Rural Revitalization Perspective: Application of Visual Design Technology Based on Style Migration Algorithm in Packaging and Illustration Design of Agricultural Products in Hebei Province

The strategy of rural revitalization provides valuable opportunities for rural development, and today’s consumer market is developing rapidly, with new types of consumer characteristics emerging one after another [1-2]. In the context of promoting the development of rural revitalization, agricultural products have successfully attracted the attention of the majority of consumers through their own green, healthy, pollution-free and other purely natural good quality, as well as affordable, nutritious and other advantages, but some of the product packaging design there are some problems, which in turn reduces the consumer’s desire to buy [3-5]. The packaging design of agricultural products is an important business card to convey the quality of agricultural products to consumers, and its advantages and disadvantages will directly affect the degree of consumer recognition of the brand and quality of the agricultural products [6-7].

Rural industrial upgrading and development cannot be separated from the help of design, and a famous designer has pointed out that “packaging design is an important part of the product, and it is a key means of realizing the value of the product and the value of use [8-9].” In the environment of homogenization of agricultural packaging design, integrating regional culture into the design elements can highlight the regional characteristics of agricultural products, create unique regional cultural symbols, enhance the brand value of agricultural products, and promote product sales and industrial upgrading [10-12].

Illustration art is a contemporary and narrative form of painting art, which is usually used in books, advertisements, comics, movies and other visual arts [13-14]. It emphasizes elements such as color, shape, line and composition, and has a strong narrative and a high degree of expressiveness and visual appeal, and thus has been widely used in packaging design [15-16]. Packaging design as a comprehensive design activity, in addition to protecting the basic function of the product, it can also create an external image and internal quality for the product that meets the market demand and consumer preferences through innovative design, so as to enhance the market competitiveness of the product [17-19]. Therefore, the application of illustration art in packaging design can give personalized and distinctive features to products, highlight the selling points of products, and better meet the needs and expectations of consumers Through the design and application of illustration, it can give distinctive images and cultural connotations to rural revitalization projects, agricultural product brands and so on, enhance the brand awareness and reputation, and promote the development of rural economy [20-23].

Agricultural packaging design has a pivotal role in rural revitalization, which can not only enhance the product image and market competitiveness, spread the rural brand culture, but also promote the quality of agricultural products, expand sales channels and market space, enhance the image and development of the countryside, and inject new vitality and power for rural economic development and social progress [24-26]. In promoting the implementation of rural revitalization strategy, strengthening the research and practice of packaging design of agricultural products has a role that cannot be ignored [27-28].

The packaging design of agricultural products to a certain extent affects the consumer’s preference and desire to buy agricultural products, Kwaku, A. R et al. elaborated that the design and packaging of agricultural products is of great significance to the market value and performance of agricultural products, and analyzed the influence mechanism in depth [29]. The regional cultural characteristics of the origin of agricultural products also have some inspiration for the packaging design of agricultural products, Zhou, C analyzed the idea of packaging design of characteristic agricultural products, proposed the organic integration of regional and traditional cultural characteristics and modern green concepts, introduced into the packaging design of agricultural products to shape its unique brand image [30]. Some scholars have also explored the feasibility of three-dimensional virtual reality technology in the packaging design of agricultural products. Fu, L et al. explored the issue of three-dimensional presentation of packaging design of agricultural products, and concluded that the application of virtual reality technology can help to display the structure of the product packaging and then design and optimization [31]. At present, the packaging design of agricultural products focuses on value, cultural characteristics and structural optimization, and less research on artistry.

Wu, Y et al. revealed that the introduction of illustration art into the packaging design of local specialties added cultural beauty Han and artistic features to the local specialties, and promoted the improvement of consumers’ three-dimensional cognition of the local specialties [32]. He, J based on the case of Nongfushanquan, analyzed the product packaging design concept of creativity+products+communication, and revealed that the integration of exquisite illustration art into product packaging effectively stimulated the consumer’s desire for consumption [33]. The introduction of illustration art into packaging design has a positive effect on product image enhancement and consumer purchase, so it is necessary to further explore the best method of integrating illustration art into product packaging design, Wang, J proposed an illustration classification technique based on weighted augmented deep generative network to promote the level of packaging design, which effectively strengthens the marketing and branding image of the product packaging design [34]. There are more studies on the integration of illustration art into product packaging design, but there are fewer studies on the integration of illustration art into packaging design of agricultural products, and in the context of rural revitalization nowadays, how to promote the better flow of agricultural products into the market has always been full of concern, so it is necessary to explore the feasibility of the integration of illustration art into the packaging design of agricultural products and the path of in-depth investigation.

