Available online at www.HighTechJournal.org  

HighTech and Innovation  
Journal 

Vol. 4, No. 4, December, 2023 

 

 

779 

 

ISSN: 2723-9535 

 

Color Analysis of Cloud Brocade Pattern by Image Style Transfer 

Chuanqi Wu 1* 

1 Changzhou Vocational Institute of Textile and Garment, Changzhou, Jiangsu 213000, China. 

Received 12 September 2023; Revised 16 November 2023; Accepted 23 November 2023; Published 01 December 2023 

Abstract 

With the continuous improvement of the level of science and technology, the design method of cloud brocade pattern has 

gradually changed from the traditional process of color halo, white, and gold stranding to the modern design process, such 

as the synthesis of cloud brocade line pattern based on the transfer of image style. But back to reality, this method still has 

problems such as blurred outline and mixed colors, which is not conducive to the transfer of cloud brocade style pictures. 

Based on this, the paper will use the cloud brocade pattern style transfer color optimization model to analyze the color of 

the cloud brocade pattern in order to get a better cloud brocade style effect map. The results show that the average similarity 

of the local migration algorithm is 0.348, while the average similarity of the local migration algorithm based on color 

optimization is 0.378, which is 8.62% higher than that of the local migration algorithm. After 1600 iterations, the average 

running time of the local migration algorithm is 13.65s, and the running time of the local migration algorithm for color 

optimization is 12.46s. It can be seen that the local migration algorithm based on color optimization has obvious advantages 

in both comprehensive similarity and running time and can provide new ideas and references for the current design of 

Yunjin pictures. 

Keywords: Style Transfer; Cloud Brocade Pattern; Color Optimization. 

1. Introduction 

As one of the three famous scenic spots in China, Yunjin has not only gained the favor of the general public due to 

its exquisite patterns, rich color matching, and fine weaving, but it also means that the traditional silk manufacturing 

process in China has reached a high level. For example, domestic scholars such as Liu et al. (2023) explored the 

inheritance and promotion path of Nanjing Yunjin based on the Internet and found that Nanjing Yunjin seized the 

opportunities of historical development and combined its own advantages. Integrating into the development of the new 

era has given rise to new vitality and provided many reference experiences for the development of intangible cultural 

heritage [1, 2]. 

In order to present the effect of different colors per flower, there are usually dozens of colors in the cloud brocade 

pattern. Traditional weaving craftsmen usually use techniques such as color halo formulas, alternating white and large, 

and gold twisted edges to match the pattern and complete the pattern design. For example, domestic scholar Pan (2023) 

took Nanjing cloud brocade as an example to explore and practice the cultivation of its intangible cultural heritage 

inheritors. We have gradually built a comprehensive education system for intangible cultural heritage inheritance based 

on "intangible cultural heritage inheritance+information technology+core literacy", integrating master inheritance and 

information technology, strengthening students' core literacy, practical skills, and innovation and entrepreneurship 

abilities, and achieving some results [3, 4]. 

                                                           
* Corresponding author: wu38028@163.com 

 
http://dx.doi.org/10.28991/HIJ-2023-04-04-07 

 This is an open access article under the CC-BY license (https://creativecommons.org/licenses/by/4.0/). 

© Authors retain all copyrights. 

https://creativecommons.org/licenses/by/4.0/


HighTech and Innovation Journal         Vol. 4, No. 4, December, 2023 

780 

 

The traditional design method of cloud brocade pattern takes a long time, has low efficiency, and is affected by the 

skill level of the process inheritor, which greatly limits the innovative design of cloud brocade pattern and the activation 

and inheritance of products. With the development of image technology and artificial intelligence, image style transfer 

has been gradually applied to the fields of arts and crafts such as porcelain, lacquer, and painting [5, 6]. 

As a pattern design feature of Yunjin, it is also very suitable for using the method of image style transfer to complete 

the design. Style transfer refers to the process of re-rendering an image with the texture, color, and other styles of another 

image while keeping the content of the image unchanged. The current mainstream image style transfer method is based 

on a convolutional neural network algorithm, which extracts the image content and style from the content map and style 

map with the help of a training network and obtains the effect map through image reconstruction [7, 8]. 

