Scholarlink

<p>Intelligent material recognition in traditional Chinese painting images is vital for digital cultural heritage preservation. However, existing methods struggle with limited samples and weak feature representation. This paper proposes a new classification framework based on the prototypical network with a ResNet18 backbone. To enhance sensitivity to fine-grained details, a cropping-based augmentation strategy is introduced during inference. Additionally, a multitask learning scheme is employed to improve generalization and enrich global representations, where the auxiliary task of dynasty classification is trained jointly with the main task of material classification. Furthermore, an ensemble voting strategy is applied to refine the final predictions. Experiments on a self-constructed dataset demonstrate the robustness of the entire strategy under few-shot scenarios, achieving 80% accuracy with merely 30 samples and outperforming comparative methods. Moreover, this work provides practical value for the development of intelligent tools to support the digital preservation of traditional Chinese art.</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

Material classification method of traditional Chinese painting image based on prototypical network

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年7月30日

查看全文
  • Zhibin Su,
  • Luyue Zhang,
  • Jinyu Liu,
  • Shuang Wang

摘要

Intelligent material recognition in traditional Chinese painting images is vital for digital cultural heritage preservation. However, existing methods struggle with limited samples and weak feature representation. This paper proposes a new classification framework based on the prototypical network with a ResNet18 backbone. To enhance sensitivity to fine-grained details, a cropping-based augmentation strategy is introduced during inference. Additionally, a multitask learning scheme is employed to improve generalization and enrich global representations, where the auxiliary task of dynasty classification is trained jointly with the main task of material classification. Furthermore, an ensemble voting strategy is applied to refine the final predictions. Experiments on a self-constructed dataset demonstrate the robustness of the entire strategy under few-shot scenarios, achieving 80% accuracy with merely 30 samples and outperforming comparative methods. Moreover, this work provides practical value for the development of intelligent tools to support the digital preservation of traditional Chinese art.