基于大数据挖掘与深度学习的中国书法仿写与智能创作的算法研究

Chinese calligraphy is the gem of Chinese traditional art. The effort to design and produce the machine intelligence for automatic processing, deep analysis, and intelligent creation of artistic images in general, and Chinese calligraphic works in particular, is one of the most challenging grand problems in artificial intelligence research. The proposed research aims to leverage big data mining and deep learning methodologies to investigate a series of intelligent algorithm design, modeling and optimization problems in the visual thinking domain that are essential to the activities of computational imitation and generation of Chinese calligraphic works. At the present big-data era, conducting the aforementioned artificial intelligence research for Chinese calligraphy is also beneficial for protecting and advancing traditional Chinese culture, utilizing the timely opportunities offered by the One Belt One Road initiative of the nation. The key distinguishing feature of this proposed study lies in its integrative approach to jointly utilize the power of classical image processing techniques and the potential of contemporary artificial intelligence in its investigation of multiple key scientific problems, which are concerned with a collection of visual thinking modeling, reasoning, algorithm design and optimization tasks undertaken for facsimileing, imitating, and generating Chinese calligraphy artwork via the computational vehicles of big data mining, analysis, and deep learning. The proposed study will sequentially conduct the following main research tasks, including intelligently identifying and extracting strokes from images of Chinese calligraphic artwork, learning representations for shapes and topological structures in calligraphic images, modeling the beauty of calligraphic writings for computational evaluation, machine appreciation and determination of calligraphic writing styles, and ultimately developing the computational intelligence capable of generating realistic imitation of calligraphic artwork for any Chinese characters and using any given writing style as well as the computational intelligence capable of creating aesthetic, novel, and consistent Chinese calligraphic artwork with high performance.

书法是中国传统文化中的瑰宝；而对以书法作品为代表的艺术作品图像进行计算机处理与深度分析，进而实现自动创作的机器智能也是人工智能研究中最具挑战性的核心问题之一。本课题将以大数据挖掘与深度学习为手段，研究计算机书法仿写与创作中一系列形象思维领域内的智能算法设计、建模与优化的问题。在当今大数据的时代，对中国书法艺术开展智能化的研究，也有利于保护与弘扬中国传统文化，并在"一带一路"的时代契机下，促进中华民族的伟大复兴。本课题的特色在于结合经典图像处理与当代人工智能技术，研究中国书法中临摹、仿写与创作过程中所涉及到的多个形象思维建模、推理、算法设计与优化的关键科学问题，先后开展对书法作品图像的智能笔划提取、书写形态与汉字结构的表达学习、书写美观度的评价建模、书写风格的机器鉴赏判定，最终实现以任意书写风格对任意汉字生成逼真的仿写效果，并能自动创作出风格美观、新颖与统一的书法作品的计算智能。

汉字书法是中华文明的瑰宝。近年来，人工智能在汉字理解、仿写与创作中取得了诸多鼓舞人心的创新成果。然而，传统算法在泛化性、鲁棒性及多样性上表现欠佳；而当代深度学习模型对数据依赖严重。具体而言，目前深度学习算法存在以下四个关键挑战：..1）中国书法的书写自由度高、形态变化多样；上述复杂视觉信息蕴含在高维特征空间中，难以被常规机器学习算法有效捕捉并形成可解释的统一形状表达。.2）不少汉字书法作品因年代久远而遭受各类污损，影响了可供机器学习样本的质量与数量。.3）经典书法作品数量与覆盖的汉字数目有限，造成了可供机器学习的中国书法样本整体数量不多，难以充分训练大数据学习算法。.4）书法作品中汉字内容与风格强耦合，共同决定了最终书写形态，加剧了算法对训练样本的依赖性。..针对上述四个瓶颈问题，本课题分别开展了以下的研究内容：..1）提出了以汉字骨架作为汉字书法作品的生成与统一表达的可解释的先验，并设计了智能骨架提取算法，可通过对少量训练数据的学习实现对风格各异、形态迥异的汉字书法的大规模、高质量地骨架提取。.2）提出了基于视觉Transformer网络的退化文档智能修复模型，可基于小样本学习，有效修复低质量文档，提升可供机器学习的书法样本的质量与数量。.3）探索了基于缺失数据填充的汉字书法样本增广新途径，并提出了基于像素块和分位数的缺失数据填充新模型，显著提高了缺失数据推断填充的准确度。.4）提出了基于深度自编码网络的正交非负矩阵分解模型，利用矩阵分解模型解耦数据潜在表征，可实现对汉字内容与风格的精准鲁棒解耦，降低模型对数据的依赖性。..以上算法显著增强了对汉字的视觉感知智能。在此基础上，课题进一步提出了基于上述视觉感知智能的新一代汉字智能仿写与生成模型；较之现有同类方法，可显著提升算法仿写与生成的质量。最后，课题将汉字视觉感知和生成模型迁移应用至跨领域的多个医学视觉感知和分析诊断任务中，验证了本课题提出智能模型的泛化性与通用性。