This study first describes the advantages of digital illustration in product packaging design, including presentation, information conveying, emotional expression, fun, and sales enhancement. It then proposes an improved GANILLA model with the addition of the SE module and uses a style decorator as a style transformation migration module. Subsequently, both pixel RGB values and color histograms are used to compare the color information of input cartoon drawings and generated pixel drawings to explore the learning ability of both GANILLA and CycleGAN models for color features. The image similarity between GANILLA and CycleGAN is then explored by comparing the PSNR and SSIM values of the results of the two groups of experiments. Finally, the applicability of visual design based on style migration algorithm in the packaging and illustration design of agricultural products in Hebei Province is demonstrated by investigating the respondents’ perception of pleasure and satisfaction evaluation of the illustrations under the two methods of GANILLA and CycleGAN.

2

Theory of Illustration Advantage and Style Migration Algorithm

2.1

Advantages of Digital Illustration in Packaging Design of Agricultural Products in Hebei Province

1)

Show diverse style characteristics

The application of digital illustration is a very important design method for product packaging, and different digital illustration styles can bring more colorful display contents to the packaging design. Diversified style characteristics to help agricultural packaging to achieve a variety of ideas and creativity, giving more personalized elements of agricultural products in Hebei Province, to meet the growing aesthetic needs of consumers and visual perception. In addition, the basic function of packaging is to protect the product, in order to adapt to a variety of products, packaging materials have also become diverse, about the printing of digital illustrations, different materials will show different printing effects, packaging materials on the display of digital illustrations also play a crucial role, which requires designers to consider the packaging material factors while designing patterns.

2)

Convey clear product information

Packaging is the basic carrier for displaying product information, and the application of digital illustration on product packaging can make intuitive and infectious visual presentation for the product. Through the integration of product and brand information, designers use digital illustration to skillfully integrate the information into the picture, together with text description, so that the appearance and quality of agricultural products and other information can be fully demonstrated, thus enhancing the publicity effect and showing the artistic connotation of packaging. With the help of digital illustration, consumers can quickly recognize the product information according to the packaging pattern and understand the ingredients contained in the product.

3)

Facilitate emotional expression

Transmitting emotions to consumers is one of the important functions of packaging design, digital illustration can convert abstract text, scenes, emotions and these emotional information into easy to read and browse the image content, so that people can get the information they want more quickly and easily. Incorporating emotional expression can make the packaging content more profound and attractive. The illustrations of agricultural products in Hebei Province can be used to convey the concept of safe and pollution-free healthy diet to consumers, reflect the freshness and safety of the products, and satisfy the emotional needs of consumers who care about the quality of agricultural products.

2.2

GANILLA model with the addition of the SE module

2.2.1

GANILLA’s Web Style Migration Principle

The goal of GANILLA is to generate a “new” image from a given image with a certain style but unchanged content by giving an image style. Adopting the “unpaired” method, GANILLA works as shown in Figure 1. GANILLA builds a new generator with jump connectivity, i.e., 2 times downsampling of the feature map at each residual layer, followed by merging low-level features with high-level features by jump connectivity and upsampling in order to better transfer content features. Low-level features usually contain edge-like information that helps to generate the image and preserve the structure of the image.

More specifically, the downsampling phase starts with a convolutional layer with a 7*7 convolutional kernel, goes through an instance-norm layer, a ReLU layer, a max-pool layer, and after that, continues to connect four layers (residual layer), each containing two blocks, each starting with a convolutional layer, an instance-norm layer, ReLU layer, a convolutional layer, another instance-norm layer, after which the output is connected to the input of that block, and finally, this connected tensor is fed into the final convolutional and ReLU layers.

