A Comprehensive Study of Art Image Style Transfer Methods Based

on Generative Adversarial Networks

Liyuan Huang

, Hanlin Liu

and Jiaqi Wen

Aberdeen Institute, South China Normal University, Foshan, Guangdong, 523000, China

Computer Science, Xu Hai College, China University of Mining and Technology, Xuzhou, Jiangsu, 221000, China

School of Information Science, Guangdong University of Finance & Economics, Guangzhou, Guangdong, 510000, China

Keywords: Art Image Style, Transfer Method, Generative Adversarial Networks.

Abstract: Image style transfer is a cutting-edge technique that seamlessly merges one image's content with another's

distinct style. The rapid progress of deep learning has led to significant advancements in image style transfer

technology. Nevertheless, this technology still encounters several issues, such as the inability to attain the

optimal expression effect of artistic attributes, and the mismatch between semantic and style characteristics.

Based on the generative adversarial network (GAN), this paper examines the improved algorithmic

applications of image style transfer technology in ink painting, animation, and oil painting. Additionally,

using quantification and comparative analysis of the outcomes of the improved style transfer algorithm

applied in diverse art forms, Foreseen are the obstacles to be tackled and the expected development path of

image style transfer technology in the future. The application of image style transfer technology in the domain

of art still demands more efficient algorithms and more artistic outputs. This study focuses on summarizing

popular algorithms in image style transfer technology and driving forward innovation in style transfer

techniques.

1 INTRODUCTION

Image style transfer constitutes a deep learning

technique, it can transfer the style of one image to

another, thereby generating a new image. In recent

years, image style transfer has been extensively

applied across various fields. In the transportation

sector, Lin has proposed day-night style transmission

for detection purposes of vehicles during the night, to

reduce the incidence of car accidents (Lin, Huang,

and Wu, et al, 2021). In medicine, Yin improves

medical image accuracy through a context-aware

framework (Xu and Li, 2020). Lv adds to the

authenticity of blood vessel image generation through

the application of deepnet (Tmenova, Martin, and

Duong, 2019). In art, migrating imagery is used more

frequently. By incorporating the Cantonese dialect

into the creation of porcelain patterns, the Cantonese

porcelain culture can be preserved and passed on

(Chen, Cui, Tan, et al., 2020).

https://orcid.org/0009-0005-7243-1963

https://orcid.org/0009-0005-1265-9923

https://orcid.org/0009-0001-9110-1039

Painting is an ancient form of artistic expression.

The most common way to create paintings with

different styles is Generative Adversarial Network

(GAN). Chinese Ink Painting with rice paper and ink.

They integrate and intermingle, creating a sense of

depth. Anime uses simple strokes to create images.

Oil paintings have rich colors and strong three-

dimensional texture. Most experiments on style

transfer employ GAN to generate new images.

Nevertheless, as time goes by, the majority of basic

GANs to doing things fail to satisfy people's

requirements for generating images. There are many

problems arising: differences in skills and styles

between ink painting and Western painting contribute

to style transfer in ink painting's suboptimal

performance. When transferring animation styles,

there exist problems regarding anime image feature

texture is missing and the generated image quality is

not good. When transferring oil painting styles, there

Huang, L., Liu, H., Wen and J.

A Comprehensive Study of Art Image Style Transfer Methods Based on Generative Adversarial Networks.

DOI: 10.5220/0013512100004619

In Proceedings of the 2nd International Conference on Data Analysis and Machine Learning (DAML 2024), pages 171-178

ISBN: 978-989-758-754-2

171

are often issues with image local minutiae are missing

and style area and content do not match.

To address these issues, several new models have

emerged. The purpose of this article is to discuss

some new GAN-based model for use in various types

of paintings and introduce their advantages and

disadvantages. The aim is to provide some basis for

the subsequent research on style transfer in paintings.

2 METHOD

In this section, the application of image style transfer

technology in the field of art will be elaborated,

focusing on three main categories: ink painting,

animation, and oil painting.

