基于投影与深度学习网络的三维人脸特征点定位方法
|
摘要
标定三维人脸模型特征点对人脸识别、人脸建模等都具有重要作用。针对人脸特征点标定需要手工干预、标定特征点个数少或不准确、标定时间长等问题,提出了一种基于投影与深度学习网络的人脸三维模型特征点标定法。基于正交投影,生成人脸三维模型二维深度图与二维特征点位置,采用以卷积神经网络为主的深度学习网络模型训练测试,将深度图上特征点映射到三维人脸模型,实现眉毛、眼睛、鼻尖、嘴巴等重要区域的特征点定位。实验表明,该方法可自动标定三维人脸模型特征点,快速、准确预测足够数量特征点位置。
Calibration of the three-dimensional face model feature points has an important role on face recognition,face modeling,and so.Some problems still exist,like that Face feature points need to be manually calibrated,the number of calibration points is small or inaccurate,the calibration time is long,and so on.A feature point calibration method of face 3 Dmodel based on projection and depth learning network is proposed.Firstly,based on the orthogonal projection,the two-dimensional depth map and two-dimensional feature point position of the human face 3 Dmodel are generated.Then we use the depth learning network model,mainly based on the convolution neural network,to train and test,and map the feature points on the depth map to the 3 D human model.Finally,we achieve facial key points on eyebrows,eyes,the nose tip and the mouth.Experiments show that this method can automatically calibrate enough 3 Dface model feature points,fastly and accurately.
Calibration of the three-dimensional face model feature points has an important role on face recognition,face modeling,and so.Some problems still exist,like that Face feature points need to be manually calibrated,the number of calibration points is small or inaccurate,the calibration time is long,and so on.A feature point calibration method of face 3 Dmodel based on projection and depth learning network is proposed.Firstly,based on the orthogonal projection,the two-dimensional depth map and two-dimensional feature point position of the human face 3 Dmodel are generated.Then we use the depth learning network model,mainly based on the convolution neural network,to train and test,and map the feature points on the depth map to the 3 D human model.Finally,we achieve facial key points on eyebrows,eyes,the nose tip and the mouth.Experiments show that this method can automatically calibrate enough 3 Dface model feature points,fastly and accurately.
引文
[1]吕士文,达飞鹏,邓星.基于区域改进LBP的三维人脸识别[J].东南大学学报:自然科学版,2015(4):64-68.
[2]王健,高媛,秦品乐,等.基于改进的LBP算法的三维人脸识别[J].计算机工程与设计,2016(12):234-238.
[3]邓星,达飞鹏,杨乔生.基于自适应人脸切割的三维人脸识别算法[J].东南大学学报:自然科学版,2016(2):34-38.
[4]刘述木,杨建,陈跃.保角特征结合改进差分进化算法的三维人脸识别[J].计算机应用研究,2016(6):304-308.
[5]李江,冉君军,张克非.一种基于降噪自编码器的人脸表情识别方法[J].计算机应用研究,2016(12):329-332.
[6]黄建,李文书,高玉娟.人脸表情识别研究进展[J].计算机科学,2016(S2):132-135.
[7]XUN GONG,GUOYIN WANG.Automatic 3Dface segmentation based on facial feature extraction[C].IEEE International Conference on Industrial Technology,2006:1154-1159.
[8]贾晖,耿国华,周明全,等.基于区域离散曲率的三维网格分水岭分割[J].计算机工程与应用,2015,51(11):182-186.
[9]CHANG K I,BOWYER W,FLYNN P J.Multiple nose region matching for 3Dface recognitlon under varying facial expression[J].IEEE Transactions on Pattern Analysis and Machine Intelligence,2006,28(10):1695-1700.
[10]王密宫,陈锻生,林超.基于局部形状图的三维人脸特征点自动定位[J].计算机应用,2010(5):121-124.
[2]王健,高媛,秦品乐,等.基于改进的LBP算法的三维人脸识别[J].计算机工程与设计,2016(12):234-238.
[3]邓星,达飞鹏,杨乔生.基于自适应人脸切割的三维人脸识别算法[J].东南大学学报:自然科学版,2016(2):34-38.
[4]刘述木,杨建,陈跃.保角特征结合改进差分进化算法的三维人脸识别[J].计算机应用研究,2016(6):304-308.
[5]李江,冉君军,张克非.一种基于降噪自编码器的人脸表情识别方法[J].计算机应用研究,2016(12):329-332.
[6]黄建,李文书,高玉娟.人脸表情识别研究进展[J].计算机科学,2016(S2):132-135.
[7]XUN GONG,GUOYIN WANG.Automatic 3Dface segmentation based on facial feature extraction[C].IEEE International Conference on Industrial Technology,2006:1154-1159.
[8]贾晖,耿国华,周明全,等.基于区域离散曲率的三维网格分水岭分割[J].计算机工程与应用,2015,51(11):182-186.
[9]CHANG K I,BOWYER W,FLYNN P J.Multiple nose region matching for 3Dface recognitlon under varying facial expression[J].IEEE Transactions on Pattern Analysis and Machine Intelligence,2006,28(10):1695-1700.
[10]王密宫,陈锻生,林超.基于局部形状图的三维人脸特征点自动定位[J].计算机应用,2010(5):121-124.