surpassed any previous period in history. The application of computers, the Internet, and artificial intelligence has made information exchange more convenient and ideas more accessible. These advancements have made it possible to express and imagine compositions that were previously challenging to achieve.

Currently, the art of decorating Chinese ceramic products has experienced a significant elevation in thought, technological innovation, and continuous improvement in technique. However, in the process of modern development, this art form faces the challenge of inheritance and protection due to limited dissemination methods and low public awareness. Consequently, enhancing the general public’s awareness of the decorative pattern styles of ceramic painting and finding ways to better preserve and develop the art of decorating Chinese ceramic products have become urgent issues that need to be addressed.

2. Research aim

In order to make the art of decorating Chinese ceramic products will not be forgotten by the public and not lose in the long river of history, it is important to preserve the typical design elements of the art style characteristic of ceramic painting in the past, and find new ways to use artificial intelligence methods to protect and inherit them, so that the traditional decorative technique of ceramic painting not only has a new creative language, but also can be carried forward in various cultural and creative products derived from it, so that not only professional designers, but also the general public can appreciate the deep cultural charm of art of decorating Chinese ceramic products. It is an important direction for this paper to take artificial intelligence technology as a new means to inherit and develop the style and cultural characteristics of ceramic painting decoration. Moreover, with the increasing global exchanges, people’s ideological understanding has changed considerably, which has also contributed to the transformation of ceramic styles and aesthetics. Ceramics are gradually evolving from traditional daily necessities to works of art with aesthetic decorative functions, which is not only a change in the appearance and shape of ceramics, but also a way of artistic presentation of ceramics. Traditional ceramic art design and modern advanced technology promote and complement each other. On the one hand, traditional ceramic art has the characteristics of “sensibility” and “art”, ensuring the aesthetic properties of products. On the other hand, modern technology brings advances in thought and productivity, which can fundamentally change the characteristics of ceramic art design, efficiently expand the applicability of traditional ceramic art design, and thus discover more application scenarios for ceramic cultural and creative products.

With the rapid development of computer science and artificial intelligence technology (Janai et al., 2020; Xu et al., 2021; Ye & Su, 2021), style transfer and generative adversarial networks are increasingly applied in various fields (Kadish et al., 2021; Wang et al., 2021; An et al., 2021). Gatys et al. (2015) first proposed the concept of style transfer of images, which uses CNN neural networks to extract the style features of the input picture, and then separates the input content from the reference image features, and uses the loss function as the target to get the reconstructed style image. However, this method is time-consuming in training, has high requirements on equipment, and is limited in application scenario. Risser et al., (2017) proposed using histogram loss method to improve the problem that content and style cannot be reasonably segmented by studying the previous methods. Subsequently (Fei et al., 2020; Li et al., 2017), respectively improved the neural network algorithm, which made the image generated by the algorithm more reasonable and gradually in line with the aesthetic taste of the public, and made it possible for artificial intelligence to replace designers to do complex and repetitive work, so that designers could better focus on innovation.

Currently, although style transfer algorithms have been widely applied in the field of art and design, there are still issues such as poor transfer effects, structural changes in the generated images, and limited diversity in the transferred styles. Therefore, this study aims to explore an improved style transfer algorithm through qualitative research methods, focusing on the decorative art of Chinese ceramic artworks. The dataset used in this research primarily relies on a database of Chinese ceramic artworks, which provides a rich collection of samples with different styles and characteristics. These samples serve as the source images and target style images for the style transfer algorithm, providing the necessary data foundation for the study.