In the upsampling phase, the output of each layer of the downsampling is used to transfer lower-level features to the sum layer prior to the upsampling phase through longer jump connections (these connections are used to preserve content). The upsampling phase consists of a convolutional layer, an upsampling layer, and a sum layer that first transmits the output of the residual layers above it through the convolutional and upsampling layers, which are used to increase the size of the feature mapping so as to match the features of the previous layer. All the convolutional filters for upsampling are l*l kernels. Finally a convolutional layer with 7*7 convolutional kernel is used to output the 3 channel translated image.

The discriminator is a 70*70 PatchGAN for the image translation model, which consists of three convolutional blocks, each containing two convolutional layers, the filter size of the first block is 64, the filter size of the next block is doubled, and its loss function has a total of three parts:

First the Ll distance between the synthesized part of the image and the real image is compared, where N is the number of samples, y is the predicted pixels, x is the pixels of the real picture, and WHC is the width-height and number of channels of the image: (1) $L_{r e c} = \frac{1}{N} \sum_{1}^{N} \frac{{‖ y - x ‖}_{1}}{W H C}$ $${L_{rec}} = \frac{1}{N}\sum\limits_1^N {\frac{{{{\left\| {y - x} \right\|}_1}}}{{WHC}}}$$

This is followed by the discriminator against loss and x is the input broken image. (2) $L_{G A N} (G, D) = E_{x - p (x)} [\log (D x)] + E_{y - P G (x)} [\log (1 - D (G (x^{'})))]$ $${L_{GAN}}(G,D) = {E_{x - p(x)}}[\log (Dx)] + {E_{y - PG(x)}}\left[ {\log \left( {1 - D\left( {G\left( {{x^\prime }} \right)} \right)} \right)} \right]$$

Finally, there are joint losses: (3) $L = λ_{1} L_{r e c} + λ_{2} L_{g - a d v} + λ_{3} L_{p - α d v}$ $$L = {\lambda _1}{L_{rec}} + {\lambda _2}{L_{g - adv}} + {\lambda _3}{L_{p - \alpha dv}}$$

Finally Cycle GAN’s idea of cyclic consistency is followed to train GANILLA’s model.

2.2.2

Network structure and principle of the improved GANILLA

The image effect generated by the traditional CycleGAN network is easy to be distorted, because the balance between content and style cannot be achieved when extracting features with high subjective abstraction in the downsampling stage, while GANILLA uses jump connection to separate the content and style well, and achieves good results in the “recognition” and “selection” of style. Therefore, this paper proposes an optimization scheme based on the GANILLA network in terms of both network performance and feature extraction.

In terms of feature extraction, the SE block is added to the Residual Block to improve the expressiveness of the overall generative network so that it focuses on the key positions of the image, and the channel attention mechanism introduced by SENet adds a small number of parameters so that the model can better acquire features on different channels, thus improving the learning level for highly abstract art style learning.

In terms of the performance of the network, adding a small number of parameters also means that the computation amount of the generative network increases. According to the Ghost bottleneck structure of GhostNet as a reference, the Ghost module is added to the Residual block, thus reducing the parameters of the network model as well as the FLOPs, and thus realizing the model’s light weight. The network parameters are designed as follows:

The Ghost module with Ghost bottleneck structure is used in Residual Blocks to enhance the performance of the whole generative network, and then the SE block based on channel attention mechanism is added to make it enhance the model’s sensitivity to channel features and learn more details from the deeper level.

The loss function of the generative adversarial network designed by SE-GANILLA includes two generators and discriminators for Minimax loss and a CycleGAN-based cyclic consistency loss.

Adversarial loss is first applied in mapping networks. In the mapping relation G: X − > Y and its discriminator D_Y, denoted as: (4) $G^{*}, F^{*} = \arg \min_{G, F} \max_{D_{x}, D_{Y}} L (G, F, D_{X}, D_{Y})$ $${G^*},{F^*} = \arg {\min {_{G,F}}}{\max {_{{D_x},{D_Y}}}}L\left( {G,F,{D_X},{D_Y}} \right)$$

G The generated image G(x) is constantly adapted to the domain Y to increase the similarity between the two, and D_Y with the goal of accurately distinguishing between G(x) and y. G with the goal of pursuing minimization against the discriminator D, i.e., satisfying $L_{G A N} (F, D_{X}, X, Y)$ $${L_{GAN}}\left( {F,{D_X},X,Y} \right)$$.