The Yun brocade pattern is rich in color, and the effect drawing generated by the original style transfer algorithm is 

prone to problems such as color mixing. The outline of the target pattern and the content are unclear, so the design effect 

is poor. In order to inherit and innovate the design and application of cloud brocade based on the characteristics of cloud 

brocade patterns and the original transfer model, this paper proposes a local style transfer method based on color 

optimization. First, the mask of the target pattern is obtained to distinguish the pattern from the background, and the 

outline of the pattern is clear. Secondly, the variance sum of the composite image pixels in the three color channels is 

taken as the color loss in the total loss, and the color difference within the pattern is reduced by optimizing the color 

loss, and the semantic clarity of the pattern is improved. By combining color loss and mask mapping, the effect map of 

Yunjin style transfer with a clear outline and easy semantic recognition is obtained. 

2. Principle of Image Style Transfer 

The Visual Geometry Group (VGG19) model based on a convolutional neural network has 16 hidden layers (divided 

into 5 stages) and 3 fully connected layers, which have powerful image features and semantic expression ability [9, 10]. 

The original style transfer algorithm uses the VGG19 model to extract the underlying texture and high-level semantic 

information of the image as style and content, respectively, uses the optimization function to minimize the loss, and then 

iteratively updates the composite image to obtain the stylized effect diagram. The specific process is shown in Figure 1. 

Style chart

Initializes the 

Gaussian white 

noise map

Content map

VGG19 was used to extract 

features and calculate style 

losses

Minimize the loss 

function by iteration
Style transfer effect

 

Figure 1. Process of image style transfer 

Since the high-level convolutional network focuses on preserving the contour, semantic content, and other 

information of the image, the feature matrix of the content graph C and the composite graph G in the fourth layer of the 

convolutional neural network is selected, and the square error of the two feature matrices is the content loss, as shown 

in Equation 1. 

𝐿𝑐𝑜𝑛𝑡𝑒𝑡 =
1

2
∑ ‖𝑎[𝑙](𝐶) − 𝑎[𝑙](𝐺)‖

2

𝑙∈{𝑙𝑐𝑜𝑛𝑡𝑒𝑛𝑡}
 (1) 

Where, 𝑎[𝑙] is the activation value matrix obtained by convolution of image I at the first layer [11], and its size is 𝑛𝐶
[𝑙]

×

𝑛𝐻
[𝑙]

× 𝑛𝑊
[𝑙]

. In terms of style, Gram matrix can describe the correlation between features by calculating the eccentric 

covariance matrix between features, such as the intensity of a feature, positive correlation or negative correlation between 

features, and so on, so as to obtain the style features of images. Transform 𝑎[𝐼](𝐼) into a matrix of 𝑛𝐶
[𝐼]

× 𝑛𝐻
[𝐼]

× 𝑛𝑊
[𝐼]

, 

denoted as𝑎[𝐼](𝐼)`, then the Gram matrix I of the image G(I) is shown in Equation 2. 

G(I) =
[𝑎[𝐼](𝐼)`][𝑎[𝐼](𝐼)`]

T

nC
[I]

×nH
[I]

×nW
[I]   (2) 

The Gram matrix of style graph S and composite graph G are calculated separately, and their square variances are 

denoted as style losses. Since the features extracted by the deep and shallow layer networks are different, in order to 



HighTech and Innovation Journal         Vol. 4, No. 4, December, 2023 

781 

 

comprehensively summarize the image style features, the style loss of all the subsampled layers is calculated and 

recorded as the style loss, as shown in Equation 3. 

𝐿𝑠𝑡𝑦𝑙𝑒 = ∑
1

2
‖𝐺(𝑆) − 𝐺(𝐺)‖2

𝑙∈{𝑙𝑠𝑡𝑦𝑙𝑒}
 (3) 

In order to ensure the controllability of the composite image, content and style loss weights α and β are set 

respectively to adjust the style transfer effect, so the total loss is shown in Equation 4. 

𝐿𝑡𝑜𝑡𝑎𝑙 = 𝛼𝐿𝑐𝑜𝑛𝑡𝑒𝑛𝑡 + 𝛽𝐿𝑠𝑡𝑦𝑙𝑒 (4) 

3. Yun Brocade Pattern Style Transfer Color Optimization Model 

In order to improve the problems such as color clutter, front and back scenes mixing, and unclear pattern outline in 

the migration effect diagram of Yun brocade lines, the paper added color loss and mask map on the basis of the original 

style transfer model and proposed a color optimization model of style transfer of Yun brocade patterns [12–14]. The 

structure of the model is shown in Figure 2. 

