多特征融合的小样本人脸检索  

Method of multiple features fusion on sparse samples face retrieval

作　　者：沈旭东[1] 蒋婷[1] 赵斌[1] 朱明[1] 

机构地区：[1]中国科学技术大学自动化系,安徽省合肥市230027

出　　处：《电子技术（上海）》2016年第6期20-23,7,共5页Electronic Technology

基　　金：面向NGB的互联网视频访问控制应用示范(KGZD-EW-103-5(5))中国科学院"NGB有线无线融合应用"重点部署项目子课题;安徽省科技攻关项目2014互联网基础信息智能搜索与数据挖掘应用平台(1301b042012)

摘　　要：在很多现实场合下,人脸检索的训练样本收集难度较大,甚至存在只有单个样本的情况。因此,找到一种有效的人脸特征描述方法成为解决此类稀疏样本人脸检索问题的关键所在。文章提出一种Gabor,LGBP,LPQ三种特征融合的人脸特征描述方法。将人脸按五官进行分块,分别提取Gabor,LGBP和LPQ特征,再通过融合得到最终的融合特征来表征每一个人脸。实验表明,在FERET,AT&T数据集以及现实采集人脸数据集上,该融合特征检索性能与当前常用的特征相比有了一定程度的提高。Face retrieval training samples is difficult to obtain in many situations. Even only single sample can get. Therefore, an effective way to describe the human face becomes more important. We have fused Gabor feature, LGBP feature and LPQ feature to describe face in this paper. Faces are divided into blocks according to features, than fusion features obtained to characterize each individual face. Taken together, our results suggest that fused features for data retrieval have a certain improvement compared with the current popular features in FERET, ATT database and our dataset.

关 键 词：特征融合 小样本人脸 人脸检索 GABOR LGBP LPQ 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]