姿势和光照变化下的人脸识别核函数及核参数的研究

项目名称： 姿势和光照变化下的人脸识别核函数及核参数的研究

项目编号： No.60873168

项目类型： 面上项目

立项/批准年度： 2009

项目学科： 金属学与金属工艺

项目作者： 陈文胜

作者单位： 深圳大学

项目金额： 30万元

中文摘要： 本项目拟解决人脸识别技术中的瓶颈问题，即姿势和光照问题。经验结果表明核判别分析（KDA）是解决此问题行之有效的方法，其关键在于对核函数和核参数的选取。目前对核函数和核参数的选取均为人为的，为避免普遍采用的人工选取的方法，我们拟研究人像识别KDA算法的泛化能力，对算法误差的离差－方差分解和稳定性进行理论分析，建立离差、方差、稳定性、核函数和核参数之间相互关系的理论框架，并用来指导构造和选取最佳核函数、确定最佳核参数，从而增强KDA方法非线性分类特征的提取能力，使人像识别KDA算法对户内外环境下人像姿势和光照的变化均是稳定的。通过在公开的人脸数据库上进行实验来评估和验证所得KDA方法最佳核函数的效果。本课题所得结果将更深入地刻划出KDA学习效果与核函数及其参数变化的特性，这对将来的基于KDA算法的实际应用提供重要理论依据，也为人脸识别技术在商业和法律上的应用开辟新途径。

中文关键词： 人脸识别；姿势和光照问题；核方法；非线性特征提取

英文摘要： This project aims to address the bottleneck problems of face recognition technology, namely pose and illumination problems. The empirical results showed that KDA is a promising approach to solve these issues. However, the performance of KDA on face recognition is sensitive to the kernel function and its parameters.The current methods select the function and its parameters manually. Instead of following the current ad hoc approaches in selecting (existing) kernel function and its parameter(s), we propose to adopt the concept of generalization ability (GA) and perform theoretical analysis based on bias-variance decomposition and stability. Through the study of GA of KDA, we will find out the relationship between bias, variance, stability, kernel function and kernel parameter(s). With these findings, we plan to develop an "optimal" kernel function(s) with optimized parameter(s) for KDA on face recognition which is robust to the change in poses and illuminations in both indoor and outdoor environments. Extensive experiments will be performed using public available face image databases to evaluate and validate the "optimal" kernel function(s) in KDA. The results of this project will provide in-depth understanding on the behaviors and characteristics of kernel functions and kernel parameters in the KDA learning process. The theories to be developed in this project will not only be useful for designing KDA-based algorithms, but also blaze a new path for face recognition technology applications in commerce and law enforcement.

英文关键词： Face Recognition;Pose and Illumination Problems;Kernel Method;Non-linear Feature Extraction