均值漂移LDA优化FSIF的三维人脸识别  

FSFF optimized by shifting mean LDA for three-dimensional face recognition

作　　者：徐屹[1] 樊晓平[2] 廖志芳[3] 

机构地区：[1]湖南工业职业技术学院信息工程系,长沙410208 [2]中南大学信息科学与工程学院,长沙410012 [3]中南大学软件工程学院,长沙410012

出　　处：《计算机工程与应用》2014年第5期160-164,194,共6页Computer Engineering and Applications

基　　金：国家科技支撑计划项目(No.2012BAH08B00);湖南省自然科学基金(No.06JJ50143)

摘　　要：针对传统的三维人脸识别算法受光照、姿态、表情及场景变化影响导致耗时过多及成本过高的问题,提出了一种基于均值漂移线性判别分析优化尺度不变特征融合(FSIF)算法。使用均值漂移线性判别分析找到五个类似于查询人脸的最佳候选类;利用尺度不变特征融合提取出候选人脸及查询人脸的融合特征描述符,并进行特征匹配得到目标人脸;根据特征描述符的匹配关键点数目完成人脸的识别。在USCD/Honda、FRGC v2及自己搜集的人脸数据集上的实验结果表明,该算法解决了降低FSIF人脸识别的计算复杂度,并在不降低识别性能的前提下大大地节约了成本,相比几种较为先进的三维人脸识别算法,该算法取得了更好的识别效果。Traditional three-dimensional face recognition algorithms take too long time and high costs due to variations of illustration, poses, expression and scene, so a fusion of scale invariant features algorithm optimized by shifting mean linear discriminant analysis is proposed. Shifting mean linear discriminant analysis is used to find five optimal candidate classes similar to query faces. Fusion of scale invariant features is applied in extracting fusion feature descriptors of candidate faces and querying face, after which feature matching is done so as to get objective face. Matching point number of the feature descriptors is referred to finish face recognition. Experimental results on USCD/Honda、FRGC v2 and face dataset collected by self show that proposed algorithm decreases the computational time of FSIF-based face recognition and saves costs clearly without decreasing the recognition performance. It has better recognition efficiency than several advanced three-dimensional face recognition algorithms.

关 键 词：三维人脸识别 尺度不变特征融合 均值漂移线性判别分析 特征描述符 查询人脸 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]