基于核函数因素分解模型的表情合成与识别

Facial expression synthesis and recognition using a kernel-based factorization model

人脸图像合成是新一代人机交互中的重要技术。传统的三维模型加生理模型的方法可以生成真实的人脸表情图像,但是其中的计算复杂度很高。该文提出了一种基于样本的方法,将不同的人和不同的表情看作影响人脸表情图像的两种变化因素,利用因素分解模型巧妙地进行人脸表情图像合成。同时,分析了因素分解模型获得的身份子空间和表情子空间的特点,提出了一种在子空间中利用余弦距离进行身份和表情识别的新思路。从实验结果来看,这里提出的方法可以仅利用一张训练集内、外的人脸图像合成出该人在不同表情下逼真的脸部表情图像,同时可以合成库内的人在新表情下的表情图像。

Facial expression synthesis is an important technique in human computer interactions. The traditional 3-D reconstruction method using a face physiology model requires very complex computations, therefore, a sample-based method was developed to synthesize realistic facial expressions. To translate facial expressions, the model treats ＂human identity＂ and ＂facial expression＂ as two factors influencing the face appearance, with a kernel-based bilinear factorization model used to decouple their interactions. The distribution features of the subspaces obtained by factorization were used to develop a method to recognize people and expressions. Test results show that the method can successfully translate realistic facial expressions from one person to another.