Graphic design style transfer and aesthetic optimization algorithm based on a generative adversarial network.

Graphic design style transfer and aesthetic optimization algorithm based on a generative adversarial network - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Sci Rep . 2026 Apr 11;16:16883. doi: 10.1038/s41598-026-46316-0 Graphic design style transfer and aesthetic optimization algorithm based on a generative adversarial network Baiyun Deng 1 , Lili Ren 1, ✉ , TikFan Chan 2 , Chunyan Tang 3 Author information Article notes Copyright and License information 1 Solux College of Architecture and Design, University of South China, Hengyang, 421001 Hunan, China 2 Academy of Arts & Design, Tsinghua University, Beijing, 100084 Beijing China 3 College of Fine Arts, Hanshan Normal University, Chaozhou, 521041 Guangdong China ✉ Corresponding author. Received 2025 Oct 27; Accepted 2026 Mar 25; Collection date 2026. © The Author(s) 2026 Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/ . PMC Copyright notice PMCID: PMC13230764  PMID: 41965396 Abstract Current graphic design style transfer technology mainly focuses on geometric or texture features, while ignoring overall beauty and artistic expression, resulting in a mismatch between style and content, poor detail processing, and a lack of artistic appeal and true style presentation in the generated design. To this end, this article proposes a graphic design style transfer and aesthetic optimization algorithm based on a generative adversarial network (GAN). First, a graphic design image database with diverse styles is constructed; the GAN architecture is improved through generators, discriminators, and pre-trained VGG (visual geometry group) networks; an efficient channel attention mechanism and optimized inversion residual blocks are applied to enhance the model’s ability to capture aesthetic features; an aesthetic scoring model is designed by combining the loss functions of content, style, and generated images to ensure the visual appeal of generated images; VGG-19 (Visual geometry group-19) networks are used for pre-training, and the neural network parameters are optimized through the Adam algorithm to achieve efficient model training. The results show that the average values of SSIM and MSE (mean square error) for the improved GAN in this article are 0.93 and 0.027, respectively, in terms of content retention; the average value of MSE is 0.020 in terms of style similarity; the average values of PSNR (Peak Signal to Noise Ratio) and SSIM are 34.33 and 0.91 respectively in terms of image clarity. The study shows that the proposed method can not only improve the aesthetic quality and diversity of style transfer but also ensure the stability of image content, providing new theoretical and technical support for the field of graphic design style transfer. Keywords: Graphic design, Style transfer, Generative adversarial network, Aesthetic-aware optimization, Perceptual loss Subject terms: Engineering, Mathematics and computing Introduction As a modern form of visual communication, graphic design always seeks innovation and enhancement of aesthetic value. The significant progress of artificial intelligence (AI) technology, especially generative adversarial network (GAN) in image generation, has brought new directions for graphic design. As a powerful image generation model, GAN can produce highly realistic images through its unique interaction mechanism between generation and discrimination networks, and shows great potential in style transfer 1 , 2 , 3 . However, the graphic design style transfer based on GAN still faces challenges. How to achieve effective style transfer and aesthetic improvement on the basis of ensuring that the image content remains unchanged is an urgent problem to be solved. This study is dedicated to exploring the graphic design style transfer and aesthetic optimization algorithm based on GAN. An aesthetic scoring model is constructed by deeply analyzing the principles and applications of GAN in style transfer and image generation, and combining the actual needs of graphic design. The aesthetic scoring model aims to achieve rapid transfer and aesthetic improvement of multiple styles, while enhancing the stability and quality of image generation. The significance of this study is not only to provide new theoretical and technical support for graphic design style transfer, but also to open up a broader road for the application of AI in graphic design, which has far-reaching theoretical value and practical significance. This study uses GAN to construct a graphic design style transfer and aesthetic optimization algorithm. By integrating content retention and aesthetic loss, the generation effect is improved, and a random module is added to stimulate creativity. Experiments show that this method performs well in content retention, style similarity, image clarity and diversity. Compared with other methods, the SSIM and PSNR values of the generated images are higher; the MSE value is lower; the aesthetic scoring error is smaller. The innovation of this study lies in integrating the generative power of GAN and the feature extraction ability of the VGG network, and significantly improving the aesthetic quality of the generated images through the optimization of the aesthetic loss function. The main contributions of this study are: The proposed GAN-based graphic design style transfer and aesthetic optimization algorithm optimizes the generation effect by integrating content retention and aesthetic loss, and improves the content retention, style similarity, clarity and diversity of the image. Experiments have shown that this method surpasses other methods in multiple indicators, bringing new technical support for graphic design style transfer. This study not only performs well in static image transfer, but also opens up new possibilities for scenarios such as dynamic art text and multimedia graphic design. By optimizing models and loss functions, it is expected to further expand applications in the future and promote the development of AI in art design. By applying an efficient channel attention mechanism and improving the inverted residual block, the method enhances the model’s recognition of aesthetic features, improves the diversity and visual appeal of images, ensures a high degree of consistency between content and style, and injects high-quality creativity into graphic design. Even though the separate elements used in this research (GAN style transfer, perceptual loss, attention systems, and aesthetic evaluation networks) have been investigated in the previous literature, the novelty of this article is that these elements have been presented in a design-focused workflow of a single aesthetic-sensitive optimization network. Compared to the currently used strategies that consider