三维表情识别中的张量表示及分解理论和算法研究

项目名称： 三维表情识别中的张量表示及分解理论和算法研究

项目编号： No.61471032

项目类型： 面上项目

立项/批准年度： 2015

项目学科： 无线电电子学、电信技术

项目作者： 阮秋琦

作者单位： 北京交通大学

项目金额： 90万元

中文摘要： 表情识别是人机交互的关键技术，它在计算机视觉、行为科学、情感计算、监控系统、社会安全、舆情分析及游戏娱乐等领域中有重要的理论意义和应用价值。但是，人脸是一个塑性变形体，表情建模难度大，在统一性特征提取、表情描述精细度及识别率方面仍存在很多亟待解决的问题。基于张量代数的方法有可能避免维数灾难和小样本等问题，并且可有效去除张量数据的冗余。基于张量子空间的建模与分解对揭示人脸表情流形以及增强表情类别间的鉴别性有较大优势。但是目前基于张量表征的表情识别尚不深入。本课题将研究以下关键问题： ①探讨三维人脸表情的张量子空间建模及正交张量流形学习算法，同时给出理论解释；②研究张量分解和流形学习相结合的张量-张量映射的降维算法，探讨三维人脸表情识别中的理论依据；③非负张量分解（NTF）与图保持的流形学习准则结合，探讨非负基图像的稀疏性表示；④搭建基于张量分解和流形学习的三维人脸表情识别验证系统。

中文关键词： 张量理论；表情识别；图保持；流形学习

英文摘要： The expression recognition is key technology on natural man-machine interaction. The expression recognition has important thery significance and application worth on the computer vision,action science, sensibility computing, surveillance and control system, social security, public feelings analysis and entertainment. But,the human face is non-rigid body . the expression modeling is very difficult. There are a lot of problems to solve on the unification feature extraction, recognition robustness, description precision of face expression and recognition ratio and so on. The problem about dimension calamity and small sample is able to be avoided by method based on tensor algebra. The redundancy of the tensor data can be effectively take out. There are big advantages in the modeling and decomposition based on tensor subspace for opening out human face expression manifold and enhancement discriminate performance between expression sorts. But, the 3D human face expression recognition based on tensor representation is not enough. A lot of theory problem also need to be solved. The project will study the theory and algorithm about tensor subspace modeling and orthogonal tensor manifold learning of 3D expression; The algorithm of reduce dimension about combining tensor decomposition and manifold learning of tensor-tensor, exploring the theory gist of 3D expression recognition based on tensor; Exploring sparseness of the non-negative basis images by combining Non-negative Tensor Factorization and the graph preserving based manifold learning; At the same time construction a 3D face expression recognition system based on Tensor Factorization and manifold learning.

英文关键词： tensor theory;expression recognition;graphics preserving;manifold learning