基于视觉感知和迁移学习的书法鉴别技术研究

The authentication of chinese calligraphy works is an important part in the inheritance and protection for chinese calligraphy. However, traditional authentication, mostly depending on limited personal experience, lacks quantitative and objective evidence, hence can't meet the pressing need of authentication for large quantities of works. This project focuses on the digital form of works and explores the authentication technologies based on visual perception and transfer learning. The main contributions of this project are as follows：(1)According to research and analysis of calligraphy style and personalized calligraphy characteristic, domain knowledge databases are constructed. (2) The digital image preprocessing methods of calligraphy works are studied. (3) In reference to human visual perception mechanisms and on the basis of statistical methods, hidden Markov models are used to build the computable calligraphy style models. Besides, sparse coding models are used to build the computable personalized characteristic models. (4) After discussing the mechanism of calligraphy authentication, the project proposes and builds a two-stage authentication model consisting of calligraphy style verification and personalized characteristic identification. (5) For the calligraphists with scarcity works left, authentication of their works is performed by using transfer learning method. This project aims to establish a universal calligraphy authentication system. Its structure and relations of different parts are investigated. Meanwhile the related technologies are studied and implemented. All the research results provide new ideas and practical support for the development of large-scale calligraphy authentication.

书法作品鉴别是书法传承和保护中的重要环节。 然而，传统鉴别只局限于有限专家的个人经验，缺乏客观量化的实证，难以应对时下大批量作品辨真识伪的迫切需求。本课题以数字图像形式的书法作品为对象，展开基于视觉感知和迁移学习的书法鉴别技术研究，具体内容包括：（1）研究书法风格知识和书家个性知识，建立领域知识库；（2）研究书法作品图像的预处理技术；（3）遵循人眼视觉感知机理，基于统计学方法，用隐马尔科夫模型建立可计算的书法风格模型，用稀疏编码算法建立可计算的书家个性模型；（4）阐明书法鉴别的内部机理，建立书法鉴别的阶段模型：书法风格鉴别和书家个性鉴别；（5）采用迁移学习方法，研究基于小样本或稀缺样本的名家书法鉴别问题。本课题以建立通用的书法鉴别模型入手，梳理书法鉴别的层次结构和关系，研究具体实现技术，旨在为大规模书法鉴别的发展提供新思路和技术实践支持。

书法作品鉴别是书法传承和保护中的重要环节。 然而传统鉴别只局限于有限专家的个人经验，缺乏客观量化的实证，难以应对时下大批量作品辨真识伪的迫切需求。项目以数字图像形式的中国书法作品为对象，以书法作品的艺术要素：形质和神采为出发点开展自动鉴别技术的研究。项目以8类书法风格的作品为对象进行分析，建立了书法风格知识库和书家个性知识库；实现了碑类和帖类作品的自动识别；针对低质量的碑类图像，提出了基于多滤波器和联通区域的去噪增强算法。基于书法作品的审美要素，项目提出了基于多通道信息和引导滤波的作品信息的提取方法，并以5层次分解模型为依据，建立了可计算的书法风格模型；提出了分区引导滤波的笔画信息提取方法，以8类笔画基元为对象建立了可计算的书家个性模型。项目实现了基于支持向量机和稀疏编码的书法风格鉴别和书家个性鉴别，并利用迁移学习方法初步实现了小样本或稀缺样本的名家书法作品的鉴别。此外，针对书法信息的特殊性，项目进一步研究了评判艺术信息提取效果的标准。