aesthetics as a post-processing metric, this work introduces an aesthetic scoring model into the GAN training loss, allowing content preservation, consistency of the style, and visual attractiveness to be optimized simultaneously when carrying out a graphic design task. Related work Techniques of style transfer have found extensive use in the field of graphic design and have been used to enhance design efficiency, allow visual exploration of a variety of styles, and personalization and visual variety 4 – 6 . Initial research concentrated more on the transfer of low-level visual attributes like color and texture, which were useful in the transformation of artistic images but were generally insensitive to layout structure, semantic consistency, and aesthetic design-oriented features. In order to deal with the visual expressiveness of graphic visualization, Shi Yang et al. introduced Vistylist that isolates graphic visualization content and style so that the visual image can be rendered in a faithful and expressive way 7 . Although this method works well in pictorial visualization, it makes use of visual mappings that are predefined and does not directly optimize aesthetic goodness in generation. Neural style transfer has been studied as a support tool for digital art and graphic design. Ye et al. 8 presented a multi-semantic preserving style transfer system that relies on the Y-channel information to guarantee the semantic consistency of the stylized outputs. In spite of the fact that this approach enhances the preservation of semantics, it involves more semantic description networks and thus, more complexity of the system regarding scalability in the real design context. Equally, Zhang et al. 9 integrated style transfer with deep residual adaptive networks in multimedia communication and urban graphic design applications. Although these approaches prove to be highly applicable, none of them pays much attention to artistic assessment and optimization, focusing mostly on structural changes and content conservation. Neural style transfer was applied to the generation of hybrid camouflage patterns by Daneshvar et al. 10 , which allows the manipulation of features between patterns, but it is task-specific and not applicable to a wider range of graphic design situations that necessitate style variation as well as aesthetic consistency across patterns. Over the past few years, generative adversarial networks (GANs) have demonstrated a high potential in image generation and style transfer because of their high-generative potential and flexibility 11 , 12 . Applications of GAN-based approaches to dynamic text generation in art and fashion transfer have been successfully done. As an example, Yang et al. 13 introduced Shape-Matching GAN++ to produce quality and controllable artistic text animation, and Liu et al. 14 presented SwapGAN, which can be used to transfer fashion style in multiple steps. The above studies reveal that GANs can effectively process more complicated style changes and increase visual realism. Nevertheless, current GAN-based style transfer systems are typically preoccupied with visual fidelity and diversity, and aesthetic quality in their approach is an implicit or post hoc attribute, and not an explicit optimization goal. Besides, there are still problems with the stability of content and aesthetic consistency in cross-domain and design-driven case scenarios. Overall, despite the achievements made in the field of style transfer and GAN-based image generation in the past, the majority of the existing models do not include a coherent way of ensuring that content, style, and aesthetics are mutually optimized, especially when working with graphic design tasks with visual appeal as a major factor. The outlined restrictions can be seen as motivation behind the given approach, which ensures that an aesthetic scoring model is incorporated into the course of GAN training and that a number of loss components are minimally optimized to attain a stable and aesthetically improved transmission of the style of graphic design. Graphic design style transfer and aesthetic optimization algorithm Graphic design style transfer model of an improved generative adversarial network Dataset construction In order to facilitate style transfer based on the graphic design orientation, a special dataset of images in the field of graphic design was created using the freely available design platforms and professionally-oriented design resources. The dataset mainly comprises layout-based graphic designs and not natural photographs, which are of various visual styles, including typography arrangements, color schemes, compositional layouts, iconographic, and graphics that are brand-focused. All photos were selected with the intention of highlighting the design-specific visual properties, and all were resized to a standardized resolution. Such images were used to train the GAN to acquire representative features of graphic design styles. Sample data of the dataset is represented in Fig. 1 . Fig. 1. Open in a new tab The diversity of layout structures, color compositions, typography, and graphic elements used for training the proposed model. Improved generative adversarial network To clarify the combination of classical neural style transfer and adversarial learning, we note that the proposed system is a hybrid GAN-based image translator in which the generator produces the stylized output and is adversarially trained with a discriminator. The iterative image synthesis does not make use of the pretrained VGG network, but only the fixed perceptual feature extractor is used when computing content and style supervision terms as part of the unified loss function. The structure has three fundamental elements, namely, the generator , a discriminator , and a pretrained VGG feature network . The generator is based on encoder decoder architecture with improved inverted block residuals and channel attention to be learned between the content image and style reference and the stylized output image. Loss of perceptual contents maintains semantic structure, whereas style and color consistency losses balance the statistics of texture and the distribution of chromas. Both the real target style images and the generated images are provided to the discriminator, which is trained to treat the new image as different from the real one, thus imposing a visual realism and domain consistency. Adversarial loss with perceptual and aesthetic constraints is used to optimize the generator and discriminator alternately, creating a single objective as opposed to a distinct NST optimization process. Fig. 2 shows the general structure of the proposed framework for transferring graphic design styles. Fig. 2. Open in a new tab Graphic design style transfer model of an improved generative adversarial network. The