【作者】 宋金晶；

【导师】 王守觉； 武妍；

【作者基本信息】 同济大学 ， 模式识别与智能系统， 2006， 硕士

【摘要】 近年来,随着社会科学、经济的不断发展,人脸识别技术获得了广泛的应用,已成为模式识别领域的一个研究热点。本文对人脸识别技术的理论和方法进行了较深入的研究,并着重就基于子空间分析的人脸特征提取方法和特征空间几何分类器进行了研究和实验验证。 基于子空间的特征提取方法具备计算效率高、表征能力强等优点,其中又以基于主元分析的特征脸方法和基于独立元分析的方法比较成功,不过由于人脸图像较易受各种外界因素的干扰,简单的一次主元特征提取或独立元特征提取很难得到人脸图像在特征空间内准确表征,本文在二次主元分析法的基础上,结合独立元分析方法能够提取图像高阶统计特征,从而更有效反映人脸面部细节的优点,提出了混合独立元特征提取方法。实验证明,使用混合独立元特征表征的人脸图像在不同分类器下的识别效果,要优于基于传统子空间分析的方法。 在识别分类器方面,本文在可调最近邻特征分类器的基础上,将可调特征子空间的构筑从二维推广到三维,提出可调最近邻特征面分类器,在三维空间的基础上提出可变等势曲面的概念,通过动态调节参数,优化分类器对于特征点在高维空间内非线性几何分布的描述,从而达到最佳分类识别效果。实验表明该方法在识别率、稳定性、应用的灵活性等方面优于传统的识别方法。 最后对本文工作进行总结,并就进一步的研究方向进行了简要讨论。更多还原

【Abstract】 The human face recognition technology has been applied in many areas with developing of social science and economy. It has become a hotspot in pattern recognition study. This thesis mainly focuses on theory and method of human face recognition, especially on feature extraction method based on subspace analysis and geometrical classifier in high dimensional feature space.Feature extraction based on subspace analysis has advantages such as high computing efficency and strong geometry feature description ability. Eigenface method based on principal component analysis and the method based on independent component analysis are two succuessful ones. While because of the disturbance coming from environment, one-order feature extraction based on PCA or ICA can hardly get the features exactly standing for identification. This thesis proposes mixed ICA feature extraction method that integrates second-order analysis with the high order infomation extraction abilitty of ICA. Experiments prove that using different classifier, mixed ICA feature extraction method has more excellent performance than traditional subspace analysis method.On recognition classifier, based on tunable nearest neighbor classifier, the tunable nearest neighbor plane classifier is proposed, which extends the concept of feature subspace building from two dimension to three dimension. Through the tunable parameter, we can optimize the description of the nonlinear distributing of feature points in high dimensional feature space, and get better performance. Through experiment, this method performs more efficient than traditional methods on recognition rate, stability and applicability.In the finality, the thesis is summarized firstly, and then the problems requiring further studies are discussed.更多还原