摘要
关键点检测是三维人脸识别过程中非常重要的一步,为了提高关键点检测的准确度,提出了一种多特征相结合的三维人脸关键点检测方法。首先对训练集中的三维人脸手工标记关键点,计算三维人脸上每一点的不同特征值,得到每类关键点关于每个特征的均值和方差,其次对关键点和非关键点上的特征进行线性判别分析,得到与每个关键点相关的分值加权向量,将前面得到的均值,方差以及分值加权向量作为线下训练的结果输出。最后对于一个输入模型,结合线下训练的结果得到每个关键点的候选点,利用这些候选点构建人脸结构模型。再根据绝对距离约束,相对位置约束,FLM模型一致性分类,自旋图描述符等方法确定最终的关键点。实验部分,从CASIA-3DFaceV1和FRGC V2.0数据库中选不同姿态,不同表情,姿态与表情混合的三个数据集,对其进行关键点检测。实验结果表明,不同姿态的检测率为94.5%,不同表情的检测率为94%,和其他文献相比,检测率平均提高了20%,并且有着较高的运算效率。
Key point detection plays an important role in the process of 3D face recognition.In order to improve the accuracy of key points detection,a new method of 3D face key points detection based on multi feature is proposed.Firstly,the key points of the 3 dface of the training set are manually marked,and the different eigenvalues of each point of 3D face are calculated,and the mean and variance of each feature are obtained for each key point.Secondly,the linear discriminant analysis is carried out on the characteristics of key points and non-critical points,thus get the score-weighted vector associated with each key point.The mean,variance,and score-weighted vectors of the previous ones are the output of the offline training.Finally,for an input model,we can get the candidate points ofeach key point combined the results of offline training,and construct the face structure model using these candidate points.According to the absolute distance constraint,relative position constraint,FLM model consistency classification,spin map and other methods we can determine the final point.In the experimental part,we select the three data sets of different postures,different expressions,gestures and expressions mixed from CASIA-3 DFaceV1 and FRGC V2.0 database to detect the key points.Experiment results show that the detection rate of different gestures was 94.5%,and the detection rate of different expressions was 94%.Compared with other literatures,the detection rate increases by 20% on average.In addition,the algorithm has higher computing efficiency.
作者
冯超
陈清江
FENG Chao;CHEN Qing-jiang(School of Resources &Surveying Engineering,Shaanxi Energy Institute,Xianyang 712000,China;School of Science,Xi’an University of Architecture and Technology,Xi'an 710055,China)
出处
《液晶与显示》
CAS
CSCD
北大核心
2018年第4期306-316,共11页
Chinese Journal of Liquid Crystals and Displays
基金
国家自然科学基金(No.61403298)
陕西省自然科学基金(No.2015JM1024)~~
关键词
有效能量
鼻尖点检测
姿态校正
测地距离
迭代最近点
主成分分析
分类器
effective energy
nasal tip detection
posture correction
geodesic distance
iterative closest point
principal component analysis
classifier