中国画的计算机分析与分类算法研究

Digital arts are becoming increasingly important over the past few years along with the enormous advancement of digital technology application upon Chinese paintings. At present, computerized analysis on Chinese paintings is limited to two points of weakness: (i) directly applying image processing techniques to carry out content-based search or analysis, and such approach often fails due to the fact that the content of Chinese paintings is limited with significant level of similarities among those basic elements, such as mountains, trees, etc. (ii) current research on computerized Chinese paintings is mostly limited to simulation of individual strokes and feature extractions, which often fail to capture the creative nature of Chinese arts as well as the varying styles. This proposal will resolve the above problems according to the following approaches, which can be highlighted as: (i) description and representation of Chinese painting styles to capture their essential creative elements and hence provide high-level characterization of Chinese paintings and artistic style features; (ii) development of new approaches in extracting multiple and integrated digital features to drive a multi-kernel one class SVM and facilitate an embedded learning process to complete automated classification and recognition. Such approach will produce powerful digital tools and enable Chinese artists to increase their effectiveness of creative arts; (iii) extensive research is to be carried out on developing self-adaptive multi-source fusion algorithms to increase the level of precision in analysing and understanding Chinese paintings to exploit the embedded learning process out of multiple and integrated feature analysis. The proposed project will have a range of impacts and values, which include: (i) combination of computerized analysis and creative arts to explore a new field of research on digital arts; (ii) digitalization of Chinese arts towards computerized analysis, recognition and management of national heritages; and (iii) provision of computing support for cultural and artistic inheritance as well as theoretical developments.

中国画数字化正迅猛发展，近年来随着数字博物馆技术的发展，国画图像分析处理越来越重要。目前针对中国画的计算机数字分析还处于探索阶段，存在两点不足1）大多是基于内容的检索和分析，而中国画艺术元素有限，仅仅依据图像内容的比较不能达到很好的识别效果；2）水墨画渲染研究中大多是单独笔道的仿真研究。针对以上问题，本项目将1）对中国画中的主要艺术元素进行分割，进而对其艺术风格在像素域和变换域进行数字量化、分析和深度优化，提炼出最能体现画家个人风格的高级语义信息；2）将复合特征提取加入嵌入式学习的过程中，提出基于嵌入式学习的多核一类支持向量机对国画的艺术风格、作者进行自动分类；3）研究自适应的多源信息融合算法，将识别结果进行有机的融合，优化计算机对中国画的分析、理解的准确度。本项目将计算机与画家的创造性思维相结合，其研究将实现国画的数字化分析、识别和管理，为文化艺术的传承、鉴赏提供理论和技术支持。

中国画数字化正迅猛发展，近年来随着数字博物馆技术的发展，国画图像分析处理越来越重要。本项目基于图像、机器学习、可视化等交叉领域的知识开展了一系列算法研究和实验。本项目1)研究了艺术目标的深度学习与融合，提出一个对中国画中的主要艺术目标进行交互式分割及识别的框架，将国画自动分割成一系列的艺术目标，如宝塔、人物等，以提取画家用来表现艺术形式和抒发情感的相对稳定单元，进而针对艺术目标进行分析研究；2)在像素域和变化域内针对线条形状、主方向墨色分布、运笔力度、图像结构与细节、图像结构稀疏化特性等进行了一系列艺术风格数字量化、分析和深度优化，提炼出最能体现画家个人风格的高级语义信息；3）对中国画的优化式着色算法进行了研究；4)中国画情感分类研究。尝试通过可视化技术研究一种新的深度学习网络结构，从而利用人类情感认知优化网络，对中国画的情感进行分析和智能优化；5) 现有中国画分类大多只考虑了画作和类标签的相关性，忽略了画作间的关联以及特征重要度的影响，从而导致分类结果受制约。课题将联合互信息理论引入嵌入式机器学习，使“友情则近， 敌对则远”的嵌入式原则受到特征选择及特征重要度的影响，提出基于互信息的改进嵌入式机器学习算法，并对中国画进行作者分类。本课题可用于中国画的数字化分析、管理、理解和识别，为中国画传承和鉴赏提供有效的数字工具。研究成果可转化应用到文化保护、建立数字图书馆、国画鉴赏、艺术培训等方面，有利于传统文化的传承创新，推动文化大发展大繁荣。

内容获取失败，请点击重试