In the research process, the watershed algorithm is first employed to segment the input and output images, ensuring that the image structure remains unchanged during the style transfer process. Furthermore, local affine transformations are introduced to preserve the points, lines, and planes of the images throughout the transfer. This improved style transfer algorithm is then applied to the design of ceramic cultural and creative products, generating design renderings. The results of the study demonstrate that the improved style transfer algorithm enhances the clarity of ceramic painting decorations, facilitating a comprehensive presentation of ceramic art. Compared to traditional algorithms, the improved algorithm provides better transfer effects. Traditional algorithms may suffer from image distortion or loss of details after the transfer, while the improved algorithm, by leveraging the segmentation of the watershed algorithm and the application of local affine transformations, better preserves the structure and details of the original images, resulting in more accurate and natural style transfer effects. Additionally, traditional algorithms often enable the transfer of a single style, limiting the diversity and creativity in artistic design. In contrast, the improved algorithm, through the application of local affine transformations, allows for a greater variety of style transfers, offering more possibilities for the comprehensive presentation of ceramic painting decorations.

In summary, based on qualitative research methods and focusing on the decorative art of Chinese ceramic products, this study improves the effectiveness of ceramic painting decorations through an enhanced style transfer algorithm. The research findings highlight the advancements of the proposed algorithm over traditional approaches. This study provides a new approach to the protection and inheritance of ceramic painting decorations, making ceramic art more accessible to emerging designers and the general public for participation in the development of artistic works.

3. Material and methods

This study incorporates a combination of qualitative and experimental research methods to comprehensively explore the art of decorating Chinese ceramic products. Firstly, qualitative research methods are employed to observe and describe this art form, facilitating a deep understanding of its characteristics and styles. The dataset utilized in this study encompasses a diverse range of ceramic painting decoration patterns obtained from the Chinese ceramic artwork database, ensuring the inclusion of different styles and features.

Subsequently, the collected data is analyzed using descriptive analysis methods. This involves segmenting the input and output images, and implementing the watershed algorithm to preserve the stability of the image structure. Local affine transformations are applied to maintain the invariance of points, lines, and planes during the style transfer process. Through these measures, the improved style transfer algorithm is effectively evaluated and optimized, leading to enhanced transfer effects and a broader array of style transfers.

Furthermore, experimental research methods are employed for comparison and summarization purposes. The improved style transfer algorithm is compared with traditional algorithms to assess differences in transfer effects, structural changes, and the diversity of style transfers. By carefully analyzing and summarizing the experimental results, valuable conclusions can be drawn regarding the performance of the improved algorithm. These findings provide guidance that is crucial for the development and preservation of the art of decorating Chinese ceramic products, ensuring its continued growth and inheritance.

3.1 Data collection and preparation

Artificial intelligence can help art designers get rid of tedious steps in some aspects, save designers’ time, and improve creation efficiency (Jong et al., 2021; Zhang & Lu, 2021). The idea of using artificial intelligence for art creation is shown in Figure 2. In terms of the decorative art design of ceramic painting, Firstly, the ceramic pattern art style is collected from the database of Chinese ceramic artworks, including: blue and white, famille rose, ancient color, underglaze red and many other categories. And then it is necessary to make a summary of the patterns, color characteristics and semantics presented through the style characteristics. By collecting a large number of pictures related to ceramic patterns, different categories are classified. After extracting the image features, according to the data characters and pattern analysis, a variety of pictures that can represent the artistic characteristics of ceramic patterns are selected from the texture, color characteristics and semantics for style reference.

Figure 2. Design idea.

3.2. The steps of improved style transfer algorithm

In terms of algorithm, this paper first uses the L-BFGS-B (Limited-memory Broyden—Fletcher–Goldfarb—Shanno) algorithm (Dalvand & Hajarian, 2020) to solve the optimal solution problem in transfer learning. L-BFGS-B is similar to the gradient descent algorithm, but it converges faster in most cases. It has the characteristics of Newton’s method, but does not need to store the Hesse matrix during operation like Newton’s method, so it can save more computing resources. The steps of the L-BFGS-B algorithm are shown in Table 1:

Table 1. L-BFGS-B algorithm steps

L-BFGS-B algorithm steps:
Step 1	Select the initial point X0, when$\epsilon >0$ ,store the data of the last m iterations
Step 2	$\mathrm{k}=0,H_0=I,r=\mathrm{\nabla }f\left(x_0\right)$
Step 3	If $∥\mathrm{\nabla }f\left(x_{k+1}\right)∥\le \epsilon$ ,then return the optimal solution x, otherwise go to step 4
Step 4	Calculate the feasible direction for this iteration ${\mathrm{p}}_k=-{\mathrm{r}}_k$
Step 5	Calculate the step ak>0, and perform a one-dimensional search on the following equation $\mathrm{f}\left(x_k+{\alpha }_{\mathrm{k}}{p}_k\right)=min\mathrm{f}\left(x_k+\alpha p_k\right)$
Step 6	Update the weight x $x_{k+1}=x_k+{\alpha }_{\mathrm{k}}{p}_k$
Step 7	If k>m ,then keep the last m vector pairs and delete $\left({\mathrm{S}}_{k-m},{\mathrm{t}}_{k-m}\right)$
Step 8	Calculate and hold ${\mathrm{s}}_{\mathrm{k}}=x_{k+1}-x_k$ ${\mathrm{t}}_{\mathrm{k}}=\mathrm{\nabla }f\left(x_{k+1}\right)-\mathrm{\nabla }f\left(x_k\right)$
Step 9	Use two loop recursion algorithms to solve ${\mathrm{r}}_{\mathrm{k}}={\mathrm{B}}_{\mathrm{k}}\mathrm{\nabla }f\left(x_k\right)$
Step 10	Calculate k=k+1 ,and go to step 3

When the transfer learning algorithm runs, it first takes two images: one is the input image and the other is the reference style image. In order to transfer the reference image style to the input image and ensure that the output image is not distorted, this paper introduces the following two methods:

① During the optimization process, affine transformation (Luan et al., 2017) is added to the objective function to constrain the local overflow of image colors after style transfer to prevent image distortion.

② The watershed algorithm is introduced, which can segment the image of the input picture to avoid the problem of content mismatch and can significantly improve the authenticity of the migrated picture.

Transfer learning first transfers the style of the reference picture S to input picture I, and then uses the minimized objective function to generate the output image O, and the algorithm formulas are as follows:

(1) ${\mathcal{L}}_{\mathrm{t}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l}}=\sum _{\mathrm{\ell }=1}^L{\alpha }_{\mathrm{\ell }}{\mathcal{L}}_c^{\mathrm{\ell }}+\mathrm{\Gamma }\sum _{\mathrm{\ell }=1}^L{\beta }_{\mathrm{\ell }}{\mathcal{L}}_s^{\mathrm{\ell }}$(1)

(2) ${\mathcal{L}}_c^{\mathrm{\ell }}\hspace{0.17em}=\frac{1}{2N_{\mathrm{\ell }}{D}_{\mathrm{\ell }}}\sum _{ij}{\left(F_{\mathrm{\ell }}[O]-F_{\mathrm{\ell }}[I]\right)}_{ij}^2$(2)

(3) ${\mathcal{L}}_s^{\mathrm{\ell }}\hspace{0.17em}=\frac{1}{2N_{\mathrm{\ell }}^2}\sum _{ij}{\left(G_{\mathrm{\ell }}[O]-G_{\mathrm{\ell }}[S]\right)}_{ij}^2$(3)

${\mathcal{L}}_{\mathrm{t}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l}}^{\hspace{0.17em}}\hspace{0.17em}$represents the total loss function, ${\mathcal{L}}_c^{\mathrm{\ell }}\hspace{0.17em}$ represents the content loss function, ${\mathcal{L}}_s^{\mathrm{\ell }}\hspace{0.17em}$ represents the style loss function; L represents the total number of convolutional layers; $\mathrm{\ell }$ represents the $\mathrm{\ell }$convolutional layer of the deep convolutional network; There are $N_{\mathrm{\ell }}$ filters in each convolutional layer, and the size of the eigenvector graph of each filter is $D_{\mathrm{\ell }}$; $F_{\mathrm{\ell }}[\cdot ]\in {\mathbb{R}}^{N_{\mathrm{\ell }}\times D_{\mathrm{\ell }}}$ represents the eigenvector; $G_{\mathrm{\ell }}[\cdot ]=F_{\mathrm{\ell }}[\cdot ]F_{\mathrm{\ell }}[\cdot {]}^T\in {\mathbb{R}}^{N_{\mathrm{\ell }}\times N_{\mathrm{\ell }}}$is the Gram matrix, which represents the inner product between different vector feature mapping functions. ${\alpha }_{\mathrm{\ell }}$ and ${\beta }_{\mathrm{\ell }}$ are the weights of the configuration layer preferences; $\mathrm{\Gamma }$ represents the weight between the styles (2) and (3) .