Theoretically, the adversarial training can learn the mapping relationship between G and D at the same time, and when the network capacity is large enough, the original image set can be mapped to the target domain, and can be arranged in any order, and the mapping learned by the network can further guide the corresponding output distribution. In order to ensure that the input x_i of the original images can reach the expected output y_i and to save the mapping storage space as much as possible, the GANILLA lossy loop consistency is verified, and the verification principle is shown in Equation (5): (5) $\begin{array}{l} L_{c y c} (G, F) = E_{x ~ P d a t a (x)} [{‖ F (G (x)) - x ‖}_{1}] \\ + E_{y ~ P d a t a (y)} [{‖ F (G (y)) - y ‖}_{1}] \end{array}$ $$\begin{array}{l} {L_{cyc}}(G,F) = {E_{x\sim Pdata(x)}}\left[ {{{\left\| {F(G(x)) - x} \right\|}_1}} \right] \\ \qquad\qquad\quad+ {E_{y\sim Pdata(y)}}\left[ {{{\left\| {F(G(y)) - y} \right\|}_1}} \right] \\ \end{array}$$

The individual loss functions are concatenated by introducing λ parameters to continuously optimize the relative importance of x and the generated data, and the larger λ is, the closer each pair of discriminating generators will be, and the concatenation formula is as follows: (6) $\begin{array}{l} L (G, F, D_{X}, D_{Y}) = L_{G A N} (G, D_{Y}, X, Y) \\ + L_{G A N} (F, D_{X}, X, Y) + λ L_{c y c} (G, F) \end{array}$ $$\begin{array}{l} L\left( {G,F,{D_X},{D_Y}} \right) = {L_{GAN}}\left( {G,{D_Y},X,Y} \right) \\ \qquad\qquad\qquad\qquad+ {L_{GAN}}\left( {F,{D_X},X,Y} \right) + \lambda {L_{cyc}}(G,F) \\ \end{array}$$

The optimization function is finally converted into a very large very small value problem and the resulting optimization function is as follows: (7) $G^{*}, F^{*} = \arg \min_{G, F} \max_{D_{x}, D_{Y}} L (G, F, D_{X}, D_{Y})$ $${G^*},{F^*} = \arg {\min {_{G,F}}}{\max {_{{D_x},{D_Y}}}}L\left( {G,F,{D_X},{D_Y}} \right)$$

2.3

SAD as a style transformation module

Style migration is a non-realistic rendering technique related to texture synthesis. Local statistical information is usually utilized to achieve effective distribution alignment and texture splicing, these methods although produce good results but distribution alignment is limited to pairs of images with similar content and is not applicable to zero-sample style migration.

In this paper, we will for the first time use a style adornment decorator (SAD) as a style transformation migration module embedded in an encoder-SAD-decoder network framework.SAD belongs to a new class of block-based feature operations, specifically transforming content features into the semantically closest style features while minimizing the differences between their overall feature distributions.

2.3.1

Style transformation by feature manipulation

The image I is feature extracted using the trained feed forward network and the extracted features are $F = E_{θ_{e n c}} (I) \in ℝ^{H \times W \times C}$ $$F = {E_{{\theta _{enc}}}}(I) \in {\mathbb{R}^{H \times W \times C}}$$, HWC are the height, width and channel of the image respectively. The extracted features are reconstructed into an image represented as $\tilde{I} = E_{θ_{d e c}} (F)$ $$\tilde I = {E_{{\theta _{dec}}}}(F)$$ by a decoder, and the encoder and decoder modules are parameterized by θ_enc and θ_dec, respectively. And labeled F_c and F_s as extracted features for content image I_c and style image I_s. By transforming F_c to the domain of F_s, we hope that the transformed feature $F_{c S} = Ψ (F_{c}; F_{s})$ $${F_{cS}} = \Psi \left( {{F_c};{F_s}} \right)$$, can infer the spatial distribution of content features and the texture representation of style features. The challenge is to design a feature transformation that both globally adapts the F_cs domain to F_s, and at the semantic level allows F_cs to be associated with F_s and F_c. 1)

Normalized interrelationships

F_cs is generated by nonlinearly exchanging blocks in F_c with their nearest neighbors in F_s. This approach moreover allows F_cs to be combined with style patterns more often and explicitly binds F_cs and F_c through spatial distribution.