2.1 The Overview of GAN

GAN, which stands for Generative Adversarial

Network, is a type of deep learning model comprised

of two primary components: the generator and the

discriminator. As illustrated in Figure 1, the generator

G takes random noise Z as input and produces an

image G(z), while x represents the input image. The

discriminator D assesses whether x is a real image and

outputs D(x) as the probability of its authenticity. A

probability closer to 1 indicates higher authenticity of

the image; conversely, a probability closer to 0

suggests lower authenticity. The objective of G is to

generate images that are highly realistic to deceive D,

while D strives to differentiate real photos accurately.

They engage in a mutual game to enhance their

respective discriminative capabilities.

2.2 Ink Painting

Due to the significant differences between Chinese

ink painting and Western painting techniques, the

direct application of existing image style conversion

methods is ineffective. Hu improved on the existing

ChipGAN model to promote the standard and white

space effect related to the generated image, replacing

the ResNet residual network in the generator with a

residual dense network (RDN), which enables the

network to reuse shallow features of the image and to

extract more feature information by combining

shallow and deep features (Hu, 2023). The PatchGAN

discriminator is substituted with a multi-scale

discriminator to enhance the discriminative ability of

images at different scales. The white space loss is

added based on the original loss function, the image

background is processed by threshold segmentation,

and the background white space is constrained using

the L1 loss and the SSIM loss to generate an image

that is more in line with the ink style.

2.3 Animation

To solve the previous problem of imperfect style

migration of anime images, derived from edge

improvement and coordinated attention Hong et al.

proposed an animation translation method, called

FAEC-GAN, to help complete the task of migrating

from real photos to anime faces (Lin, Xu, Liu, 2023).

Firstly, they introduced an edge discriminative

network consisting of an edge detection module and

an edge discriminator. The edge detection module

obtains edge information from the image and sends it

Figure 1: The structure of GAN (Photo/Picture credit: Original).

DAML 2024 - International Conference on Data Analysis and Machine Learning

172

to the edge discriminator for evaluation. Next, they

designed a novel loss function that measures the

image discrepancy between the focusing frequency

loss and the spatial loss (Lin, Xu, Liu, 2023).

Zhao proposed a new image canonisation architecture

that extracts anime image features independently,

making it controllable and scalable in practical

applications (Zhao, Zhu, Huang, et al., 2024). The

architecture is a GAN model based on multiple

decompositions. The image features generated by the

generator network training are decomposed and

extracted, and the texture module takes the extracted

texture and trains the image using a discriminator to

create textures with more cartoon image features.

Such a structural module uses the pre-trained network

to efficiently extract image structural features and

preserve image content information (Zhao, Zhu,

Huang, et al., 2024).

Li proposed an improved GAN-based style

migration algorithm for anime images (Li, Zhu, Liu,

et al., 2023). Firstly, the improved inverted residual

block and the efficient attention mechanism are used

to form a feature transformation module, which

enriches the local feature attributes of the migrated

image and improves the expression ability of style

features; secondly, the style loss function is improved

to suppress the interference of colour and luminance

on texture learning. Experiments reveal that the

introduced algorithm excels in producing images with

superior quality and enhanced visual realism (Li, Zhu,

Liu, et al., 2023).

2.4 Oil Painting

To solve the detail loss and blurring that appears

locally in the generated oil painting images problem,

Han proposed a multi-feature extractor to perform

style transfer of shape, texture, color, and space in oil

painting (Han X, Wu Y, and Wan R, 2023). The

extractor contains U-net, MiDas, texture extraction

network, and chromatic histogram. Meanwhile, the

autoencoder structure is employed as the main

framework for feature extraction. After identifying

the input and style images, DE-GAN is trained using

the model architecture, generating a network that

shares style parameters with the feature extraction

module. Finally, implement the generation of realistic

photos in the oil painting style (Han, Wu, and Wan,

2023).

