Scholarlink

<p>The creative style transfer, which enables robots to blend creative styles with image information, is an important area of computer vision. Compared to traditional hand-crafted filtering or manual texture blending, deep learning (DL) algorithms achieve superior detail preservation, structural integrity, and consistent stylistic representation. This paper proposes a DL-Based Painting Style Transfer Framework that uses generative models to achieve high-fidelity artistic transformations between different painting styles. The proposed Efficient Tuna Swarm-mutated Adaptive Style Generative Adversarial Network (ETS-AStyleGAN) model guarantees enhanced texture realism and stylistic consistency to generate high-fidelity artistic transformations across a range of painting styles. 2585 high-resolution images from open-source art collections were gathered to construct the dataset, which included both real-world images and vintage paintings for style transfer studies. To preserve structural consistency, preprocessing included image scaling, color normalization, and median filtering to remove noise. Pre-trained VGG18 was used for feature extraction in order to compute Gram matrices and extract both content and style features. While ETSO enhances the generating variables to increase convergence speed, reliability, and efficacy, AStyleGAN uses concentrated artistic style transfer to produce visually consistent and high-fidelity stylized images. Using Python 3.10, the research experimental findings show an ETS-AStyleGAN model LPIPS value of 0.142 and an SSIM value of 0.859 when compared to baseline methods. Results showed that the ETS-AStyleGAN successfully creates aesthetically pleasing images while preserving a high degree of feature coherence between the target and produced outputs. This research offers a reliable and expandable solution for digital creative applications and the creation of artistic images.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deep learning-based painting style transfer and generative models

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月5日

查看全文
  • Qing Li

摘要

The creative style transfer, which enables robots to blend creative styles with image information, is an important area of computer vision. Compared to traditional hand-crafted filtering or manual texture blending, deep learning (DL) algorithms achieve superior detail preservation, structural integrity, and consistent stylistic representation. This paper proposes a DL-Based Painting Style Transfer Framework that uses generative models to achieve high-fidelity artistic transformations between different painting styles. The proposed Efficient Tuna Swarm-mutated Adaptive Style Generative Adversarial Network (ETS-AStyleGAN) model guarantees enhanced texture realism and stylistic consistency to generate high-fidelity artistic transformations across a range of painting styles. 2585 high-resolution images from open-source art collections were gathered to construct the dataset, which included both real-world images and vintage paintings for style transfer studies. To preserve structural consistency, preprocessing included image scaling, color normalization, and median filtering to remove noise. Pre-trained VGG18 was used for feature extraction in order to compute Gram matrices and extract both content and style features. While ETSO enhances the generating variables to increase convergence speed, reliability, and efficacy, AStyleGAN uses concentrated artistic style transfer to produce visually consistent and high-fidelity stylized images. Using Python 3.10, the research experimental findings show an ETS-AStyleGAN model LPIPS value of 0.142 and an SSIM value of 0.859 when compared to baseline methods. Results showed that the ETS-AStyleGAN successfully creates aesthetically pleasing images while preserving a high degree of feature coherence between the target and produced outputs. This research offers a reliable and expandable solution for digital creative applications and the creation of artistic images.