entire training process of the proposed hybrid adversarial system is outlined in Algorithm 1, which covers forward propagation, computation of the multi-term loss, and alternating optimization of the generator and discriminator. Open in a new tab Algorithm 1 Hybrid GAN training procedure Generator model The generator model consists of three parts: feature encoding, feature decoding, and feature conversion 15 . The feature encoding module is responsible for capturing the key features of the input image and reducing its dimensionality, while the feature decoding module maps abstract features back to specific features and gradually restores the original image size. Regarding feature conversion, the use of reverse residual blocks is not conducive to the model’s extraction of image texture, so this problem needs to be solved. The reverse residual block is improved to capture global and local information more effectively and allow the output features to incorporate diverse information. The convolved features are divided into four branches in the channel dimension. The branch uses 1×1 convolution to extract features and directly transmits them to . After undergoes a 3×3 depth convolution, the features are passed to and simultaneously. and also each undergo a layer of 3×3 depth convolution. features are passed to and simultaneously, while features are passed directly to . This method uses a deep convolution method to effectively protect the bottleneck structure of the residual without changing the number of channels. To solve the problem of significant differences between samples during the style transfer process, this article treats instance normalization as a function of depth convolution normalization. Splitting feature maps along the channel dimension leads to the isolation of channel information, affecting model generalization and robustness. Therefore, channel shuffling is used after feature fusion, and the feature map is divided into 512 groups according to the channel dimension. The channels within the groups are randomly sorted and then merged to form a new feature map 16 – 18 . This method does not change the dimension of the feature map, but promotes the integration of channel information, enhances the model’s adaptability to different samples and scenarios, and reduces feature bias and over-fitting problems. Figure 3 illustrates the general scheme of the offered generator, which is an encoder-decoder in design. The multi-level content representations are extracted and aggregated by the encoder, and the middle transformation phase employs enhanced inverted residual blocks with an effective channel attention mechanism to reinforce aesthetic feature encodings. The decoder further reassembles the output gradually to create the final stylized image. With this design style, one can effectively transform the style without distorting the content’s structure and enhancing aesthetic consistency. To enhance the expressiveness of style features, an efficient channel attention (ECA) mechanism is applied. ECA is a cross-channel interaction method that does not require dimensionality reduction and can dynamically adjust the channel weights on the feature map to focus more on essential feature channels 19 , 20 . In the inverted residual block, since the output of the deep convolution contains rich high-level semantic features, the ECA mechanism is added to the deep convolution, so that the size of the convolution and the number of channels of the input feature have a non-linear relationship: 1 where represents the number of channels of input features, represents is the kernel size of the one-dimensional convolution within the ECA module and which is odd, should be the closest value of an even number to this value, so that it offers symmetric convolution. The mapping slope and the mapping offset are controlled by the parameters and , respectively, and are set empirically to and as is customary when modeling attention mechanisms based on ECA. Fig. 3. Open in a new tab Architecture of the proposed generator network for graphic design style transfer. Discriminator model To balance the discriminator’s and generator’s capabilities, a spectral normalization layer is added after the convolutional layer of the discriminator, and noise set by the mean value of the style data set is added to the generated image. Figure 4 shows the loss curves of the generator and discriminator. Fig. 4. Open in a new tab Loss curves of generator and discriminator. Figure 4 shows that the horizontal axis is the number of iterations (unit is ), and the vertical axis is the loss value, representing the generator’s and discriminator’s performance in training. The loss of the initial generator fluctuates around 50, indicating that the quality of the samples generated in the initial stage is not high, and it is difficult to deceive the discriminator. However, as the number of iterations increases, the generator’s loss gradually decreases and stabilizes after about 30000 iterations, indicating that its performance is improving and the generated samples are closer to reality. The initial maximum value of the discriminator’s initial loss is around 25, which shows that it is difficult for the discriminator to distinguish between real and generated samples at the beginning. However, as the training progresses, its loss gradually decreases and is basically stable within 10 after about 25000 iterations. The discriminator receives as input the generated stylized image and real images sampled from the target graphic design style domain. It is trained to distinguish real target style images from generated outputs, thereby enforcing visual realism and style consistency. The discriminator follows a convolutional classifier design with spectral normalization for stable optimization, producing a realness prediction for each input image. Image style transfer Convolutional neural network (CNN) is a forward neural network that combines convolution operations and deep structures, characterized by local connectivity, weight sharing, and pooling layer downsampling 21 – 23 . The VGG-19 model uses the ReLU (Linear rectification function) to train the neurons of the fully connected layer and input them to the output end 24 , 25 . The VGG-19 model is based on the image dataset and extracts image features from the image data by pre-training. To solve the problem of fewer style pictures and content pictures in actual application scenarios, the VGG-19 network structure is optimized, and the global average pooling layer is utilized to replace the fully connected layer to reduce training parameters. The parameters obtained by training are used to extract images, thereby improving training time and achieving better generation results. The implementation process of graphic design image style transfer is illustrated