In order to ensure that the original structure of the input image is not destroyed in the process of style transfer, the algorithm chosen in this paper directly processes the transformation process of the input image. By adding affine transformations in (2), it is ensured that the image is not distorted during the transmission process. Equation (4)(4) ${\mathcal{L}}_m=\sum _{c=1}^3{V}_c[O{]}^T{\mathcal{M}}_I{V}_c[O]$(4) represents the true regularization of the image, which can be minimized by a standard linear system. The system is expressed by a matrix ${\mathcal{M}}_I$ consisting of a matrix of order N×N, which relies only on the input image I and N is one pixel of image I. $V_c[O]$ represents the vectorization function (N ×1) of the image O in channel c.

(4) ${\mathcal{L}}_m=\sum _{c=1}^3{V}_c[O{]}^T{\mathcal{M}}_I{V}_c[O]$(4)

Since the Gram matrix in Equation (3)(3) ${\mathcal{L}}_s^{\mathrm{\ell }}\hspace{0.17em}=\frac{1}{2N_{\mathrm{\ell }}^2}\sum _{ij}{\left(G_{\mathrm{\ell }}[O]-G_{\mathrm{\ell }}[S]\right)}_{ij}^2$(3) implicitly encodes the exact value of the neural network output, it determines the ability of the image style transfer composition vector to be isometric (Weisstein, 2012) variations, and may distort the picture. In this paper, algorithm similar to Neural Doodle (Bae et al., 2006) and watershed (Levner & Zhang, 2007) image segmentation method is used to solve the problem.

Deep learning image segmentation algorithm needs to train a large amount of data for image recognition and then segmentation, and can only identify trained targets, such as people, cars, animals, etc. The main research subject of this paper is focused on art patterns. If deep learning is to be used for image segmentation, a lot of training is needed, which is time-consuming and may not include all the design patterns., which is more time-consuming and may not include all design patterns. Therefore, in this paper, the watershed algorithm is selected for image segmentation.

The watershed algorithm firstly determines the boundary positions of different object pixels in the image, and then performs image segmentation. Generally, the gray value inside each object is usually close, while the gray value pixels at the boundary of different objects are quite different. Therefore, the watershed algorithm first calculates the gradient map of the image, where the small gradient value is inside the object, and the large gradient value is the junction of different objects, and then finds the location of pixel with large gradient value, so as to achieve the purpose of image segmentation.

Next, the migration algorithm generates an image segmentation mask together with the processed input image and the style reference image, and uses the image segmentation mask as an additional channel of the neural network to add it to the input image to enhance the accuracy of the neural network’s style transfer. ${\mathcal{L}}_{s+}^{\mathrm{\ell }}$is the style enhancement loss function, as shown in Equation 5:

(5) ${\mathcal{L}}_{s+}^{\mathrm{\ell }}=\sum _{D=1}^D\frac{1}{2N_{\mathrm{\ell },D}^2}\sum _{ij}{\left(G_{\mathrm{\ell },D}[O]-G_{\mathrm{\ell },D}[S]\right)}_{ij}^2$(5)

(6) $F_{\mathrm{\ell },D}[O]=F_{\mathrm{\ell }}[O]M_{\mathrm{\ell },D}[I]F_{\mathrm{\ell },D}[S]=F_{\mathrm{\ell }}[S]M_{\mathrm{\ell },D}[S]$(6)

D is the number of channels in the image segmentation mask; $M_{\mathrm{\ell },D}[\bullet ]$ represents the D channel of layer 1 of the segmentation mask in the layers; $G_{\mathrm{\ell },D}[\bullet ]$ represents the Gram matrix corresponding to $F_{\mathrm{\ell },D}[\bullet ]$. In this study, down-sampling masks are used to match the feature space size of each layer of the CNN network.

