摘要
针对人脸姿态、光照、人脸表情和周围环境发生变化时人脸识别精度不高的问题，提出一种基于稀疏多级正则化剪切波网络的鲁棒人脸识别算法。该算法创新点有二：第一，采用剪切波网络（SN）多尺度方向框架去提取人脸特征，其优势在于剪切波框架具有高度稀疏表示，有利于的提取图像中鲁棒的几何内容；第二，采用了一种完善的多任务稀疏学习，利用识别阶段中正则化参数改变多个共享任务间的关系。实验结果表明：在可控数据库上，提出算法的识别率达98．5％；在不可控数据库上，比其他算法中的最好结果高5％左右。
As the recognition accuracy is not high when the face pose, illumination, facial expression orsurrounding environment changes. A new robust face recognition algorithm based on sparse multi-levelregularization shear-wave networks （SMRSN） is proposed. The main innovation is twofold： Firstly,shear-wave networks （SN） multi-scale framework is used to extract facial features. And the advantages of SNis that this framework can be highly sparse represented, which is very beneficial for extracting robustgeometric contents; Secondly, a complete multitasking sparse learning method is used, which can change therelationship between share tasks by the regularization paras. Experimental results show that the recognitionrate of the proposed algorithm can reach 98.5 percent in the controlled database; and in the uncontrolleddatabase, the result is higher than that of other algorithms by 5 percent.
出处
《控制工程》
CSCD
北大核心
2017年第1期106-111,共6页
Control Engineering of China
基金
江苏省高校自然科学研究项目(14KJB520036)
关键词
人脸识别
剪切波网络
稀疏表示
多任务稀疏学习
鲁棒性
不可控数据库
Face recognition
shear-wave network
sparse representation
multitask sparse learning
robustness
uncontrolled database
