多尺度分块协同表示的选择性集成人脸识别算法  

A Selective Ensemble Face Recognition Algorithm Based on Multi-scale Patch Collaborative Representation

作　　者：殷清燕 史加荣[1] 魏宗田[1] 岳红云[1] 

机构地区：[1]西安建筑科技大学理学院,陕西西安710055

出　　处：《信号处理》2016年第6期707-714,共8页Journal of Signal Processing

基　　金：国家自然科学基金(11526161;11471255;61403298);陕西省自然科学基础研究计划(2014JQ8323);西安建筑科技大学科技基金项目(RC1438;QN1508)

摘　　要：为了进一步改善人脸识别系统在小样本条件下的识别性能,本文在图像分块协同表示分类算法的基础上,提出了一种新的基于多尺度分块协同表示选择性集成的人脸识别算法。该算法首先通过对各个尺度下的图像子块进行总变差加权,突出具有鉴别能力的局部关键特征子块的判别作用;其次通过多尺度分块协同表示的选择性集成,显著地提高了分类器的泛化能力和稳健性。对于三种不同采集条件下涵盖各种光照、表情和姿态变化的标准人脸数据库进行数值实验,实验结果表明新算法比现有的稀疏表示分类算法具有显著的识别性能和鲁棒性。In order to further improve the recognition performance of face recognition system under the condition of small sample size, we propose a novel selective ensemble face recognition algorithm based on multi-scale patch collaborative rep- resentation. Firstly, the newly proposed method has the ability to identify key characteristic sub-blocks through total varia- tion local patch weighted. Secondly, it also effectively improves the generalization ability and robustness of the ensemble classifier through selective ensemble of multi-scale patch collaborative representation. Numerical experiments were carried out on three standard face databases under different acquisition environments with a variety of light, facial expression and posture changes. The experimental results indicate that the newly proposed algorithm has better recognition performance and robustness than the existing state of the art sparse representation classification algorithm.

关 键 词：稀疏表示分类 分块协同表示 总变差加权 多尺度选择性集成 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]