人脸表情识别与合成系统模型设计

Model designing of facial expression recognition and synthesis system

本文提出了一种新的用于人脸表情识别与合成的情感模型,该模型是基于已泛化的和非线性映射关系的五层神经网络。模型的输入和输出层有相同数目的运动单元,在中间层可以实现特征的映射和情感空间的构造。从输入层到中间层的映射是表情识别,从中间层到输出层的映射是根据情感值进行表情合成。神经网络的训练采用典型的6种表情作为训练样本,最后通过实验证明了该模型在进行表情识别与合成时的可行性。

This paper presents a new emotion model which gives a criteria to decide human＇s emotion condition from the face image. Our emorion model is based on 5-layered neural network which has generalization and non-linear mapping performance. Both input and output layer has same number of units. So identity mapping can be realized and emotion space can he constructed in the mid- dle-layer （3rd layer）. The mapping from input layer to middle layer means emotion recognition and that from middle layer to output layer corresponds to expression synthesis from the emotion value. Training is performed by typical 6 emotion patterns which are expressed by expression parameters. Subjective test of this emotion space proves the propriety of this model.