in Fig. 5 : Fig. 5. Open in a new tab Implementation process of graphic design image style transfer. In the graphic design image style transfer process, the pre-trained VGG-19 model is used to extract features from the input and target style images. The input image undergoes multiple convolutions and pooling to generate a feature map containing details such as edges and textures. To achieve style transfer, the content features of the input image are retained and combined with the target style features. The Gram matrix technology is used to measure the similarity between the two images, and the best way to apply the target style to the original image is found by minimizing the Gram matrix distance between them. Finally, a new image that combines the original image content and the target style is generated. Loss function design To prevent a black box formulation as well as achieve the optimization fully reproducible, we introduce a single unified training purpose where perceptual supervision, adversarial learning, and aesthetic guidance are being optimized together. In our model, the generator results in a stylized output of a content image Icand a style reference . A fixed perceptual feature extractor is a pretrained VGG 19 network, and a differentiable aesthetic signal is a pretrained aesthetic scoring network which is back propagated to the generator. Let is the activation map of the anchored VGG layer selected at position . The content loss maintains the high-level structure of the content image by reducing the square difference between feature responses of the generated and content images: 2 In order to pass the style statistics, we match Gram matrix correlations between the created image and the style reference. Let be the Gram matrix computed from the VGG feature map. The style loss is defined as: 3 and here are the number of channels in the feature map and the spatial size of the feature map of layer , and the weights of any given layer are denoted by . In order to promote the similarity of color statistics between the created image and the desired style distribution, we use an RGB channel pixel mean consistency loss. The RGB channels are flattened, sorted, and broken down into regions, and the means of the regions are calculated. where and denote the mean of the -th region for channel n the generated image and style image, respectively: 4 5 Let denote the aesthetic scoring network. Because is differentiable, its output provides a gradient signal that is back propagated through to update the generator parameters. The aesthetic loss is defined as: 6 where the aesthetic network is held fixed during training to provide fixed supervision, and the only gradient updates are provided to the generator using this term only. We learn both an adversarial loss through discriminator feedback in order to achieve realism and domain consistency, and a reconstruction term to achieve content level fidelity. All the terms of loss are gathered into one generator objective: 7 In our experiments, the coefficients are set to , , , and , while is reported in the implementation settings. All coefficients are treated as static during training to maintain reproducibility and to ensure consistent balancing between realism, content preservation, style strength, and aesthetic quality. Aesthetic scoring model The traditional GAN does not consider the aesthetic properties of the generated images during the training process, resulting in the generated graphic designs not being beautiful enough. To this end, this article adopts an advanced image aesthetic scoring model, based on the obtained graphic design scoring results, to establish a reasonable loss function to guide the generation of the GAN network. The image aesthetic scoring model is built based on CNN and is used to predict the image aesthetic score. Its system structure is shown in Fig. 6 . Fig. 6. Open in a new tab System structure of the aesthetic scoring model. The whole training process can be divided into three stages: taking AlexNet as the basic network and fine-tuning it. AlexNet is an image-based classification network. More than 1000 classifiers FC8 in AlexNet are removed, and the Euclidean loss regression model is used to replace the Softmax classification layer to achieve the transformation from classification to prediction 26 , 27 . After fine pre-training of AlexNet, its performance is better than that of direct training. The formula of Euclidean loss can be expressed as: 8 The model trained only by minimizing the Euclidean loss does not perform well for images with similar beauty scores. However, a more detailed aesthetic evaluation is particularly important for experiments. To improve the reliability of sorting, a model training method based on the Siamese network is adopted. The two photos are represented by and respectively, the real aesthetic scores of and are represented by and , and the predicted scores of the network model are represented by and . Then, the formula for calculating the pairwise sorting loss is: 9 Among them: - the prescribed boundary parameter The function relationship can be expressed as: 10 In terms of aesthetic prediction, using image attribute labels for aesthetic evaluation has achieved good results. To this end, based on existing research, this article adds an additional activation layer to the network to enable it to encode image attribute information. The specific implementation method is to add an attribute prediction branch based on the second fully connected layer of the aesthetic network described above. The attribute loss is expressed using the K-path Softmax loss, and combined with the scoring and ranking losses: 11 Among them: -the weight value of the attribute loss In the case of insufficient attribute labels, the branch of this attribute can be locally frozen, and only the other parts can be fine-tuned. There are different aesthetic concepts for different images. To this end, while aesthetically grading the image, the classification information is incorporated into the model to achieve comprehensive optimization and prediction of the image. Meanwhile, the image category is judged based on the image content and aesthetic evaluation is performed. On this basis, the Softmax technology is used to fine-tune the first two layers of AlexNet, train the content-oriented branches, and achieve the prediction of classification labels. In the convergence stage, the Softmax output based on the content class prediction result is utilized as the weight vector to weight the scores generated by each feature branch (aesthetic branch, attribute branch, content branch). Finally, the base network and the output branch are fine-tuned together, and the content classification branch is fixed. Combining the attribute and category information, the network learns the aesthetic features