基于形状索引的DoG特征结合GPRT的人脸关键点检测算法  

Shape Index Based DoG Feature and GPRT Face Key Detection Algorithm

作　　者：赵康[1] Zhao Kang(Shangqiu Vocational and Technical College, Shangqiu476100, China)

机构地区：[1]商丘职业技术学院,河南商丘476100

出　　处：《计算机测量与控制》2019年第5期203-206,211,共5页Computer Measurement &Control

基　　金：河南省高等学校重点科研项目(15A520118);河南省科技厅软科学研究计划项目(142400411213)

摘　　要：针对传统人脸对齐算法效率较低的问题,提出一种基于形状索引的高斯差分(DoG)特征与高斯过程回归树(GPRT)的人脸关键点检测算法;首先,由高斯过程回归树的内核测量两个输入之间的相似性,并表示为两个输入进入相同叶子的树木数;然后基于高斯过程回归树模型提取形状索引DoG特征,并进一步完成GPRT的特征设计;最后从局部视网膜模式中采集滤波回应来增加稳定性,实现对抗几何差异的鲁棒性;在LFPW人脸数据库上验证结果表明该方法能够取得良好的性能表现,证明了基于形状索引的DoG特征与GPRT的人脸关键点检测算法的有效性。Aiming at the low efficiency of traditional face alignment algorithm,a face index based Gaussian difference(DoG)feature and Gaussian process regression tree(GPRT)face key detection algorithm are proposed.First,the kernel between the two inputs is measured by the kernel of the Gaussian process regression tree and is expressed as the number of trees entering the same leaf for both inputs.Then,the shape index DoG feature is extracted based on the Gaussian process regression tree model,and the feature design of the GPRT is further completed.Finally,the filtering response is collected from the local retina mode to increase the stability and achieve robustness against geometric differences.The verification results on the LFPW face database show that the proposed method can achieve good performance,and proves the validity of the shape index-based DoG feature and GPRT face key detection algorithm.

关 键 词：高斯模型 人脸关键点检测 形状索引 鲁棒性 几何差异 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]