联合空谱信息和Gabor特征的高光谱人脸识别算法  

Hyperspectral Face Recognition with Spatial-Spectral Fusion Information and Gabor Feature

作　　者：魏冬梅[1] 张立人 胡楠楠 刘璐[1] 马娜[1] 赵曰峰[1] 

机构地区：[1]山东师范大学大学物理与电子科学学院,山东济南250014

出　　处：《北京理工大学学报》2017年第10期1077-1083,共7页Transactions of Beijing Institute of Technology

基　　金：山东省高等学校科技计划资助项目(J14LN06);山东省重点研发计划资助项目(2016GGX101016);国家自然科学基金青年科学基金资助项目(61501283);山东省科技发展计划资助项目(2014GSF116004);山东师范大学培育基金资助项目

摘　　要：提出一种采用高光谱图像的人脸识别算法.根据人脸肤色在可见光范围内的光谱特征进行波段选择并依据人脸结构特征,对选定波段的灰度图像进行Gabor特征提取.最后分别进行特征层上的融合识别和决策层上的融合识别.特征层融合的权重系数由反射率和正确识别率共同决定,决策层融合算法采用"最高票当选制"原则.利用香港理工大学的高光谱人脸数据库对进行验证.结果证明,本文算法在识别速度和正确识别率方面都得到了显著改善,在3幅训练样本情况下,正确识别率达到96.5%.相对于全波段参与识别,识别速度提高了约3倍.A hyperspectral face recognition was proposed using the HK-PolyU database in this paper.Twelve spectral bands were chosen from the hyperspectral face image data cubes according the spectral reflection property of skin,and then Gabor features were extracted for each selected spectral gray image respectively.Next,the feature fusion and decision fusion were studied.For the feature fusion,the fusion image was constructed by combining the twelve Gabor-feature vectors,and the weigh coefficients were decided by both the spectral reflection and the recognition accuracy.Maximum voting system was employed in the decision fusion.The validation experiment results show that,the recognition accuracy can reach to 96.5% when three image cubes are selected as training from each class.The recognition speed is more than 3 times of that without band selection.The extensive experiments show that the promising proposed approaches are superior both in the recognition accuracy and in recognition speed.

关 键 词：高光谱图像 人脸识别 GABOR特征 空谱信息融合 投票 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]