作　　者： ; ; ; ;

机构地区： 中南大学信息科学与工程学院

出　　处： 《系统仿真学报》 2009年第8期2229-2234,共6页

摘　　要： AAM(Active Appearance Models,主动表观模型)是一种定位人脸特征点的有效方法,它由人脸动态表观建模和拟合算法两部分组成。在多种拟合算法中,投影式反向组合算法(Project-Out Inverse Compositional Algorithm)具有快速高效的特点。但当人脸的某部分被遮掩时,算法的精度会明显下降。提出一种采用逐层细分掩模消除干扰的正规化反向组合算法,该算法既保留了反向组合算法快速高效的优点又提高了算法处理遮掩的能力。实验结果表明:在采用标准的IMM人脸库作为训练集的情况下,当脸部被遮掩0%-30%时,算法能够保证定位特征点的标准误差值介于0.01-0.1之间。 Active Appearance Models （AAM） is an efficient method for facial features location, which includes active appearance models and fitting algorithm. Within all kinds of fitting algorithms, the Project-Out Inverse Compositional Algorithm is one of the most fast and efficient algorithms. However, when the partial face is occluded, the accuracy and efficiency of algorithm become worse. An efficient Robust Normalization Inverse Compositional Algorithm with layered mask subdivision to eliminate disturbance is proposed. The algorithm not only keeps the superiority of the original lnverse Composition Algorithm but also enhances the ability of processing occlusion. When IMM Face Database is chosen as training set, the experimental results show that RMS point errors of locating features are always in the range of 0.01-0.1 when 0%-30% face is oceluded.

关 键 词： 人脸特征点定位 遮掩 掩模 主动表观模型 反向组合算法

领　　域： [自动化与计算机技术] [自动化与计算机技术]