Scholarlink

<p>We address critical limitations in current artificial intelligence (AI) systems for the authentication and attribution of paintings, specifically regarding algorithmic bias in tile averaging and the scarcity of high-fidelity training samples. As an alternative to standard patch-level techniques (such as convolutional neural networks) that independently assign a unitary classification probability to each evaluated patch, we employ attention-based multiple instance learning (MIL), which analyzes tiles collectively. The MIL framework generates a high-dimensional “bag” embedding that represents the stylistic essence of a painting under study. Operating on this high-dimensional representation, the classifier learns complex, nonlinear decision boundaries. To overcome data scarcity, we utilize optimization-based neural style transfer to synthesize comparative training images that more closely mimic the artist’s style, creating surrogate artworks to sharpen the model’s decision boundary. Combined with entropy-based patch sifting, our approach delivers superior classification accuracy and significantly improved statistical reliability across complex attribution tasks.</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

Computational art attribution: integrating attention-based multiple instance learning and style-augmented synthetic data

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月1日

查看全文
  • Steven J. Frank,
  • Andrea M. Frank

摘要

We address critical limitations in current artificial intelligence (AI) systems for the authentication and attribution of paintings, specifically regarding algorithmic bias in tile averaging and the scarcity of high-fidelity training samples. As an alternative to standard patch-level techniques (such as convolutional neural networks) that independently assign a unitary classification probability to each evaluated patch, we employ attention-based multiple instance learning (MIL), which analyzes tiles collectively. The MIL framework generates a high-dimensional “bag” embedding that represents the stylistic essence of a painting under study. Operating on this high-dimensional representation, the classifier learns complex, nonlinear decision boundaries. To overcome data scarcity, we utilize optimization-based neural style transfer to synthesize comparative training images that more closely mimic the artist’s style, creating surrogate artworks to sharpen the model’s decision boundary. Combined with entropy-based patch sifting, our approach delivers superior classification accuracy and significantly improved statistical reliability across complex attribution tasks.