保持生理特征的人脸编辑

Physiological Feature-preserved Facial Modelling

提出了一种保持生理特征的交互式人脸编辑方法。采用控制点分层策略,即以用户直接操作的控制点对(称为主控制点对)为输入层,其他控制点对(称为次控制点对)为输出层,建立人工神经网络;然后采用误差反向传播法(Error Back Propagation)学习,从而建立主、次控制点之间的约束关系;最后通过输出层将编辑信息在模型中进行插值。该编辑结果可以应用到具有相同拓扑的任意人脸模型上。实验结果表明,采用分层控制的方法不仅保持了编辑操作的方便性、精确性,同时还保持了人脸生理特征的真实性。

In this paper we present an approach for physiological feature preserved facial modelling. We employ a multilayered strategy, where user directly handled control points（ called primary control points） are taken as an input layer and the other control points（ called secondary control points） are taken as an output layer. An artificial neural network is built upon these layers. A set of face models is used to train the network using Error BackPropagation approach. The trained constraints are then transferred to the output layer to guide the interpolation of editing information over the facial model. The result can apply to any face models with the same topology. Experimental results show that the proposed multi-layered approach not only provides efficient and exact editing operation, but also preserves the fidelity of the physiological feature of facial models.