Finding nearest neighbor blocks then uses normalized interrelationships. From a statistical point of view, NCC is an efficient way to compute the correlation between two sets of data, with values in the range $[- 1, 1]$ $$\left[ { - 1,1} \right]$$. For images, each pixel point can be represented by a red, green and blue triad (RGB), so an image can be seen as a collection of triads. If a subset of one set happens to match another set, the NCC value is 1, indicating high correlation. Conversely, -1 means irrelevant. This is an effective nearest neighbor block search algorithm.

However from the formula it is found that every block of content feature needs to be compared with all the blocks of style feature because there is GPU accelerated convolution operation and NCC can be converted to one of the few steps of convolution. According to the experimental results it is found that F_cs retains the style, but it still favors the content features of F_c. Therefore matching the style block must strictly follow the local differences of the content block. As a result, the stylized image retains more of the content stuff and only a portion of the semantically aligned style patterns. In fact, two blocks can be matched as long as they are semantically aligned in some form, rather than being matched only if the blocks are similar in an overly restrictive way.

2)

The computational process of SAD

The idea of SAD is to roughly translate style patterns onto content features such that the distributions of F_cs and F_s are maximally aligned and specific style patterns are semantically visible on F_cs. This is achieved by loosening block-to-block interrelationships in a normalized coordinate system enough to whiten the texture in the original feature space of both features. The swapped standardized feature ${\bar{F}}_{C S}$ $${\bar F_{CS}}$$ corresponds explicitly to the original stylistic feature, so we are able to reconstruct the stylized feature F_CS from the stylistic feature, statistically following both the style of F_s and the spatial layout of the content feature F_c. The following is the exact computational procedure for SAD: (1)

We project F_c and F_s to the same spatial dimension, writing ${\bar{F}}_{c}$ $${\bar F_c}$$ and ${\bar{F}}_{s}$ $${\bar F_s}$$. The two projected features should be compact enough and have similar scales for subsequent processing. The projection operations of content and style features are shown in Eqs. (8) and (9): (8) ${\bar{F}}_{c} = W_{c} \otimes (F_{c} - μ (F_{c}))$ $${\bar F_c} = {W_c} \otimes \left( {{F_c} - \mu \left( {{F_c}} \right)} \right)$$ (9) ${\bar{F}}_{s} = W_{s} \otimes (F_{s} - μ (F_{s}))$ $${\bar F_s} = {W_s} \otimes \left( {{F_s} - \mu \left( {{F_s}} \right)} \right)$$

Here ⊗ represents the convolution operation and $μ (F_{s})$ $$\mu \left( {{F_s}} \right)$$ and $μ (F_{c})$ $$\mu \left( {{F_c}} \right)$$ are the mean values of the features. The whitening kernels W_c and W_s are derived using some form of whitening matrix based on the covariance matrices $C (F_{c})$ $$C\left( {{F_c}} \right)$$ and $C (F_{s})$ $$C\left( {{F_s}} \right)$$. ${\bar{F}}_{c}$ $${\bar F_c}$$ and ${\bar{F}}_{s}$ $${\bar F_s}$$ keep the original data unique, but both follow the standard normal distribution statistically.

(2)