and outputs the score of each image. After the training of the aesthetic scoring model is completed, the generated image is input into the model to maximize the aesthetic score of the generated image to assist the GAN network in training. In this way, the aesthetic loss can be expressed as: 12 Unlike other methods that use aesthetic attributes as input image aesthetic quality scoring methods, the aesthetic attribute reference layer in the network of this article shares the feature extraction structure with the score prediction network, and there is no need to train the classifier separately. Although the aesthetic scoring model is grounded on a traditional CNN architecture, what is new about it is the multi-objective formulation and its practical application in the style transfer framework. Using regression loss, pairwise ranking loss, and attribute-sensitive learning via a collaboration-based approach, and incorporating the resulting aesthetic score as a feedback mechanism throughout the training of the GAN, the model can offer effective aesthetic guidance specific to generation of graphic design. Training process The Adam algorithm is utilized to optimize the parameters of the neural network. The exponential decay coefficient of the first-order moment estimate is beta1=0.5; the exponential decay coefficient of the second-order moment estimate is beta 2=0.999; the learning rate reaches 0.0001. The Adam algorithm is an algorithm that integrates adaptive learning rate and momentum, and is derived from various optimal algorithms 28 , 29 . The Adam algorithm uses the gradient method to establish two vectors n and W, where vector n is the momentum term and vector W constructs the adaptive learning rate by accumulating the square sum of the gradients 30 , 31 . The initial values of vectors n and W are 0. The update formula is: 13 J-the vector component subscript, relying on these two values to construct the updated value of the parameter. The vectors are subscripted, and the updated parameters are built based on these two values. The parameter update formula is: 14 The Adam algorithm is an effective method that combines the first-order gradient moment with the non-central second-order moment, and has been widely used in neural network optimization 5 , 6 , 32 . The current method for neural network optimization usually uses gradient descent instead of Newton’s method. Effect verification of style transfer and aesthetic optimization Experimental design To evaluate the actual application effect of the improved generative adversarial network in graphic design style transfer and aesthetic optimization, this article takes interior graphic design drawings as experimental objects, collects interior graphic design drawings from graphic design websites as datasets, and preprocesses the images to ensure uniform format and size. Key features such as color, contrast, brightness, texture and composition rules are selected, and the selected features are quantified using OpenCV Library (Open Source Computer Vision Library) and custom scripts. Afterwards, 9 graphic design pictures are selected from the collected dataset as the main objects of this article’s test, and hybrid feature extraction, image saliency network Pool Net, and CNN are used as baseline methods. The content retention, style similarity, and image clarity of these three methods are compared with the method of this article, so as to evaluate the style transfer and aesthetic optimization effects of the model. The experimental environment of this article is displayed in Table 1 : Table 1. Experimental environment. Serial number Environment Name Configuration 1 Hardware environment Computing architecture Pytorch 1.11.0 2 Central processing unit Intel Xeon (R) 3 Graphics processing unit NVIDIA GTX1050 4 Memory 128GB 5 Software environment Operating system Windows 10 6 Compiler Jupyter, PyCharm 7 Python 3.6 Open in a new tab This experiment sets corresponding weights for different loss functions: the weight of the image reconstruction loss function is 10.0; the weight of the content encoding loss and style encoding loss is both 1.0; the weight of the adversarial loss function is also set to 1.0; the weight of the cycle consistency loss function is set to 10.0. Such weight configuration aims to balance the impact of each loss function to optimize the model’s overall performance. Experimental results Content retention Content retention is the key to evaluating the effect of style transfer. High content retention indicates that the image style is unique and attractive, and can accurately convey the original image information to meet practical and aesthetic needs. Generally speaking, content retention is usually measured by the structural similarity index. To this end, the hybrid feature extraction, image saliency network (Pool Net), convolutional neural network and SSIM under this method are compared. The results are illustrated in Fig. 7 : Fig. 7. Open in a new tab SSIM comparison under four methods. As can be seen from Fig. 7 , the method in this article has obtained the highest SSIM values on all pictures. The SSIMs of pictures 1 to 9 are 0.91, 0.93, 0.95, 0.94, 0.91, 0.95, 0.95, 0.90 and 0.96 respectively, and the average SSIM is 0.93, significantly higher than other methods. In contrast, although the hybrid feature extraction and Pool Net methods can also maintain good content retention, the SSIM values are generally lower than the method in this article. The SSIM value of the hybrid feature extraction ranges from 0.83 to 0.87, with an average SSIM of 0.85; the SSIM value of the Pool Net fluctuates between 0.80 and 0.88, with an average SSIM of 0.84. The traditional CNN method performs relatively poorly on this indicator. The lowest SSIM value is only 0.71; the highest is only 0.80; the average SSIM is 0.75. It can be seen from this that the method in this article not only has significant advantages in content retention, but also ensures the aesthetics and practicality of style transfer, providing a new solution for graphic design style transfer. In addition to SSIM, the mean square error (MSE) is also an important indicator for measuring content retention. This article also compares the MSE values under the four methods. The results are displayed in Table 2 : Table 2. Comparison of MSE values under 4 methods. Picture serial number Hybrid feature extraction Pool net CNN This article 1 0.047 0.053 0.075 0.030 2 0.064 0.067 0.077 0.029 3 0.067 0.073 0.076 0.024 4 0.049 0.058 0.077 0.023 5 0.059 0.070 0.070 0.023 6 0.060 0.059 0.083 0.030 7 0.056 0.055 0.082 0.029 8 0.046 0.054 0.068 0.022 9 0.057 0.059 0.067 0.031 Open in a new tab According to the data in Table 2 , the MSE values of this method on pictures 1 to 9 are 0.030, 0.029, 0.024, 0.023, 0.023, 0.030, 0.029, 0.022 and 0.031 respectively. The average