QQ客服 官方微博 反馈留言

题名	基于模式素材分解的纹理合成新技术及其应用
作者	古元亭
学位类别	博士
答辩日期	2008-05-30
授予单位	中国科学院研究生院
授予地点	中国科学院软件研究所
导师	吴恩华
关键词	纹理合成 模式 素材 特征纹理 分层纹理 纹理压缩 线纹理 图像类推 自类推 非线性卷积
其他题名	Pattern-Material Decomposition Based Texture Synthesis and Its Applications
学位专业	计算机应用技术
中文摘要	纹理合成是目前计算机图形学、计算机视觉和图像处理等多个研究领域中的热点问题。它源自真实感造型中的纹理映射问题，随着其内涵和形式的不断演变，逐渐成为一个应用广泛的多学科交叉研究课题。纹理合成问题可以被视为人工智能难题的一种简单表达形式，因为该问题涵盖了人工智能难题当中最重要的两大因素：分析与推理。而学习和模仿人类的分析与理解能力，是许多计算机学科目前共同的发展瓶颈。因此对于纹理合成问题的研究具有重要的意义，其进展也将对其它计算机学科的进步起到推动作用。实际上，目前已经有许多首创于纹理合成研究中的技术与思想被推广应用到其他研究方向上，取得了良好的成果。但是，和其他许多难题一样，一般都认为纹理合成问题距离真正的解决还有很远的距离。现有的方法还没有实现真正意义上的基于内容理解和推理的合成这样一个目标。未来的研究空间与研究任务尚非常巨大。纹理合成相关研究的难点和重点主要集中在三个方面：一是纹理基础理论研究；二是纹理合成方法研究，三是纹理合成应用研究。对于第一点，由于纹理概念的重要性，研究者很早就对纹理的数学基础进行过探讨。然而纹理概念本身存在的模糊性和纹理实例的多样性使得到目前为止，尚无明确统一的纹理数学概念与数学模型。这限制了纹理相关研究的深入。对于第二点，目前已有的纹理合成相关方法数以百计，成果丰富。总结其基本思路，这些方法大致可以归结到模式类方法-素材类方法-模式素材类方法这样一条发展路径上。故此充分结合模式类方法在结构刻画上的优势与素材类方法对于视觉细节表现上的优势将是未来主要的发展方向。对于第三点，由于纹理的广泛存在，其应用方面的研究内容相应得也非常丰富。即使在现有的纹理合成研究基础上，它与别的研究课题结合产生的各种衍生技术已经取得了令人瞩目的成果。如果能更好地改进合成技术，其未来的应用研究更大有可为。相应地，本文的研究也围绕这三个方面展开：在纹理理论方面，我们提出了纹理的一种新的素材-模式观点，用它来分析目前的纹理合成技术体系结构；在纹理合成方法方面，我们从上述观点出发，提出了两种新的纹理合成方法，分别解决合成时纹元的特征随机性过弱和多层次纹理合成两个问题；在纹理合成应用方面，我们分别将纹理合成技术应用到图像压缩、线状纹理编辑、图像可控类推和图像超分辨增强等研究领域，取得了很好的效果。本文算法的主要贡献和创新点在于如下几点：首次提出了纹理的素材-模式观点。该观点将传统的纹理概念分解成为纹理模式和纹理素材两个要素，更有利于理解纹理概念和纹理合成技术的基本思路。利用这种观点我们对纹理合成技术体系进行了研究和归纳，首次建立了纹理合成技术进化树模型。该模型可以清晰地表达纹理合成技术体系的全貌，避免了以前各种分类式总结中仅仅偏重于横向类别分划，忽略纵向演化关系的缺点。首次提出了纹元形状特征空间和特征分布纹理的概念。应用这些概念，可以通过收集样本纹理中各纹元的特征来重构出纹元特征空间，在合成时以采样方式获得各个合成纹元的特征值，再结合特征分布纹理将其重置而生成新的纹理。这样可以有效地避免一般方法所引起的纹元重复问题。首次提出了一种新的基于内容的纹理压缩框架。该框架不同于传统的以数据恢复为目标的压缩方式，而以内容恢复为目标。这种概念上的松弛为压缩提供了新的信息冗余空间。结合纹理合成方法，可以有效地对该内容空间进行压缩。建立了一套新的处理各向异性的线状纹理的合成框架。在研究中，我们观察到大量存在的具有方向性的线纹理编辑需求与传统纹理合成方法所要求的各向同性的面纹理之间存在着矛盾。为解决这个矛盾，可以将一般纹理合成使用的区域重组基本思想推广到线性域上，构建出新的线纹理合成框架。 首次提出了图像可控类推和图像自类推概念。前者可以在风格化学习中引入表征学习强度的调控项，对于风格化的程度控制具有重要意义；后者可以实现独立于训练集合的基于学习的图像超分辨功能，避免一般基于学习方式的超分辨方法由于对训练集合的高度依赖所产生的问题。
英文摘要	Texture synthesis is a highlighted topic in computer graphics, visual study and image processing for many years, and it becomes one of the representative and useful research areas in the fields with its evolvement. Texture synthesis can be considered as a simpler form of artificial intelligence challenge that many reseachers meet since it contains two most important intellective factors: analysis and conjecture. Since learning and simulating human ability on analysis and understanding is a shared bottle-neck for lots of research areas, texture synthesis research thus is of importantance to many areas and may also improve other artificial intelligence related studies as well. In fact, a number of original creative methods generated from texture synthesis have been extended to other directions and made great contributions there. However, study on texture synthesis is still far away from its real solution- “content-based understanding and synthesis”, so the future task is still intensive. So far, the most challenges of texture synthesis focus on three aspects: theoretical research, study on methods and applications of texture synthesis. In the first aspect, many researchers in fact have begun discussing the mathematical base of texture decades ago due to its importance and popularity. However, we’re still lack of a uniform and clear definition about texture now as its natural fuzzy property. This problem to some extent blocks the