基于数据驱动的纹理元素分析、合成与编辑技术研究

照片及视频中纹理元素的分析、合成与编辑技术是数字媒体方向的关键技术之一，具有重要研究价值。其难点在于：传统数学模型或参数化过程无法统一而准确地描述纹理元素的多样性及其属性的复杂性；现有算法很难为合成与编辑过程提供可控机制；纹理数据的规模性和高维度性使得高效计算变得耗时而困难。针对上述难点，本项目拟结合学习的思路，构建纹理元素样本库及其特征数据集，采用基于数据驱动的匹配方法获取输入图像或者视频序列中纹理元素的分类及对应的特征集标定；并以此生成纹理元素类别及属性分布控制图，指导后续的编辑与合成过程；通过改进搜索策略、设计合理的数据结构和研究硬件加速技术以实现高性能计算；并拟利用友好简单的交互平台实现样本采样的预处理工作。本课题旨在借助数据驱动模型寻求纹理元素的通用表达方式，直观高效地控制和完成复杂场景下纹理元素的编辑与材料外观的细节模拟。

Analysis, synthesis and editing of textures and materials in photographs and videos is a challenging task, playing a significant role in computer graphics, computer vision and digital multi-media industries. Its main difficulties include: traditional mathematic models or parameterize processes cannot describe the variety of texture elements and the complexity of their properties consistently and accurately; it is hard for existing algorithms to provide control mechanism to the procedure of synthesis and editing; since textures always cover large-scale and high-dimensional data, its calculation is expensive and time-consuming. Therefore, this proposal plans to construct the training database composed of texture samples by combining methods in machine learning, and to extract and generate statistical models for texture elements from input images and/or video sequences by employing data-driven based matching approaches so that the synthesis and editing operations are efficient and controllable. Then we attempt to implement high performance computations through improving search strategies, designing reasonable data structures and investigating hardware acceleration techniques, and to use simple and friendly sketch-based interaction interface to help collecting samples. In summary, this project aims to seek the universal representation of texture elements based on the methodology of data-driven, which can be applied to effectively accomplish many synthesis and editing tasks on textures and materials under large scale and complex scenes contained in photos and videos with high performance.