基于多线性子空间分析的视频特征压缩表示及应用研究

项目名称： 基于多线性子空间分析的视频特征压缩表示及应用研究

项目编号： No.61472145

项目类型： 面上项目

立项/批准年度： 2015

项目学科： 计算机科学学科

项目作者： 韩国强

作者单位： 华南理工大学

项目金额： 83万元

中文摘要： 视频数据的特征压缩表示是高效处理海量视频数据的前提。由于张量这种表示形式能够很好地描述多维度视频数据内在的自然结构和相关关系，视频数据需要用张量来进行描述。然而，多维度的视频数据往往具有高维特性，存在大量的冗余，仅用一些低维的子空间就可以进行表示。多线性子空间分析方法能够直接对张量进行操作，在降维过程中每一阶张量的冗余信息都能够得到有效的移除。传统的多线性子空间分析方法存在着一定的局限性，如：要求基本向量是正交的，难以对投影矩阵进行解释，而且只能获取局部最优解等等。在本项目中，我们拟研究新的基于稀疏表示的多线性子空间分析方法，运用高阶奇异值分解法、带约束条件的目标函数和稀疏表示来克服传统方法的局限性，探讨其在多维度海量视频数据特征压缩表示和物体识别中的应用，并将新方法应用于平安城市、智慧交通、智慧医疗等领域。

中文关键词： 数字图像处理；视频处理；稀疏表示；特征提取；压缩感知

英文摘要： The compact representationof video data is a prerequirement for processing a large number of video data efficiently. Since tensors can capture the natural structure and the corelation relationship of multi-dimensional video data, they are often used to describe video data. But multi-dimensional video data has high dimentional property, and contains a lot of reduncdant data. The structure of multi-dimensional video data can be captured by using some low dimensional space. Multiple subspace analysis is able to operate tensors directly, and reduces the reduncdant information in each mode of tensors during the dimensional reduction. Traditional multiple subspace analysis method has several limitations: (1) it require that basic vectors are orthogonal.(2) Traditional method is difficult to explain the projection matrix. (3) It is easy to fall into the local optimal. In this project, we will propose some new spase multiple subspace analysis approaches, and solve the limitations of convencient approaches by adopting higher-order singular value decomposition and the objective function with the constraints. We explore the applications of the proposed approaches on the compact representation of multi-dimensional video data and object recognition. We also consider apply the proposed approach to the areas of safe city, intelligent transportion and intelligent medicine.

英文关键词： digital image processing;video processing;sparse representation;feature extraction;compressed sensing