further improvement of texture research. In the second aspect, hundreds of methods have been presented till now. They are allocated on a develop path along such three phases as Material Method-Pattern Methods-Material and Pattern Methods. This path may indicate that the future direction of texture synthesis would be integrating both the advantage of Pattern Methods on modeling structures and the benefits of Material Methods of describing the visual detail. With regard to the third aspect, even based on current research level, existed applications have been well developed. It is believed that future texture related applications will be more attractive as synthesis techniques further advances. Accordingly, this paper focuses on the investigation to the three aspects indicated above. In the theoretical study, we present a new approach by material-pattern decomposition about texture, and use it to analyze the technical system of texture synthesis. In the synthesis algorithm study, we promote two novel methods solving the weak randomness problem and multi-level texture synthesis problem respectively. In the applications study, we propose to have various studies such as content-based image compression, strip texture synthesis, controllable image analogies, and image super resolution in corporation with texture synthesis. The main contributions of this thesis are mainly in the following aspects: First, attempt on dividing texture appearance into pattern and material is conducted. With this decomposition, we can better understand “what is texture” and “what is texture synthesis”. Also by this view, we present a new system model called technology evolvement tree to construct a complete and detailed image for texture synthesis technology development, to clearly address the relationship between each methods instead of simply making categorization with traditional way. A new concept of “texton feature space” and “feature texture” is proposed. By the proposal, we can reconstruct the feature space to describe texton distribution and sample from it to generate new textons preventing from obvious repetition problem. A new content-based compression concept is presented. This framework of image compression explores new information redundancy in images’ content space surpassing common data space. This kind of relaxation and adoption of patch-based texture synthesis algorithm help realizing a content compression method. Construction of a new synthesis framework to handle inhomogeneous strip textures, has been made. The strip textures are widely present, but it is hard to manipulate them in the common way. We extend the trivial area-texture synthesis ideas to tackle the problem in success. Promotion controllable image analogies and self image analogies extended from image analogies has been made. Controllable image analogies introduce the intensity parameter into image analogies to realize style learning in various levels, and the self analogies can be used to execute super resolution without training set support to alleviate the highly dependence on exterior information.
公开日期	2011-03-17
内容类型	学位论文
源URL	[http://124.16.136.157/handle/311060/5718]
专题	软件研究所_计算机科学国家重点实验室 _学位论文
推荐引用方式 GB/T 7714	古元亭. 基于模式素材分解的纹理合成新技术及其应用[D]. 中国科学院软件研究所. 中国科学院研究生院. 2008.

个性服务

查看访问统计

相关权益政策

暂无数据

收藏/分享

除非特别说明，本系统中所有内容都受版权保护，并保留所有权利。