【作者】 李一哲；

【导师】 陈国良；

【作者基本信息】 中国科学技术大学 ， 计算机软件与理论， 2006， 博士

【摘要】 纹理合成以人工生成纹理为目的，是计算机图形学与图像处理的重要研究领域。基于样图的纹理合成是近些年来出现的一种新技术，它以小块纹理图像作为输入而合成任意大的同类纹理。该技术大量用于制作虚拟现实中的场景，在电影、游戏娱乐等产业中有着巨大的应用需求。随着近些年来计算机三维绘图能力的飞速提高和对大尺寸高质量纹理的强烈需求，基于样图的纹理合成成为热门研究领域之一。基于样图的纹理合成以高质量的输出图像、满足实时需要的合成速度与全自动的合成过程为目标。随着近些年来研究者们的大量工作，基于样图的纹理合成技术取得了很大进展。目前该领域的难点主要体现在纹理合成的速度与合成图像的质量上。针对以上两点，本论文以高质量快速纹理合成算法为目标，所作的工作主要内容包括以下4个方面。(1)结构性纹理全局特性的识别和提取。本文首次提出一种用于识别结构化纹理中特征点分布的结构模式分析方法。该方法首先以邻域相似性为准则，对纹理中的像素进行聚类，然后以像素邻域差别为依据选择一组像素作为特征像素组，并分析该组中像素的分布，最终识别整个特征点网格，以其为基础，可以很方便地对纹理中的全局特性进行捕捉，对纹理合成相关应用起到很大的辅助作用。大量实验证明，该方法对周期及伪周期结构性纹理具有强大的通用性，能准确识别高随机度纹理中的特征点网格，并根据网格的统计数据对纹理的宏观随机度进行量化，量化后的随机度指数准确反映了纹理图像宏观结构的整齐程度。(2)基于特征点定位的纹理合成算法。本文以传统的基于块的纹理合成框架为基础，提出一种使用特征点定位的纹理合成算法。该算法在预处理阶段对样图进行结构模式分析，根据得到的统计数据对输出图像进行特征点全局模糊定位；在块匹配阶段，将特征点分布与邻域相似性一起作为匹配标准；在块边界修复阶段，基于传统的块边界切割技术，提出和使用一种强化的边界修复算法，对图像中的高频与低频分量进行分离处理，以获得更加平滑的过渡效果，从而提升输出图像的质量。理论及实验证明，使用特征点定位可以保证输出图像不会出现大的结构性错误，同时块匹配时的搜索空间根据纹理随机度的差别有不同程度的缩小，对大多数结构化纹理，搜索空间至少可以减小一到两个数量级，直接加速了纹理合成的过程。(3)基于结构坐标的纹理合成算法。本文以传统的基于像素的纹理合成框架为基础，给出一种基于结构坐标匹配的纹理合成算法。该算法在结构模式分析的基础上首次提出纹理结构坐标的概念。在像素合成阶段，结构坐标差异与像素邻域色彩差异一起作为匹配准则。通过对结构坐标匹配时的阈值进行设置，该算法还可以对输出纹理的随机度进行控制。理论及实验证明，使用结构坐标匹配明显改善了合成纹理的结构，有效弥补了传统的基于像素的合成难以再现纹理大范围特性的不足。同时，像素全局匹配时的搜索空间根据纹理随机度的差别有不同程度的缩小，有效加速了纹理合成的过程。(4)使用动态特征点匹配的纹理优化算法。本文以基于优化的纹理合成算法为基础，提出一种加入特征点匹配要素的纹理合成算法。该算法在预处理阶段以特征点周期位置为准则，生成一组经过优化的初始块集合。在块匹配阶段中，算法随时检查和保证不同块中特征点之间应有的相对位置，并对特征点网格进行动态调整。理论及实验证明，使用特征点匹配生成的初始块集合可以大大加快算法的收敛速度，而块匹配时对特征点相对位置的检查则可以有效保证输出图像结构的正确性。本研究工作的主要创新点为：(1)首次所提出的一种用于分析纹理中特征点分布的结构模式分析方法，可以有效识别和定位纹理中的全局特性，对传统的纹理合成算法是一个有力的辅助工具。(2)所提出的基于特征点定位的纹理合成算法，通过特征点定位直接排除了会导致输出图像结构性错误的匹配位置，在确保了输出图像结构正确的同时，也大幅度减少了块匹配时的搜索空间，直接提升了合成速度。(3)为了定位纹理中的全局特征，首次提出了结构坐标的概念，并将其应用于基于像素的纹理合成框架中。通过结构坐标的匹配，将纹理全局结构信息准确传递给合成算法，有效改善了输出图像的结构，同时大幅度减少了像素匹配时的搜索空间，提升了合成速度。更多还原