Roll how 
many 

pieces 1

Roll how 
many 

pieces 2

Roll how 
many 

pieces 3

Roll how 
many 

pieces 4

Roll how 
many 
pieces 

5

Style chart

Base map

Content map

copy

Content loss

Style loss

Mask pattern

Composite graph

Color loss

 

Figure 2. Color optimization model of Yunjin pattern style transfer 

According to the structure of the color optimization model of Yunjin pattern style transfer, it can be seen that, first, 

the mask map of the content map can be obtained as the input with the help of the quick selection tool, and the content 

map can be copied as the base map. The Yunjin style map, line draft content map and base map can be respectively input 

into the VGG19 network after the pre-training, and the feature matrix can be extracted through convolution. Secondly, 

the content loss is calculated by the feature matrix of the content map and the base map. Calculate the color loss by using 

the base color map pixels; The synthesized base map is divided into three color channels R, G and B, and all pixels of 

the base map are traversed in each channel, denoted as 𝑅𝑜𝑖𝑝𝑛. Each layer synthesizes the pixel variance andcolor loss in 

the three-channel base map. Since color belongs to the underlying texture information, the composite base map with the 

same number of layers is used for color loss and style loss (a total of 5 layers), and the specific calculation formula is 

shown in Equation 5. 

𝐿𝑐𝑜𝑙𝑜𝑟 = ∑ ∑
∑ (𝑅𝑜𝑖𝑝𝑛

[𝑐]
−𝑅𝑜𝑖𝑝𝑛

[𝑐]̅̅ ̅̅ ̅̅ ̅̅ ̅
)

2

𝑛

𝑛𝑐∈{𝑅、𝐺、𝐵}𝑙∈{𝑙∈{𝑙𝑠𝑡𝑦𝑙𝑒}}
  (5) 

The total loss is the weighted sum of content loss, style loss, and color loss, as shown in Equation 6. 

𝐿𝑡𝑜𝑡𝑎𝑙 = 𝛼𝐿𝑐𝑜𝑛𝑡𝑒𝑛𝑡 + 𝛽𝐿𝑠𝑡𝑦𝑙𝑒 + 𝛾𝐿𝑐𝑜𝑙𝑜𝑟 (6) 

In Formula 6, 𝛼, 𝛽 and 𝛾 are the weights corresponding to the losses respectively, which can be adjusted according 

to the style needs. Finally, the Adam optimizer is selected to optimize the total loss, feedback and update the pixels of 

the base image to ensure that the iteratively generated base image is closer to the style of Yunjin image while retaining 

the original content, and the color difference of the composite image is minimized. Then combined with the mask map, 

output the cloud brocade style effect map with relatively clear contour and pattern semantics. 

4. Experiment on Color Optimization of Yunjin Pattern by Image Style Transfer 

Before carrying out the experiment of color optimization of cloud brocade pattern, it is necessary to make full 

preparation. Specifically, the Style migration optimization experiment was conducted using the Pytorch framework on 



HighTech and Innovation Journal         Vol. 4, No. 4, December, 2023 

782 

 

a desktop with an Intel(R)core(TM)i7-117002.5GHz processor, NVIDIA GeForce RTX 3080Ti graphics card, and 32GB 

of RAM[15]. The Yunjin sample is used as the style diagram, and the line pattern is used as the content diagram. Since 

the migration model does not limit the image size, the experiment sets the size of the style map and the content map to 

224*224 pixels, 300*300 pixels, 400*400 pixels, 512*512 pixels for comparison, and achieves the migration effect as 

shown in Figure 3. 

224*224 300*300 400*400 512*512
 

Figure 3. Effect of pixel migration with different specifications 

When the size is set to 224*224 pixels, the obvious migration effect can be obtained directly, and the consumption 

time is shortest. Considering the quality and processing time of the effect image, the size of the style image, content 

image and mask image is set to 224*224 pixels, and the size setting can be adjusted according to the size and clarity of 

the original image. Adam was selected as the model optimizer, and the learning rate was set to 5*10-3. The experimental 

results show that the absolute values of content loss, style loss and color loss are quite different. In order to balance the 

various losses, the three losses were balanced by using the coefficient, and the loss values were 1*10-3, 5*105 and 

1*102, respectively, and the loss values were in the similar scale range, and the migration effect was good. In order to 

determine the number of iterations, the experiment selected three Yunjin style diagrams and line drawings as objects, set 

the initial iteration number of the optimization model to 2000, and output the loss value of content, style and color once 

every 100 iterations. The loss trend drawn is shown in Figure 4. 