In the already regularized domain, align all elements in ${\bar{F}}_{c}$ $${\bar F_c}$$ with the closest elements in ${\bar{F}}_{s}$ $${\bar F_s}$$ in order to reconstruct ${\bar{F}}_{c}$ $${\bar F_c}$$ by recombining the corresponding elements in ${\bar{F}}_{s}$ $${\bar F_s}$$. In this case, the recombined regularized style features ${\bar{F}}_{c s}$ $${\bar F_{cs}}$$ contain the specific regularized style patterns, along with the spatial layout from the regularized content features. Thus, the matching and recombination operations seamlessly merge the desired content and style features into the result ${\bar{F}}_{c s}$ $${\bar F_{cs}}$$. We utilize regularized interrelationships for matching and recombination, a process that can be translated into multiple convolutional operations: (10) ${\bar{F}}_{c S} = Γ {({\bar{F}}_{s})}^{T} \otimes β (\bar{Γ} (\bar{F_{s}}) \otimes {\bar{F}}_{c})$ $${\bar F_{cS}} = \Gamma {\left( {{{\bar F}_s}} \right)^T} \otimes \beta \left( {\bar \Gamma \left( {\overline {{F_s}} } \right) \otimes {{\bar F}_c}} \right)$$

Where $Γ ({\bar{F}}_{s}) \in ℝ^{P \times P \times C \times (H \times W)}$ $$\Gamma \left( {{{\bar F}_s}} \right) \in {\mathbb{R}^{P \times P \times C \times (H \times W)}}$$ is the style kernel consisting of the blocks in ${\bar{F}}_{s}$ $${\bar F_s}$$, P is the size of the blocks, and P also affects the rendering, P the larger the image generated the more it will favor the global style pattern. In addition, $\bar{Γ} ({\bar{F}}_{s})$ $$\bar \Gamma \left( {{{\bar F}_s}} \right)$$ is the regularized style kernel which is divisible by the channel width ℓ₂ regularization of each block. The block matching process first convolves the regularized content features ${\bar{F}}_{c}$ $${\bar F_c}$$ and $\bar{Γ} ({\bar{F}}_{s})$ $$\bar \Gamma \left( {{{\bar F}_s}} \right)$$, and then β(x) is a binarization of the resulting scores such that the maximum value along the channel is 1 and the rest is 0. Then, recombination is applied by $\bar{Γ} {({\bar{F}}_{s})}^{T}$ $$\bar \Gamma {\left( {{{\bar F}_s}} \right)^T}$$ applying an anti-convolution operation to regenerate the style pattern from the binary scores.

The block matching between ${\bar{F}}_{s}$ $${\bar F_s}$$ and ${\bar{F}}_{c}$ $${\bar F_c}$$ exhibits complete correspondence because ${\bar{F}}_{c}$ $${\bar F_c}$$ and ${\bar{F}}_{s}$ $${\bar F_s}$$ have the largest overlap for each other, so that every element in ${\bar{F}}_{c}$ $${\bar F_c}$$ can find a suitable correspondence in ${\bar{F}}_{s}$ $${\bar F_s}$$, and every element in ${\bar{F}}_{s}$ $${\bar F_s}$$ can be found. Similar to Figure 3.7. The red boxes represent the results after regularization, while the green ones show the regularized interrelationships. The generated ${\bar{F}}_{c s}$ $${\bar F_{cs}}$$ applies style patterns for all gray parts of the content features. The tendency toward stylistic richness can be reduced to keep it simple and integrated. The matching and reorganization steps actually add two style control convolutional layers in the feedforward network, and this implementation is effective in existing deep learning platforms.

(3)

We reconstruct the reorganized regularized style block ${\bar{F}}_{c s}$ $${\bar F_{cs}}$$ into the domain of style feature ${\bar{F}}_{z}$ $${\bar F_z}$$. F_cs is accomplished by operations performed by ${\bar{F}}_{c s}$ $${\bar F_{cs}}$$ and F_s. The specific procedure is that we apply the coloring transformation about F_s to the reorganized features, and the reconstruction formula for the stylized features is: (11) $F_{c s} = C_{s} \otimes {\bar{F}}_{c s} + μ (F_{s})$ $${F_{cs}} = {C_s} \otimes {\bar F_{cs}} + \mu \left( {{F_s}} \right)$$

C_s in Eq. (11) is the coloring kernel, which is derived from the covariance matrix $C (F_{s})$ $$C\left( {{F_s}} \right)$$.

Similar to AdaIN and WCT, both the mapping and reconstruction processes match the 2D data of the stylized features F_CS with the stylized features of F_s. This process is similar to $G r a m (F_{s}) = G r a m (F_{c s})$ $$Gram\left( {{F_s}} \right) = Gram\left( {{F_{cs}}} \right)$$.