To address the challenges of generating and

reconstructing image details in oil painting, Liu

proposed a method that combines the LBP algorithm

with an improved loss function (Liu, 2021). Liu

initially applied the Wasserstein metric to the GAN's

objective function to enhance training stability(Liu,

2021). Then add gradient penalty to loss function

(WGAN-GP), it can deal with gradient vanishing

during training. By incorporating noise control in the

CycleGAN loss function, boundary details and

surface patterns are enhanced after oil painting style

transfer, along with the addition of LBP structural

features and total variation.

3 RESEARCH

3.1 Quantitative Results of Ink

Painting Experiments

Hu's experiments found that the ChipGAN model,

which incorporates a residual dense network and

introduces a multi-scale discriminator, can largely

improve the quality of the generated images (Hu,

2023). The quality of an image is assessed by

quantitatively analyzing the image using Peak Signal

Noise Ratio (PSNR), Structural Similarity (SSIM),

Frechette's Distance (FID), and Cosine Similarity

(CosSim) as metrics (Hu, 2023). PSNR is used to

measure the generation quality of an image, and the

higher its value, the closer the style-migrated image

is to the original content, maximizing the retention of

content information. FID is used to compute the

similarity between two images in terms of feature

vectors, the smaller its value, the more similar it

represents the distribution of the generated image and

the real image in space. CosSim is used to measure

the pixel-wise similarity of two images, with higher

values indicating a smaller angle between the two

vectors represented by the image features and a

higher degree of similarity. SSIM is a measure of the

structural similarity of the two images, where higher

values indicate that the reconstructed image is more

similar to the original image in terms of brightness,

contrast, and structural inverse.

Hu uses four methods, method 1 is ChipGAN,

method 2 is ChipGAN based on RDN, method 3 is

ChipGAN using a multi-scale discriminator, and

method 4 is ChipGAN based on RDN and multi-scale

discriminator. Among them, method 1 is for the

ChipGAN model without any improvement, which is

a variant of GAN. The basic principles include:

generative network generates IC layout according to

the input design parameters; discriminative network

accepts the generated layout and the real layout and

outputs the true-false discriminative results;

optimizing the generative network and discriminative

network through adversarial training, so that the

quality of the generated layout is continuously

A Comprehensive Study of Art Image Style Transfer Methods Based on Generative Adversarial Networks

173

improved, and at the same time, the discriminative

ability of the discriminative network is improved.

Finally, the quality of the generated plates is

gradually improved by optimizing the loss function.

For method 2, ChipGAN based on RDN is combined

with the basic principle that the generative network

utilizes a deep convolutional neural network, which,

integrated with the architecture of RDN, can enhance

the multilevel feature extraction capability to

generate accurate and high-quality plat maps. For

method 3, which introduces a multi-scale

discriminator, the basic principle is that the generated

IC layout is evaluated for multi-level and multi-scale

fidelity, and combined with the adversarial training of

the GAN, the generative network is continuously

optimized to generate a layout that is realistic at

different scales, to achieve automatic generation of IC

layouts with higher accuracy and quality. For method

4, both combines RDN and multiscale discriminators.

For the performance comparison of these four

methods (Hu, 2023), as shown in Table 1, it can be

concluded that the SSIM, PSNR, and CosSim are

greater than the base ChipGAN model when RDN or

multiscale discriminator is introduced alone. The

reconstructed image is closer to the original image in

terms of brightness, contrast, and structure, and the

generated image is of higher quality, with less

difference from the original one and higher similarity

of image or text features. At the same time, the FID is

all lower than the base ChipGAN model, representing

a closer distribution of the generated image to the real

image in the feature space. The performance data is

further improved by combining both residual dense

networks and multi-scale discriminators (Hu 2023).

Hu's experimental results demonstrate that the ink-

style images generated by the improved model have

significant improvement in both visual quality and

white space effect (Hu, 2023). As in Figure 2, it can

be seen from the examples of apples, flowers, and fish

that the ink-style images generated by the improved

model have significant improvement in visual quality

and white space effect, and have good generalization

ability on different datasets.