 

Figure 4. Content, style, color loss trend 

It can be seen that the content loss hardly changes, and the style loss gradually decreases to a stable level with the 

increase of the number of iterations. As the style loss decreases, the style of the base map is closer and closer to the style 

of the style map, and the color is more complex, so the color loss presents a small value and then decreases to a stable 

trend. Combining the loss trend and training time, the experiment selects the synthesis graph of 1600 iterations to transfer 

the output graph. 



HighTech and Innovation Journal         Vol. 4, No. 4, December, 2023 

783 

 

4.1. Determination of Style Transfer Output Diagram 

Based on the above basic Settings, the color optimization local migration algorithm proposed in this paper is 

compared with the original migration algorithm and local migration algorithm, and the resulting style migration output 

diagram is shown in Figure 5. 

Yunjin style diagram Content map Original model Local migration Color optimization local migration
 

Figure 5. Style migration output diagram 

It can be seen that thecolor of the composite image directly generated by the original migration model is mixed, and 

it is difficult to distinguish the shape and content of the main pattern. This is because the convolutional neural network 

learns the color features of the image, including the brightness, saturation and distribution of the color, when extracting 

the features of the style map and the content map, resulting in the corresponding transfer between the color pixels. The 

content map of line sketch only contains the contour color information, which can not form the corresponding transfer, 

resulting in the color mixing of the composite effect image and the semantic ambiguity of the target pattern. In the local 

migration results, the local migration model with mask map can improve the definition of contour by separating the 

pattern from the background. In the local transfer of color optimization, color loss is added to the optimization model 

based on the mask diagram. The color of lotus petals is mainly pink, and the color of stamens is mainly green. The 

outline of the bell is dark red, the blank part is light; The mixed color of the turtle body is reduced, and the main color is 

green. The optimized model not only improves the definition of the outline, but also reduces the color difference within 

the pattern and improves the semantic recognition of the pattern. 

4.2. Analysis of Experimental Results 

For the analysis of the experimental results, there are two stages here: firstly, the experimental results are analyzed 

from a subjective perspective, using a questionnaire survey method. 20 volunteers are recruited to generate style transfer 

output maps based on three algorithms: the original transfer algorithm [15], local transfer algorithm [16, 17], and color 



HighTech and Innovation Journal         Vol. 4, No. 4, December, 2023 

784 

 

optimization local transfer algorithm. The clarity of the semantic content of the pattern is evaluated, with a score set at 

1-5 points, The easier the pattern content is to recognize, the higher the score. Calculate the mean of the output graphs 

of the three algorithms for 20 volunteers, and the results are shown in Figure 6. 

 

Figure 6. The output graphs of the original migration, local migration and color optimization local migration are used to 

calculate the survey scores 

It can be seen that the mean values of the original migration, local migration and color optimized local migration 

models are 2.17, 3.61 and 3.87, respectively. Therefore, the color-based local transfer model proposed in this paper 

improves the semantic content clarity of the main pattern through color optimization on the premise of realizing local 

style transfer. Second, the experimental results are analyzed on the objective surface, mainly to analyze the time 

consuming of the algorithm for the quality of the output image of style transfer. The method used here is structural 

similarity (SSIM), which is defined by pixel image as the combination of three different factors: brightness (𝑙), 

contrast (𝑐) and structure (𝑠)[18-20]. The mean is used as an estimate of brightness, standard deviation as an estimate 

of contrast, and covariance as a measure of structural similarity. Each time, the window with D on the picture is 

calculated, and the average value of all Windows is obtained as the structural similarity of the whole image. The 

structural similarity value ranges from 0 to 1, and the closer the value is to 1, the higher the similarity between the two 

images. The calculation formula is shown in Equation 7. 

SSIM(a, b) =
(2μaμb + c1)(2σab + c2)

(μa
2 + μb

2 + c1)(σa
2 + σb

2 + c2)
 (7) 

In the formula, the mean of sample a is represented by μa, and the mean of sample b is represented by μb; The 

variance of sample a is represented by 𝜎𝑎, and the variance of sample b is represented by 𝜎𝑏; The covariance of a and b 

is represented by σab, and c1 and c2 are two constants to avoid division by zero. In order to comprehensively analyze 

the migration effect, the paper calculates the SSIM value (𝐺𝑠, 𝐺𝑐) of the output result, style graph and content graph 

respectively, and takes the mean value of 𝐺𝑠 and 𝐺𝑐 as the comprehensive similarity index. The size of the composite 

effect map, style map and content map is set uniformly at 224*224 pixels, and the structural similarity of the migration 

output map is calculated using SSIM algorithm. The results obtained by local migration algorithm and color optimization 

local migration algorithm are shown in Table 1. 