Figure 2 shows a comparison of AdaIN, WCT, and SAD principle effects. From Fig. 2 we can see the difference between these three. Different style conversion modules have different feature distributions, AdaIN does not completely discard the texture from the content features, and in turn converts this texture feature to F_cs. WCT has a similar Gram matrix for F_cs and F_s, but we can also see that strange features appear. The converted features all retain slight content features. The SAD matches regularized features ${\bar{F}}_{s}$ $${\bar F_s}$$ and ${\bar{F}}_{c}$ $${\bar F_c}$$, and the resulting ${\bar{F}}_{c s}$ $${\bar F_{cs}}$$ applies blue style features for both gray parts of the content features. This idea comes from the paper whose core idea is to propose an efficient and effective Avatar-Net that enables visually plausible multi-scale migration of arbitrary styles.

SAD not only matches the global style distribution, but also accurately recovers the details of undistorted style patterns. This method reaches a good breakthrough in terms of visual quality and efficiency.

2.3.2

SAD-based network design for rapid style migration

Fig. 3 shows the design of SAD based fast style migration network. Style migration requires a content image and a style image of any style as inputs and then extracts F_c and F_s from the encoder module. Then follow the steps described in the previous section to first perform feature regularization, then recombination and finally reconstruction operation. The changed content features are put into the decoder to generate the target resultant image.

3

Application of GANILLA model in illustration

3.1

Comparison of GANILLA and CycleGAN style migration models

3.1.1

Pixel Image Enhancement Research

CycleGAN can also be a model that generates texture features with a more pixel art style, this section focuses on comparing the color information of GANILLA and CycleGAN. It further explores the learning ability of different models for color features. In this section, both pixel RGB values and color histograms are used to compare the color information of the input cartoon drawings and the generated pixel drawings. Table 1 presents the statistics of RGB values using the two comparison methods.

Table 1.

Comparison method RGB value statistics table

Graphics	Part 1	Part 2	Part 3	Part 4	Manhattan distance
Image 1	(98,73,83)	(156,103,84)	(263,148,106)	(284,251,196)	-
CycleGAN	(68,54,63)	(127,87,65)	(238,127,85)	(262,232,173)	38.20
GANILLA	(80,67,78)	(139,98,82)	(255,145,101)	(279,245,188)	78.33
Image 2	(81,55,41)	(128,96,88)	(193,181,154)	(247,250,238)	-
CycleGAN	(66,49,37)	(117,89,80)	(179,172,138)	(226,227,205)	40.67
GANILLA	(73,51,37)	(119,92,86)	(182,176,151)	(235,241,230)	81.52

From Table 1, it can be seen that the RGB values of both images of GANILLA are closer to the RGB values of the input image, and the Manhattan distance is much smaller than CycleGAN. This indicates that GANILLA does a much better job of retaining image color information due to the multi-scale feature multiplexing and the color discriminator setting, which makes it more capable of learning color features, and produces pixel images that are better than CycleGAN in terms of visual quality.

Fig. 4, Fig. 5 and Fig. 6 show the comparison of the RGB three-channel color histograms of the first set of images, respectively. As the color histogram represents the proportion of different colors in the whole image. From the figure, we can see that in the three channels of R, G, and B, the peak part of GANILLA’s histogram is closer to the position of the peak of the cartoon image, which represents that most of the pixel points of GANILLA’s grayscale values in the three channels are closer to the grayscale values of the cartoon image, and the brightness of the image is more consistent. So GANILLA performs better than CycleGAN in the retention of color features, which proves the effectiveness of setting a color discriminator in GANILLA.

3.1.2

Image similarity study

The PSNR and SSIM values of the experimental results of the two groups of models, GANILLA and CycleGAN, are shown in Table 2.For the PSNR metric, which reflects the distortion, the larger its value, the greater the effective information of the image and the less the content is lost. For the SSIM metric, it is a measure of image similarity, taking into account visual characteristics. The larger this value is, the closer the two images are to each other and the less style loss. When the SSIM value is 1, it means that the two images are exactly the same. Although both migration times are shorter, the SSIM values in CycleGAN are both negative and the similarity is poorer.The GANILLA model has a higher similarity and is closer to the real effect image with a better migration effect.