MSE value is 0.027, which is significantly lower than other methods. The MSE value of the hybrid feature extraction method ranges from 0.046 to 0.067, with an average MSE value of 0.056; the Pool Net method ranges from 0.053 to 0.073, with an average MSE value of 0.061; the traditional CNN method ranges from 0.067 to 0.083, with an average MSE value of 0.075. This shows that the method in this article can better maintain the content and structure of the original image during the image generation process, and generate images with minimal differences from the original image. This is mainly due to the method in this article that combines the excellent generation ability of GANs and the feature extraction ability of the VGG network. It ensures the uniqueness of style and the accuracy of information transmission by optimizing content and style loss. In addition, applying an efficient channel attention mechanism and an improved inversion residual block further improves the model’s information capture capability, thereby significantly improving content retention. It can be seen that the method in this article has excellent performance in content retention and provides strong support for design style transfer and aesthetic optimization. Style similarity Style similarity is the key to evaluating the effect of style transfer, ensuring that the generated image is consistent with the target style and meets aesthetic needs. Style similarity is usually compared using MSE. To this end, this article calculates the difference between generated images and style images under 4 methods. The results are illustrated in Fig. 8 : Fig. 8. Open in a new tab Comparison of MSE results of style similarity between generated images and style images. As can be seen from Fig. 8 , the difference between the generated image and the style image under this method is minimal. For picture 1, the MSE value of this method is only 0.017; the hybrid feature extraction method is 0.071; the Pool Net method is 0.069; the CNN method is 0.052. The proposed method is much lower than the other three methods, showing a clear advantage in maintaining style consistency. For picture 2, the MSE value of this method is 0.013, while the MSE of hybrid feature extraction, Pool Net method and CNN method are 0.074, 0.070 and 0.058, respectively. For picture 6, the MSE value of this method is 0.024, which is still better than the 0.080, 0.066 and 0.063 of other methods. Overall, the average MSE value of the image and style image generated by this method is 0.020; the average MSE values of the hybrid feature extraction, Pool Net method and CNN method are 0.071, 0.067, and 0.058, respectively. This means that the style similarity generated by the image generated by this method is the highest. This is because this method uses the powerful generation ability of GAN and precise control of details, and strengthens the capture of style features through a cleverly designed loss function. Subsequently, this article also tests the aesthetic evaluation model under different methods. The prediction accuracy of the evaluation model is evaluated by testing on images, and the error between the actual and predicted scores is evaluated by MSE. The results are shown in Table 3 : Table 3. Comparison of actual scoring and predicted scoring errors under different methods. Picture serial number Hybrid feature extraction Pool net CNN This article 1 0.036 0.046 0.042 0.024 2 0.037 0.043 0.032 0.018 3 0.039 0.057 0.039 0.017 4 0.035 0.051 0.038 0.024 5 0.046 0.053 0.048 0.024 6 0.054 0.040 0.038 0.018 7 0.048 0.055 0.049 0.022 8 0.039 0.048 0.045 0.026 9 0.043 0.042 0.041 0.023 Open in a new tab As shown in Table 3 , the MSE values of the actual scores and predicted scores of this method on all test images are significantly lower than those of other methods, showing a significant advantage in the prediction accuracy of the aesthetic evaluation model. From picture 1 to picture 9, the MSE value of this method ranges from 0.017 to 0.026. Compared with the hybrid feature extraction method (0.035 to 0.054), the Pool Net (0.040 to 0.057), and the CNN (0.032 to 0.049), this method yields a smaller prediction error. This means that the prediction accuracy of the aesthetic evaluation model under this method is significantly better than that of other models. To supplement the quantitative analysis, visual comparisons are provided to provide a qualitative assessment of the performance of various approaches. The comparison between the suggested method and the baseline algorithms, such as hybrid feature extraction, PoolNet, and CNN-based style transfer, is given side-by-side in Fig. 9 . The comparison is based on the content conservation, the consistency of the style, and the quality of the whole aesthetics. As demonstrated in the figure, the suggested approach is more structurally sound and yields more consistent style transfer and more appealing outcomes than the baseline algorithms, which are marred by artifacts such as content distortion, style inconsistency, or lack of visual clarity. Fig. 9. Open in a new tab Visual comparison of graphic design style transfer results across different methods. Image clarity Image clarity is crucial for evaluating the effect of style transfer, which ensures that the generated image has high resolution and clear details, improving the viewing and practicality of the work. High definition means that the image retains more details and avoids image blur and distortion. This article uses Peak Signal to Noise Ratio (PSNR) and SSIM to evaluate image clarity. The results are shown in Fig. 10 : Fig. 10. Open in a new tab PSNR and SSIM results under four methods ( A ): PSNR results under four methods; ( B ): SSIM results under four methods. As can be learned from Fig. 10 (A) and Fig. 10 (B), the method proposed in this article is significantly better than other methods in both PSNR and SSIM indicators. For picture 1, the PSNR value of this method is as high as 35.65, which is far better than the 26.91, 32.14, and 22.31 of the hybrid feature extraction, Pool Net method, and CNN method. For picture 6, the PSNR value of this method also reaches the highest 36.69, which is significantly improved compared to 29.82, 28.69 and 22.28 of other methods. Overall, the average PSNR of this method is 34.33; the average PSNR of hybrid feature extraction, Pool Net method and CNN method are 28.59, 30.21, and 23.33 respectively. This shows that the method in this article has significant advantages in maintaining image high resolution and detail clarity. On the SSIM index, the performance of this method is also better. For pictures 2 and 7, the SSIM values of the method in this article reach 0.94 and 0.89 respectively, and the overall average SSIM value is 0.91; the SSIM values of the hybrid feature extraction method are 0.81 and 0.84 