Table 1: Comparison of different model performance

Evaluation

metrics

Method

SSIM 0.8064 0.8127 0.8092 0.8191

PSNR 13.9860 14.5329 14.2367 14.7879

FID 172.44 167.04 166.95 165.62

CosSi

0.9733

0.9825

0.9785

0.9841

Figure 2: Examples of apples, flowers, and fish in the

improved model of image generation representation (Hu,

2023)

3.2 The Result of Animation

3.2.1 FAEC-GAN

LIN used two datasets, self2anime, and ce2anime, to

propose the FAEC-GAN method, a style migration

method based on edge enhancement and coordinated

attention mechanisms (Lin, Xu, Liu, 2023).

To validate the FAEC-GAN model's

effectiveness, LIN used FID and KID measures in the

experiments. On self2anime, FAEC-GAN reduced

FID by 1.95 and KID by 0.37 compared to the best-

performing ACL-GAN in the baseline (Lin, Xu, Liu,

2023). On the ce2anime dataset, FAEC-GAN reduces

FID by 3.03 and KID by 0.61 compared to

SpatchGAN. By delivering top results on several

datasets, FAEC-GAN proves its capacity to learn

from different data distributions and generate highly

realistic anime faces. In addition, the method

achieved the lowest scores on both evaluation

metrics, which further demonstrates that FAEC-GAN

performs very well regardless of which metric is used.

Table 2: Comparison of six methods for FID and KID under

two datasets (Lin, Xu, Liu, 2023)

Dataset

Model

Self2anime Ce2anime

FID KID*100 FID KID*100

FAEC-GAN 92.92 2.91 60.28 2.71

cle-GAN 114.54 3.80 70.90 3.67

U-GAT-IT 105.78 3.93 68.28 3.21

NICE-GAN 112.62 5.41 68.04 3.61

ACL-GAN 94.87 3.28 63.69 2.81

atchGAN 98.78 3.71 63.31 3.32

3.2.2 GAN Expansion of Animation 1

Zhao proposed a GAN-based method to combine the

features used in the generator for extracting deep

networks with the attention mechanism, and in terms

of content, real-world photographs were used as test

data. Stylistically, cartoon images from the films of

DAML 2024 - International Conference on Data Analysis and Machine Learning

174

Table 3: Uses FID to assess the similarity between real and synthetic anime images and the stability of different styles of

results relative to baseline results (Zhao, Zhu, Huang, et al., 2024).

Methods Photo CycleGAN CartoonGAN GDWCT (Zhao, 2024)

FID to cartoon 165. 70 140. 37 145. 21 136. 21 110. 39

FID to photo N/A 121. 11 86. 48 100. 29 35. 79

Methods Photo Ha

ao St

le Pa

rika st

le ShinkaiSt

(

Zhao, 2024

)

FID to cartoon 165. 70 127. 35 127. 05 129. 52 110. 39

FID to

hoto N/A 86. 48 118. 56 37. 96 35. 79

two directors, Hayao Miyazaki and Makoto Shinkai,

are used respectively, and the proposed method is

compared with Cycle GAN, CartoonGAN, and

GDWCT (Zhao, Zhu, Huang, et al., 2024).

To find the differences between the methods more

precisely, Zhao employed FID to assess the similarity

between authentic and generated anime images, as

well as the consistency of results across various styles

compared to the baseline. As in Table 3, Zhao's

methods all have minimum errors (Zhao, Zhu, Huang,

et al., 2024).

To compare the visual quality of the images

generated by each method more objectively, TABLE

4 shows the percentage of each method selected as the

best method, with higher percentages indicating more

popularity, which is used to assess the visual quality

of the images.

Table 4: Percentage of the four methods selected as the best

method under different styles (Zhao, Zhu, Huang, et al.,

2024).