【Abstract】 The goal of texture synthesis is to produce texture. It is an important research area in both computer graphics and image processing. Sample-based texture synthesis was a new technique proposed in recent years. It takes small texture image as input to synthesis same kind of texture but in arbitrary size. This technique is widely used in producing virtual reality scene and satisfies the great need in application areas such as movie industry and computer game. With the rapid development of computer 3D graphics ability and the great need of large size and high quality texture in recent years, sample-based texture synthesis became a hot research area.The goal of sample-based texture synthesis is high quality output image, synthesis speed that satisfies real time application and full automatic synthesis process. With the large amount work of researchers these years, sample-based texture synthesis technique has achieved much. The main challenge of this area currently lies in synthesis speed and the quality of synthesized image. In order to solve these problems, this thesis aims at fast and high quality texture synthesis algorithm.Main work of this thesis includes the following:(1) Global feature recognition and extraction for structured texture. A new structural pattern analysis method which is used to recognize feature point distribution pattern in structured texture is firstly proposed in this thesis. The method first clusters pixels in the texture according to the rule of neighborhood similarity, then selects a pixel group as the feature point group due to pixel neighborhood difference and analyses the distribution of pixels in this group, and finally, the method recognizes the feature point mesh. With the mesh, it is convenient to capture the global characteristics of texture. Thus the method can assist much in texture synthesis related applications. Many experiments shows that the method exhibit high generality to periodic and pseudo-periodic structured textures and can accurately recognize feature point mesh in a highly stochastic texture. Bases on the mesh statistic data, the method also quantizes macro scope randomness of texture. The quantized texture randomness parameter faithfully reflects the regularity of texture macro scope structure.(2) Feature point locating based texture synthesis algorithm. Bases on the traditional framework of patch-based texture synthesis, this thesis proposes texture synthesis algorithm using feature point locating. In the preprocessing phase, the algorithm uses structural pattern analysis on the sample image and performs global approximate feature point locating for the output image based on. the statistic data get. In the border fix phase, based on the traditional border cut technique, the algorithm proposes and uses a enhanced border fix method. The method splits the high and low frequency image component and deal with them separately to get smoother transition effect. It helps to improve image quality. It can be proved both in theory and in experiments that feature point locating assures the elimination of large scale structural error in the output image and at the same, searching space when performing patch matching is reduced more or less due to texture randomness. For most structured texture, searching space can be reduced not less than 1 to 2 decimal levels. This accelerates the texture synthesis process directly.(3) Structural coordinate based texture synthesis algorithm. Based on the traditional framework of pixel-based texture synthesis, this thesis proposes a texture synthesis algorithm which bases on structural coordinate matching. The algorithm firstly brings forward the concept of structural coordinate. In the pixel synthesis phase, differences of both structural coordinate and neighborhood color are used as matching principle. By the means of setting threshold value for structural coordinate matching, the algorithm offers the control on output texture randomness. Both theories and experiments shows that the using of structural coordinate matching remarkably improves the structure of synthesis texture and effectively makes up the lack in reproducing large scale texture characteristic. At the same time, searching space when performing global pixel matching decreases more or less due to texture randomness. This effectively accelerates the texture synthesis process.(4) Texture optimization algorithm using dynamic feature point matching. Bases on the optimization based texture synthesis algorithm, this thesis proposes a texture synthesis algorithm with feature point matching. In the preprocessing phase, the algorithm uses the periodic location of feature point as a rule to generate an optimized initial patch set. In the patch matching phase, the algorithm checks and assures the reasonable distance between feature points from different patches, and adjusts feature point mesh dynamically. Theories and experiments shows that initial patch set generated using feature point matching can accelerate the convergence of algorithm a lot. And feature point relative position check when matching patches can effectively assure the structure correctness of the output image.Main innovation points of the thesis are:(1) A firstly proposed structural pattern analysis method. It can effectively recognize and locate global characteristic of texture. It is a powerful assisting tool for conventional texture synthesis algorithms.(2) The proposed feature point locating based texture synthesis algorithm. The algorithm excludes the matching position which leads to structural errors in output image by means of feature point locating. This assures the correct structure of output image and greatly reduces the searching space when matching patches. It improves the synthesis speed directly.(3) In order to locate texture global characteristic, this thesis firstly brings forward the concept of structural coordinate, and then uses it in the framework of pixel-based texture synthesis. Through the matching of structural coordinate, global structural information of texture can be accurately passed to synthesis algorithm. This effectively improves output image structure and greatly decreases the searching space when matching pixels. It improves synthesis speed.更多还原