Table 1. Structural similarity of output graph under local migration algorithm and color optimization local migration 

algorithm 

Picture 

number 

Local migration algorithm 
Color optimization local migration 

algorithm 

𝐆𝐬 𝐆𝐜 
𝟏

𝟐
(𝐆𝐬 + 𝐆𝐜) 𝐆𝐬 𝐆𝐜 

𝟏

𝟐
(𝐆𝐬 + 𝐆𝐜) 

A 0.05 0.45 0.25 0.06 0.65 0.35 

B 0.14 0.71 0.42 0.17 0.71 0.44 

C 0.07 0.53 0.30 0.08 0.58 0.33 

D 0.08 0.70 0.39 0.07 0.71 0.39 

E 0.06 0.71 0.38 0.06 0.70 0.38 

0

1

2

3

4

5

6

A B C D E

S
co

re

Rendering number

Primary migration Local migration Color optimization local migration



HighTech and Innovation Journal         Vol. 4, No. 4, December, 2023 

785 

 

It can be seen that there is little difference between the two migration models in terms of style similarity, but in terms 

of content similarity, the optimization algorithm is generally higher. The average similarity of the local migration 

algorithm is 0.348, while that of the local migration algorithm based on color optimization is 0.378, which is 8.62% 

higher than that of the local migration algorithm. In addition, the local migration algorithm and the color optimization 

local migration algorithm were respectively used on the RTX3080TI to perform 1600 iterations on the experimental 

images, and the required processing time is shown in Table 2. 

Table 2. Running time of the two algorithms after 1600 iterations 

Picture coding 
Local migration 

algorithm 

Color optimization local 

migration algorithm 

A 13.42 12.31 

B 13.56 12.56 

C 13.81 12.68 

D 13.61 12.45 

E 13.83 12.34 

Mean value 13.65 12.46 

It can be seen that the average running time of the local migration algorithm is 13.65s, and the average time of the 

local migration algorithm for color optimization is 12.46s when color processing is added under the same configuration. 

On the surface of the data results, the optimized migration algorithm can quickly obtain a stylized image that is more 

similar to the structure of the content map, which is conducive to realizing the migration of the Yunjin style of the line 

manuscript pattern. 

5. Conclusion 

This study mainly found that the average comprehensive similarity of the local transfer algorithm is 0.348, while the 

average of the local transfer algorithm based on color optimization is 0.378, which is 8.62% higher than the local transfer 

algorithm. After 1600 iterations, the average time required for the local migration algorithm to run is 13.65 seconds, 

while the color optimization local migration algorithm runs in 12.46 seconds. The time difference between the two is 

1.19 seconds. Although the time difference is not significant, the color optimization local migration algorithm still has 

some improvement, which can confirm the feasibility of the Yunjin pattern style migration color optimization algorithm. 

Compared with other studies, the article uses empirical comparison to test the advantages and disadvantages of color 

optimization local transfer algorithms, breaking the traditional framework of theoretical analysis. This is relatively avant-

garde in promoting the development of cloud brocade patterns, with certain advantages, but also has certain limitations. 

If the number of iterations is too high, it will inevitably cause some interference to the algorithm operation and require 

subsequent optimization and adjustment. Overall, this study provides new options for the selection and development of 

algorithms for transferring the style of Yun brocade patterns, which can further promote the development of Yun brocade 

patterns. This has a certain role and contribution to the inheritance and preservation of intangible cultural heritage, and 

it is recommended to apply it in practice in the future. In addition, in the future, algorithms will be continuously optimized 

to ensure the quality of image migration while reducing runtime. 

6. Declarations  

6.1. Data Availability Statement 

The data presented in this study are available in the article. 

6.2. Funding 

The author received no financial support for the research, authorship, and/or publication of this article. 

6.3. Institutional Review Board Statement 

Not applicable.  

6.4. Informed Consent Statement 

Not applicable.  

6.5. Declaration of Competing Interest 

The author declares that they have no known competing financial interests or personal relationships that could have 

appeared to influence the work reported in this paper. 



HighTech and Innovation Journal         Vol. 4, No. 4, December, 2023 

786 

 

7. References 

[1] Chen, Y., & Gan, Z. (2023). Research on Development Mode of Intangible Cultural Heritage in E-commerce Platform: Taking 

Nanjing Yunjin as an Example. Frontiers in Business, Economics and Management, 9(3), 233-237. 