Methods Hayaostyl

Paprika

style

Shinkai

style

CycleGAN 16.20% 7.37% 15.24%

CartoonGAN 32.40% 39.69% 40.36%

GDWCT 10.88% 3.1% 2.3%

(Zhao, 2024) 40.50% 49.84% 49.84%

3.2.3 GAN Expansion of Animation 2

The content image set selected for Li's experiments is

from real images on the Flickr website, and the style

image set is from four sets of anime film frames

produced by Hayao Miyazaki and Makoto Shinkai

(Li, Zhu, Li, et al., 2023). The improved method uses

a channel blending operation combined with a

modified inverted residual block to form a feature

transformation module to enhance the local feature

attributes of the images, and in the meanwhile,

invokes an efficient attention mechanism to further

improve the stylistic feature expression capability.

Li combined the proposed algorithm with three

algorithms, Cycle GAN, CartoonGAN, and

AnimeGAN, to perform anime style migration on two

content images in the test set (Li, Zhu, Li, et al.,

2023). According to the experimental results, the

CycleGAN algorithm generates over-stylized images

loses the semantic content of the input images, and

cannot accurately determine the object information.

CartoonGAN algorithm generates anime images with

green artifacts and more distortions in the level of

detail; at the same time, it loses the color information

of the original images due to the lack of constraints of

the color loss function. AnimeGAN algorithm still

has image details on local areas, but it is not possible

to determine the object information accurately. local

area still exists image content detail loss and point

artifacts, and there is a problem of structural adhesion

for the portrayal of the distant region. Li's proposed

algorithm incorporates an enhanced inverted residual

block within the feature transformation module,

which focuses on emphasizing local image details

while effectively maintaining global information

through feature reuse and fusion techniques (Li, Zhu,

Li, et al., 2023). At the same time, an efficient

attention mechanism is introduced to help the model

better focus on the stylistic feature information in

anime images. In the style loss function part, the color

and luminance information of the image is erased, so

that the generated image presents obvious anime

texture and avoids color shifting.

Table 5: Stylistic and Semantic FIDs of Images Generated

by Different Algorithms (Li, Zhu, Li, et al., 2023)

algorithm Hayao style Shinkai style

Style Semantic Style Semantic

FID FID FID FID

original 179.

— 135.

—

CycleGAN 123.

163. 55 106.

107. 32

CartoonGAN 157.

90. 64 114.

79. 58

AnimeGAN 160.

89. 12 116.

87. 73

(Li, 2023) 154.

71. 97 115.

63. 48

A Comprehensive Study of Art Image Style Transfer Methods Based on Generative Adversarial Networks

175

Li also computes the FID scores for images

produced by various algorithms on the test set. This

helps assess how closely the generated images match

the content of the content image and the style of the

style image. Separate FID scores are calculated for

content and style (Li, Zhu, Li, et al., 2023).

As shown in Table 5, the initial values of the style

FID of the original image are 179. 16 and 135. 81, and

the style FID values of their proposed algorithm are

second only to the CycleGAN algorithm and

CartoonGAN algorithm in Miyazaki and Makoto

Shinkai styles, whereas there is a significant decrease

in the semantic FID values, which reaches 71. 97 and

63. 48. The CycleGAN algorithm, compared with the

others, has the style FID score the lowest, this is

because the CycleGAN algorithm only matches the

stylistic information of the target image and ignores

the semantic information of the original image, so the

generated image produces semantic distortion (Li,

Zhu, Li, et al., 2023).

3.3 Oil Paintiong

3.3.1 DE-GAN

Han and three others proposed DE-GAN, 1. a style

transfer scheme based on GAN, capable of deeply

extracting the artistic style from artistic works to

target image (Han, Wu, and Wan, 2023). For

verifying the efficacy of this approach, four indicators

are used for a single picture as an evaluation criterion.

There are feature similarity index (FSIM), mean

SSIM index (MSSIM), image average gradient, and

average reasoning time. SSIM ranges from 0 to 1,

where larger scores signify greater image similarity

(Han, Wu, and Wan, 2023). FSIM is an extension of

SSIM, its evaluation system is similar to SSIM.