Table 2.

PSNR and SSIM values of GANILLA and CycleGAN

Method	Image	PSNR	SSIM
CycleGAN	a	7.254	-0.054
	b	5.654	-0.088
	c	9.587	-0.011
	d	6.331	-0.107
GANILLA	a	11.543	0.321
	b	8.603	0.152
	c	12.467	0.476
	d	9.572	0.197

3.2

Evaluation of Illustration Design Effectiveness under Different Methods

3.2.1

Illustration pleasantness

Figures 7 and 8 show for the respondents’ perception of the pleasantness of the illustrations under the GANILLA and CycleGAN methods, respectively.

In general, the research group thinks that the illustrations under the GANILLA method can bring more relaxation, pleasure and comfort, calmness and healing, and can easily cause emotional resonance, while under the CycleGAN method, it is more likely to cause thinking about life, and some of the research group will feel fearful because of the picture elements under the CycleGAN modeling method, but at the same time, there is also a 30.7% of the group thinks that CycleGAN modeling can make them feel very happy and relaxed, which is due to the differences in individual perception.

3.2.2

Illustration satisfaction

Figure 9 shows the satisfaction research of the illustrations under the GANILLA method. From Figure 9, it can be seen that the viewers are most satisfied with the healing perception, followed by the color matching, and 60.2% of the people are most satisfied with the color matching of the illustrations under the GANILLA method. More than half of the people think that the visual presentation of the illustrations using the GANILLA method is good and the images are healing.

The satisfaction survey of illustrations under CycleGAN method is shown in Figure 10, the highest score is the visual effect of illustrations, for the research of healing of illustrations under CycleGAN method, 30.2% and 32.2% of groups think that illustrations under CycleGAN method are more healing and very healing, but there are still 6.9% of groups can not feel healing in illustrations.

Of the illustrations under the two methods, the illustration under the GANILLA method touches the group more.

4

Conclusion

In this study, the following conclusions are obtained by comparing the two models, CycleGAN and GANILLA, in terms of both pixel image enhancement and image similarity, as well as evaluating the effects of different illustration designs under the two methods. 1)

The RGB values of both images of GANILLA are closer to the RGB values of the input image, and the Manhattan distance is much smaller than CycleGAN. This indicates that GANILLA is more capable of learning color features, and the generated pixel images are better in terms of visual quality than CycleGAN.

2)

In the R, G, and B channels, the crest part in the histogram of GANILLA is closer to the crest position of the cartoon image, which proves the effectiveness of setting the color discriminator in GANILLA.

3)

The GANILLA migration process uses the least amount of time and has a wider range of application scenarios, it costs less to acquire data samples compared to CycleGAN, and the final generated image has a better overall effect and a better evaluation.

4)

The research group thinks that the illustrations under the GANILLA method can bring more relaxation, pleasant and comfortable, calm and healing feeling, and easily cause emotional resonance. The illustrations using the GANILLA method touch the group more.

The above series of experimental results prove the applicability and effectiveness of the visual design based on the style migration algorithm proposed in this paper for the packaging and illustration design of agricultural products in Hebei Province.

Język:: Angielski

Częstotliwość wydawania:: 1 razy w roku
Dziedziny czasopisma:: Nauki biologiczne, Nauki biologiczne, inne, Matematyka, Matematyka stosowana, Matematyka ogólna, Fizyka, Fizyka, inne

Kanał RSS czasopisma

Rural Revitalization Perspective: Application of Visual Design Technology Based on Style Migration Algorithm in Packaging and Illustration Design of Agricultural Products in Hebei Province

Xu Han

Linna Zhao

Data publikacji: 24 mar 2025

Otrzymano: 17 paź 2024

Przyjęty: 07 lut 2025

DOI: https://doi.org/10.2478/amns-2025-0777

Słowa kluczoweStyle migration algorithm, GANILLA, CycleGAN, Digital illustration, Hebei Province agricultural products

© 2025 Xu Han et al., published by Sciendo

This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Słowa kluczowe
Style migration algorithm, GANILLA, CycleGAN, Digital illustration, Hebei Province agricultural products