具有统计不相关性的广义图像投影鉴别分析  

Statistical Uncorrelated Generalized Image Projection Discriminant Analysis

作　　者：高秀梅[1] 杨静宇[1] 陈才扣[1] 杨健[1] 

机构地区：[1]南京理工大学计算机科学系,江苏南京210094

出　　处：《小型微型计算机系统》2004年第7期1119-1123,共5页Journal of Chinese Computer Systems

基　　金：国家自然科学基金( 60 0 72 0 3 4)资助;国家教委博士点基金资助

摘　　要：提出了一种新的图像投影鉴别分析方法 .首先 ,与 L iu投影鉴别分析方法相比 ,具有能够消除投影特征分量之间相关性的优点 .其次 ,该方法从整体上考虑投影集的可分性 .即样本在图像最佳鉴别矢量上的投影集从整体上具有最佳的可分性 .另外 ,所提出的方法是直接基于图像矩阵的 ,与以往的基于图像向量的鉴别方法相比 ,它的突出优点是大大地提高了特征抽取的速度 .最后 ,在 ORL标准人脸库上的试验结果表明 ,所提出的图像投影鉴别分析方法较 L iu的方法在识别性能上有了较大幅度的提高 ,在普通的分类器下达到 95 %识别率 .该识别率明显优于颇有影响的 Fisher-faces方法 ,其特征抽取的速度提高了近 19.6 8倍 .In this paper, a novel image projection analysis method is developed for image feature extraction. First, with contrast to the Liu's projection analysis method, the proposed one has a desirable property that the projective features vectors are mutual uncorrelated. Second, our method has that the separability of the projected set of the samples is considered from global view when calculating the image optimal set of discriminant vectors,that is the projected set of the samples on the image optimal set of discriminant vectors have the best separability in global sense.Furthermore, the proposed method is directly based on image matrices. That is to say, it needs not to convert the image matrix into high dimensional image vector like the previous linear discriminant methods based on image vectors. So much computational time would be saved if using our method for feature extraction. Finally, the proposed method is tested on ORL face databases. The experimental results indicate that the proposed method outperforms Lius', and a recognition rate of 95% on ORL are achieved with ordinary classifiers. The experimental results also indicate that the proposed method is more powerful than Fisherfaces, and more importantly, its speed for feature extraction is nearly 19.68 times faster.

关 键 词：图像投影鉴别分析 图像最佳鉴别矢量集 图像最佳投影矢量集 图像特征抽取 人脸识别 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]