MSSIM is the greater the average gradient of the

image, the sharper the image, and the better the

texture details. According to this indicator, DE-GAN

was compared with StyleGAN and CycleGAN.

StyleGAN is a GAN-based model that uses style to

influence the face and body shape in the generated

images. CycleGAN is a model based on GAN that can

transform photos into oil painting styles. The

comparison results of these three methods are shown

in Table 6.

Table 6: Evaluation and comparison of different models

(Han, Wu, and Wan, 2023)

Methods StyleGAN CycleGAN

DE-

GAN

FSIM1 0. 72 0. 68 0. 74

FSIM2 0. 62 0. 60 0. 64

FSIM3 0. 64 0. 62 0. 66

MSSIM1 0. 52 0. 52 0. 54

MSSIM2 0. 48 0. 51 0. 53

MSSIM3 0. 50 0. 52 0. 53

avera

radient1 0. 31 0. 23 0. 41

avera

radient2 0. 45 0. 41 0. 52

average gradient3 0. 52 0. 58 0. 63

Average

Reasoning Time

for a Single

Picture

(

)

15. 64 26. 78 42. 63

These three methods can all effectively transfer

the painting style. However, it is clear from Table 6

that the migration images obtained by DE-GAN all

show small improvements in FSIM, MSSIM, and

average gradient. From this, it may draw such

conclusion, that the artistic images generated by this

method have better performance in structural

features, image distortion, image clarity, and texture

detail compared to others. This method is inferior to

StyleGAN and CycleGAN in speed.

3.3.2 Expansion of WGAN-GP

Liu proposed an Improved GAN for gradient

punishment to solve the difficulty of algorithm

training high, and the loss gradient of the generator

and discriminator disappears in the transfer of oil

painting style (Liu, 2021). WGAN-GP is a model that

uses Deep Convolutional Neural Network

architecture. It can solve the problem of loss gradient

disappearance of the generator and the discriminator

is difficult to train. This method is to transform a

model based on WGAN. To further assess the

efficacy of this approach., Contrast, SSIM, Entropy,

PSNR, MSE, and speed were used for evaluation.

Contrast and Entropy are two important indicators for

measuring image information in style transfer. MSE

is employed to evaluate the disparity between the

generated output and the target picture. Speed can be

seen as the algorithm efficiency. Liu uses CycleGAN

+ L1 regularization and CycleGAN to evaluate these

six indicators using this method (Liu, 2021). WGAN-

GP was only compared with this speed method. The

image quality of CycleGAN can be further improved

by CycleGAN + L1 regularization. A comparison of

these four methods is shown in Table 7.

DAML 2024 - International Conference on Data Analysis and Machine Learning

176

Table 7: Performance comparison of different methods

Methods CycleGAN CycleGAN+L1

(Liu Y, 2021) WGAN-GP

Contrast 28374.4 28736.2 29, 215. 9

SSIM 0.223 0.129 0. 251

Entropy 5.839 6.411 7. 216

PSNR 11.553 12.187 12. 638

MSE 3826.8 3769.6 3544. 7

Speed (Landscape 1) 2837.38s 2736.46 s 2637. 33 s 1384. 32 s

Speed (Landscape 2) 2365.32s 2645.56 s 2736. 73 s 1463. 23 s

eed

(

Landsca

e 3

)

3026.83s 2937.73 s 3173. 78 s 1183. 28 s

eed

(

Landsca

e 4

)

2653.43s 2736.82 s 2557. 38 s 1583. 41 s

In Table 7, most of the indicators of this approach

are better than CycleGAN + L1 regularization and

CycleGAN. This suggests that the visual quality of

this approach surpasses other methods and offers

higher practicality. In terms of efficiency, Original

WGAN-GP is more efficient than CycleGAN. The

results of time complexity for all are similar.

Therefore, to improve the speed of operation of the

law becomes a new problem.