respectively, and the overall average SSIM value is 0.83; the SSIM values of the Pool Net method are both 0.86, and the overall average SSIM value is 0.84; the SSIM values of the CNN method are 0.77 and 0.78 respectively, and the overall average SSIM value is 0.75. This shows that the images generated by the method in this article are highly consistent with the original images in structure and content. This is because the method in this article achieves high-resolution, detailed image generation by optimizing the loss function, and gives the images higher visual quality and aesthetic effects. Finally, this article also evaluates the diversity of generated images during style transfer. Diversity is the key to evaluating the effectiveness of style transfer, reflecting the diversity and creativity of the generated images. High diversity avoids single duplication and enhances design innovation. By evaluating diversity, it can be ensured that the algorithm generates high-quality images and provides a variety of styles to meet design needs. To this end, this article uses Euclidean distance and cosine similarity average to evaluate the feature diversity between image sets generated by four methods. The results are described in Table 4 : Table 4. Average values of euclidean distance and cosine similarity. Serial number Methods Euclidean distance average Average cosine similarity 1 Hybrid feature extraction 13.45 0.68 2 Pool Net 16.13 0.83 3 CNN 15.88 0.77 4 This article 20.13 0.59 Open in a new tab As shown in Table 4 , the method in this article achieves excellent performance across all indicators, with an average Euclidean distance of 20.13, far exceeding other methods, indicating that the generated images exhibit significant differences in feature space and promote diversity. The average cosine similarity under this method is only 0.59, further proving that the direction difference between images is significant and the creativity is strong. In contrast, the hybrid feature extraction method, Pool Net, and CNN methods are not as good as this method. The Euclidean distance and cosine similarity of the hybrid feature extraction method are 13.45 and 0.68, respectively, indicating that the diversity of its generated images is limited. Pool Net performs reasonably well in Euclidean distance, but the cosine similarity is as high as 0.83, indicating insufficient image direction difference. The CNN method is also inferior to this one, meaning the images generated by this method are more diverse and more creative. Table 5 demonstrates that the proposed technique has the highest SSIM and PSNR values, but by far lowers the LPIPS, which means that the structural preservation is better and the perceptual similarity. Moreover, the aesthetic score proves that the combination of adversarial and aesthetic supervision results in high visual quality in contrast to current SOTA approaches. Table 5. Quantitative comparison with current SOTA style transfer methods on the graphic design dataset. Method SSIM ↑ PSNR (dB) ↑ LPIPS ↓ Aesthetic score ↑ AdaIN 33 0.742 21.36 0.214 6.18 StarGAN-V2 34 0.781 22.95 0.187 6.47 StyleGAN-transfer 35 0.796 23.41 0.173 6.58 Proposed method 0.832 25.67 0.141 7.12 Open in a new tab Conclusions This study uses GAN as the basis for an improved GAN model that combines content retention and aesthetic loss to enhance generation quality, and adds a random generation module to promote innovation. This article trains a unique library of graphic design images through deep learning to capture and transfer image styles. The findings indicate that this approach excels in terms of aesthetics, diversity, and stylistic coherence. Compared with traditional methods, this method is better in content retention, style similarity, and image clarity. Especially in terms of content retention, the method in this article achieves the highest SSIM and the lowest MSE, and it performs well in preserving the content and structure of the original image. In addition, the method in this article also has higher style similarity and clarity, proving its ability to optimize aesthetics. The research in this article brings a new direction to graphic design, which not only enhances design efficiency but also stimulates innovation and better meets the diverse needs of customers. This method cleverly combines the generative power of GAN and the feature extraction capability of the VGG network, and generates high-resolution, detailed images by optimizing the loss function. However, research still has limitations. Additional support is needed when processing complex semantics, and information retention needs to be improved. Author contributions Conceptualization,B.D.; methodology,L.R.; investigation,T.F.; resources,C.T.; writing—original draft preparation,B.D.; writing—review and editing,L.R.; supervision,L.R.; project administration,T.C.; funding acquisition,C.T. All authors have read and agreed to the published version of the manuscript. Funding This work was supported by the General Research Project of Humanities and Social Sciences of the Ministry of Education in 2023(NO.23YJC890029); the Project of Hunan Social Science Achievement Evaluation Committee in 2021,2022 (NO: XSP21YBC046, XSP22YBZ027); Teaching Reform Research Project of Hunan Province in HNJG-2023(0585). Data Availability The data in the manuscript can be obtained by contacting the corresponding author upon reasonable request. Declarations Competing of interest The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References 1. Zhang, H., Sindagi, V. & Patel, V. M. Image de-raining using a conditional generative adversarial network. IEEE Trans. Circuits Syst. Video Technol. 30 (11), 3943–3956. 10.1109/TCSVT.2019.2920407 (2019). [ Google Scholar ] 2. Alotaibi, A. Deep generative adversarial networks for image-to-image translation: a review. Symmetry 12 (10), 1705. 10.3390/sym12101705 (2020). [ Google Scholar ] 3. Liu, M.-Y., Huang, X., Yu, J., Wang, T.-C. & Mallya, A. Generative adversarial networks for image and video synthesis: algorithms and applications. Proc. IEEE 109 (5), 839–862. 10.1109/JPROC.2021.3049196 (2021). [ Google Scholar ] 4. Feng, W., Zhu, Y., Xiao, J., Duan, J. & Zhou, W. Research on style transfer of modern home design images for interior decoration. Comput. Appl. Software 37 (7), 170–175 (2020). 5. Wang, Q. et al. Full waveform inversion based on Adam algorithm with optimized parameters. Chin. J. Geophys. 66 (11), 4654–4663. 10.6038/cjg2022Q0879 (2023). [ Google Scholar ] 6. Wang, T., Ma, Z., Zhang, F. & Yang, L. Research on wickerwork patterns creative design and development based on style transfer technology. Appl. Sci. 13 (3), 1553. 10.3390/app13031553 (2023). [ Google Scholar ] 7. Shi, Y., Liu, P., Chen, S., Sun, M. & Cao, N. Supporting expressive and faithful pictorial visualization design with visual style transfer. IEEE Trans. Vis. Comput. Graph. 