图像三维重建是从记录真实场景的照片、视频等信息中还原出具有立体感、深度感三维场景的关键技术，是混合现实领域中场景建模必不可少的前提和基础。当前大多数图像三维重建方法通过分析图像中各种特征信息进而逆向重建，这类方法受各种外部条件影响，真实感不太理想且效率不高。为提高重建场景的真实感，本项目提出一种基于纹理概率语法的图像三维重建机制。该方法首先研究基于视觉注意的图像纹理特征提取方法，实现纹理信息与其他信息的分离，在减少计算区域的同时保证了纹理信息的完整性和有效性。进一步结合概率语法的规则和统计特性，提出基于纹理概率语法的三维重建，利用规则将基于统计学的纹理特征进行三维信息的映射、重构和装配。并利用纹理质量来区分渲染计算模式，对不同注意区域进行快速计算，提高渲染效率和显示速度。此项目的研究能为低维度和海量信息的图像场景重建开拓新思路，能大力促进场景建模的进一步广泛应用。

Image three-dimensional reconstruction is the key technology of restoring three-dimensional scene with stereo feeling and deep feeling from a record of reality scene photos, video and other information，which is the premise and foundation for scene modeling in the mixed reality field. Nowadays, most of image 3D reconstruction techniques focus on analyzing various features of image information and then reverse reconstruction, which is poor sense of reality and low efficiency with influence by various external conditions. Therefore, this project proposes a novel image reconstruction mechanism based on texture probabilistic grammar. Firstly, this project do the research about the extraction method of image texture features which based on visual attention in order to separate texture from other information, which can be used to reduce computational area and ensure integrity and effectiveness of texture information. Secondly, combing rule and statistical properties of grammar, this project proposes a 3D reconstruction based on texture probabilistic grammar, and then according rules to realize three-dimensional information mapping, reconstruction and assembly texture features based on statistics. Finally, this framework uses texture quality to distinguish rendering computing model and calculate rapidly for different attention regions, in order to improve rendering efficiency and the speed of display. The research would open up the new ideas for low dimension and abundant information of image scene reconstruction, which can push the widely use of scene modeling further.