高级检索

面向中国古画鉴定的多模态可视分析方法

Multi-Modal Visual Analytic Approach for Attributing and Authenticating Ancient Chinese Paintings

  • 摘要: 中国古画的传统鉴定很大程度上依赖于专家的推理论证, 不仅需要考虑题款、印章、画作内容等多方面的信息, 还需大量的文献资料辅助. 基于深度学习的方法提高了传统鉴定工作的效率, 但其主要聚焦于图像分析, 同时深度神经网络的黑盒特性不利于人机高效协作. 为此, 提出了一种多模态语义提取的可视分析方法. 首先将古画图像分割为题款、印章、画作内容 3 类语义对象; 接着通过图像匹配和文本识别构建对应古画之间及与相关文献条目的关联结构并编码; 最后由专家进行古画片段的选取以生成多模态语义网络关系图进行查阅考证; 此外, 设计并实现人机协同驱动的中国古画辅助可视鉴定系统. 通过案例研究和专家评估并以 5 点李克特量表量化专家反馈, 得分范围在 0.78~1.78, 验证了所提方法在古画分析和考证关联中的高效性和准确性.

     

    Abstract: The traditional workflow of authenticating ancient Chinese paintings heavily relies on experts’ reasoning and justification, who need to consider various information, such as inscriptions, seals, and painting content, along with extensive literature materials. Although deep learning-based methods have improved the efficiency of traditional authentication workflow, they predominantly focus on image analysis while the black-box nature of DNNs hampering the effective human-machine collaboration. To address this issue, we propose a visual analysis approach that extract multi-modal semantics to facilitate the easy analysis of ancient Chinese paintings. Firstly, ancient painting images are segmented into three semantic objects: inscriptions, seals, and painting content. Then, an association structure is constructed between the corresponding segments of ancient paintings through image matching and text recognition, as well as with relevant literary references. Finally, experts select segments of ancient paintings to generate a multi-modal semantic network graph for reference and verification. Furthermore, we develop a mixed-initiative system for assisting the visual attribution analysis of ancient Chinese paintings. Through case studies and expert evaluations, the effectiveness and accuracy of ancient painting analysis and authentication were validated using a 5-point Likert scale for expert feedback, with scores ranging from 0.78 to 1.78.

     

/

返回文章
返回