基于多尺度SDM模型的人脸对齐算法  

Face Alignment Based on Multi-scale SDM Model

作　　者：姚文韬[1] 沈春锋[1] 顾志松[1] 董文生[1] 

机构地区：[1]上海宝信软件股份有限公司,上海201203

出　　处：《控制工程》2015年第S1期110-114,共5页Control Engineering of China

基　　金：上海市科委科技攻关项目:基于全景视觉的室内公共场合监控系统研发与应用(13511502500);上海市科委科技人才计划项目:工业海量数据分析系统的研发(13QB1400700)

摘　　要：提出一种基于改进的多尺度SDM模型的人脸对齐算法。通过分析SDM模型的缺陷,将SDM模型扩展到多个图像尺度上,利用SIFT特征计算特征点处人脸特征,计算出每个特征点的位移方向,使得人脸特征点能够从初始位置逐渐向最优位置收敛。从粗到细的多个尺度使得特征点在迭代初期具有较大的步长,加速迭代进程,且避免陷入局部最优;而后期步长较小,容易实现特征点的精确定位。实验结果表明,算法在人脸对齐精度和计算效率上均优于现有的SDM模型。算法在4个尺度,每个尺度只迭代1步的情况下,取得了0.030 8的均方根误差,相比SDM模型的准确率提高了11.2%,同时检测速度比SDM模型加快了21.8%。An improved multi-scale version of SDM model is proposed in this paper. Based on the defect of SDM model, it is extended to run on multiple image scales in our model. SIFT feature is used to compute face feature at each face point. Moving direction for each face point is calculated using our model, which drive face points from initialized location to the optimized location. Multiple scales using coarse-to-fine strategy will lead to large steps at the beginning of iteration, which accelerates the iteration and avoid local optimum, while small steps at the end of iteration, which makes face point localization more accurate. Experimental results show our model is better than SDM with respect to both accuracy and speed. Our method has RMSE error of 0.0308 using 4 scales and 1 iteration for each scale, which is 11.2 % more accurate than SDM model, and 21.8 % faster than SDM model.

关 键 词：人脸对齐 人脸特征点定位 SDM模型 多尺度 SIFT特征 梯度下降 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]