面向人脸光照合成的光照参数精确估计算法  

An Accurate Algorithm to Estiamte Illumination Parameter for Facial Relighting Synthesis

作　　者：刘剑毅[1] 何苗[1] 马瑶[1] 刘跃虎[1] 

机构地区：[1]西安交通大学人工智能与机器人研究所,西安710049

出　　处：《西安交通大学学报》2013年第6期18-24,共7页Journal of Xi'an Jiaotong University

基　　金：国家自然科学基金资助项目(60903122;91120009);教育部高等学校博士学科点专项科研基金资助项目(20090201110029)

摘　　要：为提高人脸图像光照合成的精度,提出了一种基于商图像的光照参数估计新算法。首先分析并指出传统商图像算法中存在恒定反射系数假设难以严格成立这一缺陷,它给光照参数估计引入了固有误差;然后在参数化光照子空间框架下设计了一个改进的目标函数,用商图像本身去替代原算法中的反射系数比,从而弱化了假设条件的限制,允许人脸各像素点拥有独立的反射系数;设计了一个迭代算法,通过线性系统来求解目标函数,能够实现对光照参数的高精度估计。在OFD人脸库和AI&R人脸库上的实验结果表明,所提光照参数估计算法具有良好的收敛性,光照参数估计精度的迹商指标可以获得25.49～43.09dB的提升,光照合成图像中的峰值信噪比可以获得4.51～9.24dB的提升。A novel algorithm to estimate illumination parameter based on quotient image（QI） is proposed to improve the accuracy of facial image illumination synthesis.Analysis indicates that the hypothesis of uniform albedos in traditional QI algorithm can not be strictly hold,and that inherent error in illumination parameter estimation will be incurred.Then,an improved objective function is derived in the parametric illumination subspace framework,and the albedo ratios is replaced by the quotient image itself,so that the impacts of the hypothesis is weakened,and independent albedo per facial pixel can be permitted.An iterative algorithm is designed to solve the objective function through linear system,and illumination parameter can be estimated at high accuracy.Experiments on the oriental face database（OFD） and the AIR face database show that the proposed algorithm has good convergence ability,and that the trace ratio for illumination estimation and the peak-signal-to-noise-ratio in illumination synthesis images are improved by 25.49~43.09 dB and 4.51~9.24 dB,respectively.

关 键 词：人脸图像 光照参数 商图像 光照重绘 

分 类 号：TN914.42[电子电信—通信与信息系统]