Finally, ${\mathcal{L}}_c^{\mathrm{\ell }}$、${\mathcal{L}}_s^{\mathrm{\ell }}$ and ${\mathcal{L}}_m^{\hspace{0.17em}}$ are combined to form an enhanced style transfer algorithm. The formula is shown in (7):

(7) ${\mathcal{L}}_{\mathrm{t}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l}}=\sum _{l=1}^L{\alpha }_{\mathrm{\ell }}{\mathcal{L}}_c^{\mathrm{\ell }}+\mathrm{\Gamma }\sum _{\mathrm{\ell }=1}^L{\beta }_{\mathrm{\ell }}{\mathcal{L}}_{s+}^{\mathrm{\ell }}+\lambda {\mathcal{L}}_m$(7)

4. Experimental results and analysis

The experiment is run on a 64-bit Windows 10 operating system and Python 3.6 is chosen as the software development language. The development environment for data processing is jupyterlab and sklean library is used for data preprocessing, and TensorFlow is adopted as the machine learning development framework. The input image, reference image, and training output image of the algorithm are shown in Figure 3.

Figure 3. Training image of the transfer algorithm.

As shown in Figure 3, with the increase of the number of iterations, the reference image style of the output image becomes stronger and stronger on the premise of retaining the original image structure. When the iteration reaches 4000 times, the image effect is the best, which not only retains the input structure but also has the reference image style.

Figure 4. Training loss value.

In this paper, Content loss, TV loss, Affine Loss, and Total loss are used to evaluate and analyze image training loss in more detail. Content Loss function is used to calculate the difference of pixel value and evaluate the change of pixel before and after training. TV Loss (Total Variation Loss) can reduce image blur, reduce the difference of adjacent pixels in the image by reducing TV Loss, so as to maintain the smoothness of the image. After training to 2000 times, affine transformation is added to perform linear transformation and translation transformation of the output image to enhance the authenticity of the image. Therefore, it can be seen that Content loss, TV loss and Total loss in Figure 3 have obvious changes at 2000 times. In addition, it can be seen from Figure 3 that the clarity of the image has been significantly improved after 2000 training sessions. Finally, the total loss is obtained by multiplying each loss value by the weight and summing. The loss values are shown in Figure 4.

Figure 5. Style transfer process.

In this paper, the semantic information of the input image and the reference image is extracted by the watershed image segmentation method, and the reference image style is introduced by the improved transfer learning algorithm on the basis of retaining the structure of the input image, which can effectively constrain the generation effect and better control the content of the generated image. As shown in Figure 5, the image generated by the improved algorithm in this study has clear structure and texture, and the overall style is very close to the reference image, meeting the design requirements.

In order to visually analyze the advantages and disadvantages of various algorithms in ceramic art creation, this paper selects various types of input images and reference images for style transfer and compares them with traditional algorithms, as shown in Figure 6.

Figure 6. Comparison between the proposed algorithm and the traditional algorithm.

As shown in Figure 6, the content information of the image generated by the traditional migration algorithm is distorted, and the shape and line of the original objects are changed, which cannot be well adapted to the actual needs. In contrast, the algorithm designed in this paper incorporates the style of the reference image while retaining the appearance of the original image shape, which can not only transfer and create the decorative pattern of ceramic painting, but also develop the appearance of ceramic shapes twice.