29 (1), 236–246. 10.1109/TVCG.2022.3209486 (2022). [ DOI ] [ PubMed ] [ Google Scholar ] 8. Ye, W., Zhu, X. & Liu, Y. Multi-semantic preserving neural style transfer based on y channel information of image. Vis. Comput. 39 (2), 609–623. 10.1007/s00371-021-02361-6 (2023). [ Google Scholar ] 9. Zhang, Z., Wei, F., Liang, G. & Wang, X. Graphic style transfer technology in “multimedia communication”: application of deep residual adaptive networks in graphic design. Nat. J. Antennas Propag. 6 (3), 109–117. 10.31838/NJAP/06.03.14 (2024). [ Google Scholar ] 10. Daneshvar, E., Tehran, M. A. & Zhang, Y. J. Hybrid camouflage pattern generation using neural style transfer method. Color Res. Appl. 47 (4), 878–891. 10.1002/col.22767 (2022). [ Google Scholar ] 11. Qian, W., Li, H. & Mu, H. Circular LBP prior-based enhanced GAN for image style transfer. Int. J. Semant. Web. Inf. Sys. (IJSWIS) 18 (2), 1–15. 10.4018/IJSWIS.315601 (2022). [ Google Scholar ] 12. Zhou, Y. et al. Speckle noise reduction for OCT images based on image style transfer and conditional GAN. IEEE J. Biomed. Health Inform. 26 (1), 139–150. 10.1109/JBHI.2021.3074852 (2021). [ DOI ] [ PubMed ] [ Google Scholar ] 13. Yang, S., Wang, Z. & Liu, J. Shape-matching GAN++: scale controllable dynamic artistic text style transfer. IEEE Trans. Pattern Anal. Mach. Intell. 44 (7), 3807–3820. 10.1109/TPAMI.2021.3055211 (2021). [ DOI ] [ PubMed ] [ Google Scholar ] 14. Liu, Y., Chen, W., Liu, L. & Lew, M. S. SwapGAN: a multistage generative approach for person-to-person fashion style transfer. IEEE Trans. Multimed. 21 (9), 2209–2222. 10.1109/TMM.2019.2897897 (2019). [ Google Scholar ] 15. Hu, M., Zuo, X. & Liu, J. A review of deep generative models. Acta Autom. Sinica 48 (1), 40–74. 10.16383/j.aas.c190866 (2022). [ Google Scholar ] 16. Jiang, D. & Kim, J. Image retrieval method based on image feature fusion and discrete cosine transform. Appl. Sci. 11 (12), 5701. 10.3390/app11125701 (2021). [ Google Scholar ] 17. Mohan, C. K. M-FFN: multi-scale feature fusion network for image captioning. 10.1007/s10489-022-03463-x (2022). 18. Chen, Y., Xia, R., Yang, K. & Zou, K. MFFN: Image super-resolution via multi-level features fusion network. Vis. Comput. 40 (2), 489–504. 10.1007/s00371-023-02795-0 (2024). [ Google Scholar ] 19. Mishra, S. K., Rai, G., Saha, S. & Bhattacharyya, P. Efficient channel attention based encoder–decoder approach for image captioning in Hindi. Trans. Asian Low-Resour. Language Info. Process. 21 (3), 1–17. 10.1145/3483597 (2021). [ Google Scholar ] 20. Yechuri, S. & Vanambathina, S. A nested u-net with efficient channel attention and d3net for speech enhancement. Circuits Syst. Signal Process. 42 (7), 4051–4071. 10.1007/s00034-023-02300-x (2023). [ Google Scholar ] 21. Dubey, S. R. et al. diffGrad: an optimization method for convolutional neural networks. IEEE Trans. Neural Netw. Learn. Syst. 31 (11), 4500–4511. 10.1109/TNNLS.2019.2955777 (2019). [ DOI ] [ PubMed ] [ Google Scholar ] 22. Krichen, M. Convolutional neural networks: a survey. Computer 12 (8), 151. 10.3390/computers12080151 (2023). [ Google Scholar ] 23. Li, Z., Liu, F., Yang, W., Peng, S. & Zhou, J. A survey of convolutional neural networks: Analysis, applications, and prospects. IEEE Trans. Neural Netw. Learn. Syst. 33 (12), 6999–7019. 10.1109/TNNLS.2021.3084827 (2021). [ DOI ] [ PubMed ] [ Google Scholar ] 24. Manjula, P. & Priya, S. B. An effective network intrusion detection and classification system for securing WSN using VGG-19 and hybrid deep neural network techniques. J. Intell. Fuzzy Syst. 43 (5), 6419–6432. 10.3233/JIFS-220444 (2022). [ Google Scholar ] 25. Habib, N., Hasan, M. M., Reza, M. M. & Rahman, M. M. Ensemble of CheXNet and VGG-19 feature extractor with random forest classifier for pediatric pneumonia detection. SN Comput. Sci. 1 (6), 359. 10.1007/s42979-020-00373-y (2020). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 26. Swapna, M., Sharma, Y. K. & Prasadh, B. Cnn architectures: alex net, le net, vgg, google net, res net. Int. J. Recent Technol. Eng. 8 (6), 953–960. 10.35940/ijrte.F9532.038620 (2020). [ Google Scholar ] 27. Ullah, A., Elahi, H., Sun, Z., Khatoon, A. & Ahmad, I. Comparative analysis of AlexNet, ResNet18 and SqueezeNet with diverse modification and arduous implementation. Arab. J. Sci. Eng. 47 (2), 2397–2417. 10.1007/s13369-021-06182-6 (2022). [ Google Scholar ] 28. Jais, I. K. M. & Ismail, A. R. Adam optimization algorithm for wide and deep neural network. Knowl. Eng. Data Sci. 2 (1), 10. 10.17977/um018v2i12019p41-46 (2019). [ Google Scholar ] 29. Khan, A. H., Cao, X., Li, S., Katsikis, V. N. & Liao, L. BAS-ADAM: an ADAM based approach to improve the performance of beetle antennae search optimizer. IEEE/CAA J. Autom. Sin. 7 (2), 461–471. 10.1109/JAS.2020.1003048 (2020). [ Google Scholar ] 30. Yi, D., Ahn, J. & Ji, S. An effective optimization method for machine learning based on ADAM. Appl. Sci. 10 (3), 1073. 10.3390/app10031073 (2020). [ Google Scholar ] 31. Khaire, U. M. & Dhanalakshmi, R. High-dimensional microarray dataset classification using an improved Adam optimizer (iAdam). J. Ambient. Intell. Humaniz. Comput. 11 (11), 5187–5204. 10.1007/s12652-020-01832-3 (2020). [ Google Scholar ] 32. Li, M., Luo, F., Gu, C., Luo, Y. & Ding, W. Adams algorithm based on adaptive momentum update strategy. J. Univ. Shanghai Sci. Technol. 45 , 112–119 (2023). [ Google Scholar ] 33. Huang, X. & Belongie, S. Arbitrary style transfer in real-time with adaptive instance normalization. In Proc. IEEE international conference on computer vision (2017). 34. Choi, Y., Uh, Y., Yoo, J. & Ha, J.-W. Stargan v2: Diverse image synthesis for multiple domains. In Proc. IEEE/CVF conference on computer vision and pattern recognition , (2020). 35. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J. & Aila, T. Analyzing and improving the image quality of stylegan. In Proc. IEEE/CVF conference on computer vision and pattern recognition , (2020). Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Data Availability Statement The data in the manuscript can be obtained by contacting the corresponding author upon reasonable request. Articles from Scientific Reports are provided here courtesy of Nature Publishing Group ACTIONS View on publisher site PDF (5.1 MB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• Generative AI in Action: Transforming Art, Music, and Design#75005
• Exploring the evolution of generative adversarial network architectures for text to image synthesis a comprehensive review#637749
• EmbeddGAN: A Novel GAN Framework Using an Embedding Network and Gini Distance Correlation#1037191
• The Computational Search for Unity: Synthesis in Generative AI#718241
• Reinforcement Learning and Generative AI: Training Machines to Be Creators#95267