基于多个再生核希尔伯特空间的多角度人脸识别  

Multi-angle Face Recognition Algorithm Based on Multi-RKHS

作　　者：林硕[1] 龚志恒[1] 韩忠华[1,2] 史海波[2] 

机构地区：[1]沈阳建筑大学信息与控制工程学院,沈阳110168 [2]中国科学院沈阳自动化研究所,沈阳110016

出　　处：《光子学报》2013年第12期1436-1441,共6页Acta Photonica Sinica

基　　金：国家自然科学基金(No.61164012);国家重大科技专项(No.2011ZX02507-006)资助

摘　　要：针对传统谱算法在人脸识别中的局限,提出一种基于多个再生核希尔伯特空间的多角度人脸识别算法.首先,利用landmark标记法对图像进行预处理,得到训练图像的角度;其次,通过人脸数据的多次核化迭代,使其在构建的再生核希尔伯特空间中呈线性,针对不同类型的人脸数据,建立多个再生核希尔伯特空间;最后,通过对比训练图像,判断待检测人脸图像的再生核希尔伯特空间归属,实现多角度人脸识别.选取FERET和CMU-PIE两类数据集进行对比实验,实验结果表明:所提出的算法不仅在平均识别率上高于传统算法5%,平均识别效率也较传统算法提高20%.The traditional spectrum algorithms are limited in face recognition problem. For its characteristics of problem, a novel method based on multi-reproducing Kernel Hilbert space was proposed. Firstly, the images were processed by the landmark method, and the angle of training images could be obtained. Secondly, the face data was iterated by the Kernel, then face data expressed linearly in the reproducing Kernel Hilbert space. Thereafter, for many types of face data, the multi-reproducing Kernel Hilbert space were established. Finally, the reproducing Kernel Hilbert space belonging of human face image was judged by the comparison of training images, and the multi-angle face recognition achieved. The two classes of data sets were selected as the experimental data, which consisted of FERET and CMU-PIE. A large number of experiments were carried out. The results show that the proposed method has great effect to recognise multi-angle face. The average recognition rate and efficiency are 5% and 20% higher than the traditional algorithms, respectively.

关 键 词：谱算法 多角度 预处理 核希尔伯特空间 迭代 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]