Scholarlink

<p>This study focuses on image style transfer, aiming to generate images with the desired style while preserving the underlying content structure. Existing models face challenges in accurately representing both content and style features. To address this, an integrated method for image style transfer is proposed, utilizing a parallel CNN and Vision Transformer (CaVIT). It combines a Convolutional Neural Network (CNN) with a Vision Transformer (VIT) to achieve enhanced performance. Our method utilizes VGG-19 with residual blocks to encode style features for enhanced refinement. Additionally, the PA-Trans Encoder Layer is introduced, inspired by the Transformer Encoder Layer, to efficiently encode content features while preserving the complete content structure. The fused features are then decoded into stylized images using a CNN decoder. Qualitative and quantitative evaluations demonstrate that our proposed method outperforms existing models, delivering high-quality results.</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

CaVIT: An integrated method for image style transfer using parallel CNN and vision transformer

  • 期刊论文

  • 发表日期：

    2025年1月13日

查看全文
  • ZaiFang Zhang,
  • ShunLu Lu,
  • Qing Guo,
  • Nan Gao,
  • YuXiao Yang

摘要

This study focuses on image style transfer, aiming to generate images with the desired style while preserving the underlying content structure. Existing models face challenges in accurately representing both content and style features. To address this, an integrated method for image style transfer is proposed, utilizing a parallel CNN and Vision Transformer (CaVIT). It combines a Convolutional Neural Network (CNN) with a Vision Transformer (VIT) to achieve enhanced performance. Our method utilizes VGG-19 with residual blocks to encode style features for enhanced refinement. Additionally, the PA-Trans Encoder Layer is introduced, inspired by the Transformer Encoder Layer, to efficiently encode content features while preserving the complete content structure. The fused features are then decoded into stylized images using a CNN decoder. Qualitative and quantitative evaluations demonstrate that our proposed method outperforms existing models, delivering high-quality results.