[2] Liu, T., Shi, L., & Yang, D. (2023). Research on the path of inheritance and promotion of intangible cultural Heritage in the 

context of Internet: A case study of Nanjing Yunjin. Cultural Relics Identification and Appreciation, 2023, 161-164. 

[3] Pan, H. (2023). Exploration and Practice of Cultivating Non-Genetic Inheritors of "Nanjing Yunjin". China Training, 2023, 34-

36. 

[4] Wang, W., Zhang, Y., & Yang L. (2021). Analysis on Semiotics of Nanjing Yun Brocade and Design of Silk Scarf Pattern from 

the Perspective of Court Aesthetics. Silk, 58, 116-123. 

[5] Xu, X., Yang, Z., Yan, Y., Li, Z., & Zhan, G. (2023). Research on Style Transfer Art Word Based on Convolutional neural 

Network. Journal of Dalian Minzu University, 25, 69-72.  

[6] Cai, X., Li, Z., Xi, M., & Sun, H. (2023). Color Rendering and Style Transfer of Hand-Drawn Clothing Pattern Based on Neural 

Network. Journal of System Simulation, 35, 604-615.  

[7] Chen, D., Yuan, L., Liao, J., Yu, N., & Hua, G. (2020). Explicit filter bank learning for neural image style transfer and image 

processing. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(7), 2373-2387. doi:10.1109/TPAMI.2020.2964205. 

[8] Gatys, L. A., Ecker, A. S., & Bethge, M. (2016). Image style transfer using convolutional neural networks. In Proceedings of the 

IEEE conference on computer vision and pattern recognition, 2414-2423. doi:10.1109/CVPR.2016.265. 

[9] Nie, L., Lin, C., Liao, K., Liu, M., & Zhao, Y. (2020). A view-free image stitching network based on global homography. Journal 

of Visual Communication and Image Representation, 73, 102950. doi:10.1016/j.jvcir.2020.102950. 

[10] Han, B., Du, J., Jia, Y., & Zhu, H. (2021). Zero-watermarking algorithm for medical image based on VGG19 deep convolution 

neural network. Journal of Healthcare Engineering, 2021. doi:10.1155/2021/5551520. 

[11] Wang, Y., Li, Y., Song, Y., & Rong, X. (2020). The influence of the activation function in a convolution neural network model 

of facial expression recognition. Applied Sciences, 10(5), 1897. doi:10.3390/app10051897. 

[12] Xuelin, Q., Jue, H., Ying, S., & Zheng, L. (2021, November). Digital style design of Nanjing brocade based on deep learning. In 

2021 International Conference on Culture-oriented Science & Technology (ICCST), 339-342, Beijing, China. 

doi:10.1109/ICCST53801.2021.00077. 

[13] Ajmi, C., Zapata, J., Elferchichi, S., Zaafouri, A., & Laabidi, K. (2020). Deep learning technology for weld defects classification 

based on transfer learning and activation features. Advances in Materials Science and Engineering, 2020. 

doi:10.1155/2020/1574350. 

[14] Jin, M., Zheng, Y., Li, Y. F., Gong, C., Zhou, C., & Pan, S. (2021). Multi-scale contrastive SIAMESE networks for self-

supervised graph representation learning. arXiv preprint arXiv:2105.05682. doi:10.48550/arXiv.2105.05682. 

[15] Huang, Y. Z., Yuan, Z. H., & Chen, J. R. (2022). Research on CNN based facial recognition methods under the PyTorch 

framework. Information and Computer (Theoretical Edition), 34, 193-195. 

[16] Wang, R. (2019). Research on image generation and style transfer algorithm based on deep learning. Open Journal of Applied 

Sciences, 9(8), 661-672. doi:10.4236/ojapps.2019.98053. 

[17] Cui, J. (2022). Image Style Migration Algorithm based on HSV Color Model. In 2022 IEEE International Conference on 

Advances in Electrical Engineering and Computer Applications (AEECA), 111-114. doi:10.1109/AEECA55500.2022.9919019. 

[18] Huang, J. M. (2018). Application of Style Transfer Algorithm of Convolutional Neural Networks in Scene Processing. Journal 

of Yibin University, 18, 1-5. 

[19] Sun, J.S., Huang, Z.K., & Wang, H. (2023). Analysis of Image Mosaic Technology Based on Structural Similarity and Template 

Matching. Electronic Technology, 52, 38-41. 

[20] Liu, L., Fan, S., Ning, X., & Liao, L. (2017). An efficient level set model with self-similarity for texture segmentation. 

Neurocomputing, 266, 150-164. doi:10.1016/j.neucom.2017.05.028.