局部统计不相关非线性鉴别变换  

Local Uncorrelated Non-linear Discriminant Transform

作　　者：丁洁[1,2] 荆晓远[1] 姚永芳[1] 李力[1] 

机构地区：[1]南京邮电大学自动化学院,江苏南京210023 [2]南京邮电大学计算机学院,江苏南京210003

出　　处：《计算机技术与发展》2014年第12期105-108,共4页Computer Technology and Development

基　　金：国家自然科学基金资助项目(61272273)

摘　　要：子空间学习方法是研究人脸识别问题的经典算法。然而,在实际中,人脸图像通常是高维线性不可分的,此时线性方法很难提取出更有效的特征;同时,人脸图像不满足理想的高斯分布,而传统的统计不相关方法使用所有样本的均值来估计期望,这种估计方法在样本数很少的情况下会出现较大的偏差。为了解决上述问题,文中提出了局部统计不相关非线性鉴别变换(LUNLDT)方法。该方法在核空间中,通过改良的统计不相关约束,迭代地求解出最优的鉴别向量集。AR人脸库和CAS-PEAL人脸库的实验结果验证了所提算法的有效性。Subspace learning method is the classical algorithm of face recognition research. However,in the real world,face images are usually high- dimensional and non- linear,it is difficult for linear approach to extract more effective feature. In addition,face images do not meet the ideal Gaussian distribution. However,traditional statistical uncorrelated methods employ the mean of sample- sets to estimate expectations,which may cause large deviations in the case of small sample size. In order to address the aforementioned problems,propose a novel feature extraction approach named local uncorrelated non- linear discriminant transform（ LUDNLT）. In the kernel space,iteratively calculate the optimal discriminant vectors under the corresponding statistical uncorrelated constraints. Experimental results on AR and CAS- PEAL face databases demonstrate the effectiveness of the proposed method.

关 键 词：特征提取 核空间 统计不相关约束 局部统计不相关核鉴别变换 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]