4 CONCLUSIONS

Image style migration, as an emerging image

processing technique, has been widely used in several

fields This paper provides an overview of GAN-based

art image style translation techniques in three art

types: ink painting, animation, and oil painting, and

focuses on highlighting the significance of the

application, research implications and results of these

improved new models, as well as the advantages over

existing techniques.

In terms of ink painting, due to the special

characteristics of the style, such as white space and

other features that need to be fragmented to learn the

drawing, optimized based on ChipGAN is more

efficient in solving this problem, because it can

flexibly control the effect of the generated ink

paintings by inputting different conditional

information, such as brush strokes, ink color,

composition, etc., and perfecting the details of the ink

intensity, penetration patterns, etc. can also make the

generated ink paintings more realistic. Additionally,

the model can learn and master the characteristics of

ink drawings from ink drawing datasets, enabling it

to generate new works with strong generalization

abilities. Moreover, training ChipGAN models

requires significant computational power.

There are not many papers on anime as an image

style, and three methods are mentioned in this paper

which are applied to migrate two image styles,

landscape and people respectively. FAEC-GAN

addresses the issue of image edge distortion caused

by the migration process. In contrast, Zhao employs a

new architecture that efficiently extracts structural

features using a pre-trained network while preserving

image content. Additionally, Li's proposed algorithm

offers significant advantages in terms of quality and

visual realism in the generated images. Quality and

visual realism, Li's proposed algorithm has

significant advantages in terms of the quality and

visual realism of the generated images. Each of the

three approaches solves different problems faced by

the application, and it is a challenge to implement a

more comprehensive approach to get better results for

anime images.

AUTHORS CONTRIBUTION

All the authors contributed equally and their names

were listed in alphabetical order.

REFERENCES

Chen, S. S. C., Cui, H., Tan, P., Sun, X., Ji, Y. and Duh, H.,

2020. Cantonese porcelain image generation using

user-guided generative adversarial networks. IEEE

Computer Graphics and Applications, 40(5), pp.100-

107.

Dicke, R.H., Beringer, R., Kyhl, R.L. and Vane, A.B., 1946.

Atmospheric absorption measurements with a

microwave radiometer. Physical Review, 70(5-6),

pp.340.

Han, X., Wu, Y. and Wan, R., 2023. A method for style

transfer from artistic images based on depth extraction

generative adversarial network. Applied Sciences,

13(2), p.867.

Hong, L., Xu, C. and Liu, C., 2023. Computer animation

and virtual worlds. 2023.

Hu, B., 2023. Improvement study of ChipGAN, an image

inking style migration network. Master's thesis. Xi'an

Petroleum University.

A Comprehensive Study of Art Image Style Transfer Methods Based on Generative Adversarial Networks

177

Li, Y., Zhu, J., Liu, X., Chen, J. and Su, X., 2023. A style

translation algorithm for animation images based on

improved generative adversarial network. Journal of

Beijing University of Posts and Telecommunications,

pp.1-7.

Lin, C. T., Huang, S. W., Wu, Y. Y. and Lai, S.-H., 2021.

GAN-based day-to-night image style transfer for

nighttime vehicle detection. IEEE Transactions on

Intelligent Transportation Systems, 22(2), pp.951-963.

Liu, Y., 2021. Improved generative adversarial network and

its application in image oil painting style transfer.

Image and Vision Computing, 105, p.104087.

Tmenova, O., Martin, R. and Duong, L., 2019. CycleGAN

for style transfer in X-ray angiography. International

Journal of Computer Assisted Radiology and Surgery,

14, pp.1785-1794.

Xu, Y., Li, Y. and Shin, B.S., 2020. Medical image

processing with contextual style transfer. Human-

Centric Computing and Information Sciences, 10, p.46.

Zhao, W., Zhu, J., Huang, J., Li, P. and Sheng, B., 2024.

GAN-based multi-decomposition photo cartoonization.

Computer Animation and Virtual Worlds, 35(3),

p.e2248.

DAML 2024 - International Conference on Data Analysis and Machine Learning

178