完善频谱脸人像识别的分类器设计  

Finishing the Classifier Design of Spectroface Human Face Recognition

作　　者：赖剑煌[1] 颜鑫弘[1] 邓东皋[1] 

机构地区：[1]中山大学数学与计算科学学院计算机视觉研究中心,广州510275

出　　处：《中国图象图形学报（A辑）》2002年第5期466-471,共6页Journal of Image and Graphics

基　　金：国家自然科学研究基金重点项目 (1963 10 80 );中山大学基金项目 (0 0 -0 16-4 2 9163 )

摘　　要：频谱脸方法是一种利用小波变换和 Fourier变换有效地提取人像的位移不变特征和表情相对不变特征的方法 .该文着重讨论了频谱脸方法系统化的预处理方法和相似性度量选择这两个关键性问题 .其中 ,矩的方法被用于对人像进行预处理 ,因为它能有效地对人像的伸缩和平面旋转进行矫正 ;通过对最近邻法、平均法、Hausdroff距离法和修正的 Hausdroff距离法等 4种典型的相似性度量方法中进行比较和分析的结果表明 ,最近邻法、平均法和修正的 Hausdroff距离法都是频谱脸方法进行相似性度量的有效方法 ,其中 ,最近邻法是最有效的方法 ,它对诸如位移、伸缩、平面旋转、少许遮掩及少许姿势、表情和光照条件的变化等多种影响人像识别的因素均具有最佳的容错性 ,并在 Yale和 Olivetti人像数据库上进行了识别试验 ,分别取得了 97%和Spectroface is a face representation method using wavelet transform and Fourier transform and have been prove to be invariant to translation and tolerant to expression variety. In this paper, the two important issues on Spectroface system is studied. One is how to preprocess system, another is the selection of similarity measurement. The moment is employed to preprocess system that it is good method to normalizing the scale and rotation of human face. The similarity measurement has been selected by comparing four typical kinds of similarity measurement, i.e., nearest neighbor method, averaging method, Hausdorff distance method and modified Hausdorff distance method(MH). Nearest neighbor method, averaging method and modified Hausdorff distance method are good for Spectroface. Nearest neighbor method is the most effective method in the recognition of frontal faces with translation, scale, rotation, different facial expressions, small pose, small occlusion and different illumination condition. It gives high accuracy as 97% and 99% in Yale and Olivetti face image databases respectively.

关 键 词：人像识别 频谱脸 分类器 设计 特征提取 计算机视觉 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]