Scholarlink

<p>Artworks convey rich emotional and stylistic information, and jointly predicting artistic style and audience emotion is essential for understanding artistic perception. Traditional methods struggle to capture the abstraction and complexity of art images. To address this, we propose a unified multi-task learning framework integrating Convolutional Neural Networks (CNN) and Vision Transformers (ViT). The CNN module extracts fine-grained local and emotional features, while the ViT models global artistic structures through self-attention. By fusing these complementary representations, the model performs simultaneous style classification and emotion prediction with adaptive loss weighting to enhance collaboration. Experiments on four benchmark datasets—ArtEmis, Affective Image, ADE20K, and Abstract Art—show up to 7.8% higher accuracy and 6.5% improvement in F1-score over state-of-the-art methods, demonstrating the robustness and efficiency of the proposed CNN–ViT multi-task framework.</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

Emotion-style dual prediction: a multi-task deep learning approach for artistic images

  • 期刊论文

  • 发表日期：

    2025年11月14日

查看全文
  • Yao Lu,
  • Dan Shao,
  • Xin Ni,
  • Hanrui Yan,
  • Manxing Shi

摘要

Artworks convey rich emotional and stylistic information, and jointly predicting artistic style and audience emotion is essential for understanding artistic perception. Traditional methods struggle to capture the abstraction and complexity of art images. To address this, we propose a unified multi-task learning framework integrating Convolutional Neural Networks (CNN) and Vision Transformers (ViT). The CNN module extracts fine-grained local and emotional features, while the ViT models global artistic structures through self-attention. By fusing these complementary representations, the model performs simultaneous style classification and emotion prediction with adaptive loss weighting to enhance collaboration. Experiments on four benchmark datasets—ArtEmis, Affective Image, ADE20K, and Abstract Art—show up to 7.8% higher accuracy and 6.5% improvement in F1-score over state-of-the-art methods, demonstrating the robustness and efficiency of the proposed CNN–ViT multi-task framework.