【作者】 王琦；

【导师】 鲁东明；

【作者基本信息】 浙江大学 ， 计算机科学与技术， 2011， 博士

【摘要】 以敦煌为代表的古代壁画历经千年,内容涵盖宗教、历史、地理、艺术、风俗、服饰等诸多领域,是一个巨大的图像资料库。由于其本身具有很强的多元性和综合性,成为各领域研究的重要资料来源。随着壁画资料的数字化,壁画资料的计算机检索利用成为可能。由于古代壁画内容丰富、构成复杂,而目前的检索技术缺乏对壁画图像绘画特性的体现,因此难以满足研究工作对检索全面性和准确性的要求。针对上述问题,本文深入研究了古代壁画的内容表达和度量,提出了基于主题和基于场景的语义检索技术用于壁画资料的检索,结合古代壁画辅助研究和保护的应用验证了本文方法的有效性。本文的主要研究工作包括以下几个方面：(1)综述和分析了图像语义检索技术的发展和现状。目前检索技术难以满足壁画检索的需求,原因在于没有考虑壁画的构图学特征,而且缺乏对复杂内容语义的处理能力。同时,面向古代壁画检索及应用的研究,还对文本信息检索、图像修复等相关研究进行了综述。(2)研究了古代壁画图像的内容表达和度量问题。由于目前的检索方法不能很好地表达壁画内容的布局信息,本文首次根据壁画的构图学特征,对壁画内容的位置、语义及内容间的关联关系进行了分析,提出了融合构图和语义的相关度模型,用构图显著度、语义相关度和主题相关度三个指标描述壁画内容与检索意图的符合程度,实现了壁画内容相关度的量化度量。(3)研究了两类古代壁画图像的语义检索,提出了基于主题的检索和基于场景语义的检索算法。其中,基于主题的检索采用查询扩展的方法,通过内容间的语义相关关系和主题相关关系对检索进行优化。然后,通过相关度排序,获得检索结果。基于场景语义的检索结合领域知识的本体表达,将场景语义转化为检索意图,引入编码和索引的方法,实现了场景的快速死别和提取,得到按相关度排序的壁画场景。(4)实现了语义检索技术在古代壁画辅助研究与保护中的应用。将语义检索技术用于辅助古代壁画的类型学研究,实现了基于语义检索的古代壁画整理,减少了人工方法造成重要资料遗漏的问题,提高了资料分类的效率。采用语义检索技术用于古代壁画图像的修复研究,可有效区分候选图像在造型和风格上的差异,结合ASM,PCA及图像融合技术,实现了古代壁画人物面部的虚拟修复。结合敦煌壁画研究的实际应用表明,本文的方法具有较高的实用性,提高了研究、保护工作的质量和效率。本文的研究工作仅仅是一个起步,今后可以继续在融合壁画和文献的资料关联检索,基于数据挖掘的壁画特征关联分析以及壁画构图模式的演变等方面展开深入的研究。更多还原

【Abstract】 Ancient murals as represented by the Dunhuang murals, provide the researchers with rich and vital resources for studying religion, history, geography, art, folk customs and costumes etc. The development of ancient murals digitization makes it possible for computerized retrieval. However, the present image retrieval technologies have difficulties in retrieving ancient murals, since they lack of the abilities to handle complex semantic and features of layout in painting.Aiming at the above-mentioned problems, this thesis discusses the essential techniques of mural contents representation and measuring, and proposes a novel semantic retrieval approach for ancient murals. The real applications in Dunhuang murals show the efficiency and effectiveness of our approach.The research of this thesis includes the following topics:1. Review of the development tendency of semantic image retrieval is given. Furthermore, text information retrieval, image completion technologies are introduced, which provide ideas and inspiration for this thesis.2. A relevance ranking model for relevance evaluation is proposed. The relevance ranking model measures the relevance of mural images from three aspects which are layout, semantics and topic. To the best of our knowledge, this is the first time the layout features in the view of painting are taken into similarity measure.3. Topic based retrieval and scene based retrieval are proposed for murals retrieval. By query expansion, topic based retrieval optimizes query according to its semantics and topic. Then, the results are sorted by the relevance ranking model. With the aid of domain knowledge, scene based retrieval converts query into the real intention of user. Then, a fast content indexing schema based on encoding is proposed to speed up the scene distinguish and extraction process. Finally, the results scenes are sorted by a comprehensive ranking mechanism.4. Based on the work above, a prototype system is designed and implemented for ancient murals research and preservation. In the case of ancient murals research, semantic retrieval technology is applied to murals classification. Our method can avoid the missing of important murals and improve the efficiency. By applying semantic retrieval technology to the virtual completion of ancient murals, the most relevant images with similar sculpt and style are selected as candidate images. Then, ASM, PCA and region blending method are used to produce a seamless completion.The applications in Dunhuang murals show that our method can be applied on the real applications for ancient murals research and preservation. In future, we will also explore the mural-scriptures cross-media retrieval, the association analysis of mural images and the evolution of mural layout.更多还原