5. The application

5.1. Theme selection of ceramic cultural and creative products

Chinese ceramics have a long history, and the decorative patterns of ceramic painting are varied, incorporating social style, culture, religion and other artistic connotations. The inner beauty and formal beauty shown by the decorative patterns have a profound impact on modern ceramic cultural and creative design. From the perspective of visual communication, the words or patterns contained in the design of artistic cultural and creative products should have a positive impact on the public’s thoughts, and convey the designer’s thoughts to the public through vision, so as to spread excellent Chinese traditional culture.

Different decorative patterns of painting represent a wide range of meanings. In ancient times, woven patterns were commonly used, and later evolved into geometric patterns, which can present artistic beauty through the simplest patterns (Ren, 2022). With the passage of time, landscape and plant patterns appeared. In the design of household ceramic cultural and creative products, plant patterns are widely used as an elegant decorative pattern and peony is a common plant pattern in ceramics, representing wealth and auspiciousness. In this paper, peony pattern of underglaze red style generated by improved transfer learning is combined with ceramic cultural and creative products. As shown in Figure 7, the peony pattern of underglaze red style generated by the algorithm shows the characteristics of fresh and elegant, which allows users to feel the natural flavor of the work. The application of it in ceramic cultural and creative products can not only improve the product grade, but also meet the preferences of consumers.

Figure 7. The image after the style transfer.

5.2. Functional design and style integration

The application of plant theme elements in ceramic painting design needs to refer to the composition structure of traditional plant paintings. On the one hand, it is necessary to consider the coordination of the primary and secondary structure of the painting, the distinction between virtual and real, and the correspondence of simplicity and complexity. After determining the “shape”, the design theme is clarified, the composition is arranged around the main body, and the theme is set off through scenario interaction. On the other hand, it should also be noted that the composition of ceramic art is different from that of ordinary plant painting. Because ceramic artworks are mostly curved surfaces, it is necessary to compose the composition according to the characteristics of ceramic artworks, and it also needs to consider the influence of harmony, symmetry, balance on artistic works, and appropriately use ceramic artworks that correspond to the pictures for composition.

Figure 8. Design sample.

Figure 8 is a series of renderings designed in this paper, which simulates the product effect through software modeling, rendering and other methods, in which the improved style transfer algorithm greatly shortens the design time of art of decorating Chinese ceramic products, and can also provide visual effects for the mass production of ceramic cultural and creative products. As shown in Figure 8, this series of ceramic design works have a reasonable subordinate relationship, which can optimize the aesthetics of ceramic cultural and creative products and enhance their taste and connotation. In addition, the pattern size is moderate, which brings aesthetic effect to the works through reasonable arrangement. The pattern is designed in the most prominent place of the ceramic work, and the sight moves with the work, allowing the audience to have a deep sense of artistic conception. In terms of color style, this series of ceramic design works are influenced by Confucianism and have a graceful temperament. The unique light color and thick ink style of plant painting can make ceramic cultural and creative products show a sense of euphemism and haze, and with appropriate white space, it leaves the audience with an artistic space for free imagination.

6. Conclusion

By harnessing style transfer algorithms to explore innovative approaches to traditional ceramic styles, we have introduced a multitude of ideas and possibilities for ceramic cultural and creative design. Such innovation serves to foster the diversified development of cultural and creative products, fulfilling the modern society’s desire for various forms of artistic expression.

Moreover, the utilization of style transfer algorithms in digital preservation and inheritance offers novel solutions for safeguarding and passing down the craftsmanship of ceramic painting decoration. This technology enables the preservation and inheritance of the aesthetic characteristics found in traditional Chinese ceramic art, seamlessly integrating them into modern cultural and creative art products. This transformation enhances the competitiveness of ceramic cultural and creative designers within the contemporary social landscape and addresses the market’s demand for traditional cultural and artistic features.

In conclusion, our research highlights the remarkable value and allure of ceramic painting decoration art in today’s society through the application of style transfer algorithms. This approach not only fosters the diverse development of cultural and creative products but also facilitates the innovation and preservation of traditional ceramic styles. It generates fresh and imaginative ideas for ceramic cultural and creative design, empowering designers with enhanced market competitiveness, and meeting the public’s expectations for the preservation of traditional culture and the advancement of ceramic art forms. These research findings hold great significance in promoting the development and inheritance of ceramic painting decoration art and elevating the cultural and creative industries to new heights.

