基于多点对齐的年龄估计方法  

Age estimation based on alignment of multiple landmarks

作　　者：冯倩[1] 邢玲[1] 臧婷[1] 穆国旺[1] 

机构地区：[1]河北工业大学理学院,天津300401

出　　处：《河北工业大学学报》2016年第6期41-47,共7页Journal of Hebei University of Technology

基　　金：河北省自然科学基金(E2014202124)

摘　　要：针对两眼对齐的人脸图像归一化只能做到不同图像的两眼位置对齐,而其他特征点如鼻子、嘴巴、下巴等的位置不同,提出一种基于多点对齐的人脸图像归一化算法.该算法首先标注或检测人脸图像中眼睛、鼻子、嘴巴等附近的一些特征点,利用双眼位置进行归一化操作;然后计算图像各个特征点的平均位置进行Delaunay三角剖分得到平均形状;对于双眼对齐的每幅图像,将特征点按照平均形状的连接关系连接成三角形,再通过仿射变换将每个三角形中的图像映射到平均形状对应的三角形中,即得到多点对齐的人脸图像.将该方法与基于两眼对齐的支持向量机回归年龄估计方法在FG-NET数据库上进行实验,对于常用的一些人脸特征,包括BIF,Gabor,HOG,LPQ,平均绝对误差分别降低了0.52,0.66,0.16,0.12.实验结果表明,基于多点对齐的年龄估计方法均优于基于双眼对齐的年龄估计方法.The normalization of facial images based on alignment of two eyes can ensure only the position of the eyes are aligned,and the position of other landmarks,such as nose,mouth,forehead,hin,are different generally,therefore a normalization method based on alignment of multiple landmarks is proposed.Firstly,landmarks near the eyes,nose,mouth,forehead and chin on the face images were detected or annotated,and then images were aligned based on the eyes,after that,the mean position of every landmark was calculated using normalized training images,and the Delaunay Triangulation was applied to the obtained mean landmarks,which we called mean shape of the images.In each image aligned with two eyes,the landmarks were connected into triangles with the same topology in the mean shape.Then,the image in each triangle was mapped to the corresponding triangle in mean shape with an affine transformation.After above-mentioned procedure,the image aligned with multiple landmarks was obtained.The experiments were carried out on the database of FG-NET with some common facial features,BIF,Gabor,HOG and LPQ.And with our method based on alignment of multiple landmarks,the Mean Absolute Error can be reduced by 0.52,0.66,0.16,0.12 respectively.Experiments show that our proposed method for age estimation based on alignment of multiple landmarks is more effective than the method based on alignment of two eyes.

关 键 词：计算机视觉 年龄估计 图像归一化 DELAUNAY三角剖分 仿射变换 支持向量机回归 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]