项目名称: 基于Mercer核的非负矩阵分解关键问题研究及其在人脸识别中的应用
项目编号: No.61272252
项目类型: 面上项目
立项/批准年度: 2013
项目学科: 自动化技术、计算机技术
项目作者: 陈文胜
作者单位: 深圳大学
项目金额: 80万元
中文摘要: 本项目拟解决基于Mercer核的非负矩阵分解(NMF)人脸识别技术中的三个瓶颈问题,即核相容性、计算复杂度和增量学习问题。通过对核矩阵性质进行深入研究,拟从理论上建立核相容及性能评价的一般准则,以指导构造和选取最佳NMF相容核函数,突破目前核函数与NMF不相容的现状;研究和使用稀疏编码、鉴别信息来增强核空间中NMF局部特征提取能力,使用投影梯度技术来加快核NMF的收敛速度并提高算法的稳定性,进而提高计算效率;拟使用鉴别分块策略来解决目前核NMF不能进行增量学习的问题;通过在人脸数据库上进行实验来评估所得相容核NMF方法的效果。最终将所取得的成果应用于开发户外环境下高效实用的相容核NMF人脸识别系统,使该系统不但能够抵抗人脸姿势和光照变化,而且具有极高的识别率。本项目的结论不但可用于人脸识别领域,还可用于所有基于核空间中NMF机器学习领域,这将为核NMF技术在商业和法律上的应用开辟新途径。
中文关键词: 模式识别;核方法;非负矩阵分解;;
英文摘要: This project aims to address three bottleneck problems including kernel compatible problem, computational complexity problem and incremental learning problem which are encountered in Mercer kernel based non-negative matrix factoriztion (NMF) method for face recognition. By conducting in-depth investigation of the properties of kernel matrices,this project will theoretically establish the common criterion of kernel-NMF compatibility and kernel performance evaluation.It will be used as a guide to constructing and selecting the optimal NMF-compatible Mercer kernel functions. We expect to achieve a breakthrough in current compatibility problem between kernel functions and NMF method. The sparse coding,discriminant information will be exploited in this project to enhance the power of the useful local feature extraction of NMF method in kernel space. We propose to utilize gradient projection algorithm to accelerate the convergence rate and improve the stability of the proposed compatible kernel NMF algorithm, and thus promote its computational efficiency. This project will further adopt discriminant block strategy to solve the incremental problem of kernel NMF. With these findings, we plan to develop a highly efficient and practical compatible kernel NMF face recognition system which is not only robust to the variat
英文关键词: Pattern recognition;Kernel method;Non-negative matrix factorization;;