7. Discussion

By researching relevant articles, we can see that there have been some related works in the field of style transfer applications. Chen et al. (Cheng & Xu, 2020) applied style transfer algorithms to lacquer art cultural creative products, providing a solution for the digital preservation and inheritance of lacquer art culture. However, the generated pattern effects were unsatisfactory and required further modifications. Hou et al. (Hou et al., 2020) used transfer neural networks to transfer traditional ethnic pattern designs, resulting in good-quality new style ethnic patterns. However, this method required pre-designing the transfer content framework, which was time-consuming. Ren et al. (Ren et al., 2022) used transfer learning techniques to transfer blue-and-white porcelain patterns to landscape paintings, achieving visually pleasing experiences. However, this algorithm could only learn a single style, making it relatively limited.

In comparison to these related works, our research has the following differences and unique aspects in the protection and inheritance of ceramic painting decoration techniques. Firstly, we propose an improved style transfer algorithm that optimizes the preservation and expression of ceramic painting decoration patterns. This algorithm maintains the image structure unchanged and performs local affine transformations during the transfer process, ensuring that the points, lines, and surfaces of the image remain unchanged. This approach can better preserve the expression of ceramic painting decoration patterns and further enhance the quality of generated patterns. Our research focuses on combining the intelligent development and promotion of ceramic painting decoration techniques. Through our algorithm design, we can not only provide more opportunities for novice designers and the general public to participate in the development of artistic works but also make the techniques of ceramic painting decoration more accessible for inheritance and protection. Our research aims to inject new vitality into traditional ceramic painting decoration by integrating artificial intelligence technology with ceramic art, thereby promoting its development and innovation in modern society.

Ceramic art is a treasure of traditional culture, carrying rich history and cultural connotations. However, due to factors such as the labor-intensive production process and high skill requirements, the ceramic painting decoration technique faces challenges in protection and inheritance. Our research aims to find a new path for the protection and inheritance of ceramic painting decoration techniques by combining artificial intelligence technology. In our research, we will delve into the characteristics and styles of ceramic painting decoration to better understand its uniqueness. By adopting an improved style transfer algorithm, we can maintain the image structure while performing local affine transformations during the transfer process, thereby better preserving the expression of ceramic painting decoration patterns. We will also explore how to combine artificial intelligence technology with ceramic art to enhance the level of intelligent development in ceramic painting decoration. Additionally, we will strengthen collaboration with experts and practitioners in the field of ceramic art, drawing from their experience and insights. Their understanding of ceramic painting decoration will provide valuable guidance for our research and ensure that our algorithm can better meet practical application needs. We will also actively engage in technology promotion and training activities to introduce our research findings to more designers and art enthusiasts, promoting the popularization and inheritance of ceramic painting decoration techniques.

In summary, our research has unique advantages in the protection and inheritance of ceramic painting decoration techniques, and has the potential to promote the intelligent development of ceramic painting decoration design. By combining artificial intelligence technology with ceramic art, our research is expected to have a positive impact on the development and inheritance of ceramic painting decoration techniques. We believe that through our efforts, ceramic painting decoration techniques will continue to evolve, bringing more artistic charm and cultural value to people.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Notes on contributors

Yongtao Zhao

Yongtao Zhao has been engaged in cultural creative product development for many years, with research focusing on cultural feature analysis and creative product development of cultural creative products, research and achievement transformation of the cultural creative product industry chain construction, new ideas, new technologies, and new materials for the production of cultural creative products. Based on the designed improved style transfer algorithm in this study, features can be extracted from images in the field of art. The feature information of the style image is transferred and transformed onto the target content image, resulting in a synthesized image that preserves the shape and structural information of the target content image while incorporating the color and texture information of the style image.

Mr. Yongtao Zhao, male, 44 years old, is born in Luoyang, Henan Province, China.