Abstract: Aiming at the problems of mostly relying on data source driver,low generation efficiency and poor authenticity of the existing facial expression synthesis methods,this paper proposes a new method for expression animation synthesis based on the improved CycleGan model and region segmentation.This new method can synthesize new expression in real time and has good stability and robustness.The proposed method constructs a new covariance constraint in the cycle consistent loss function of the traditional CycleGan model,which can effectively avoid color anomalies and image blurring in generation of new expression images.The idea of zonal training is put forward.The Dlib face recognition database is used to detect the key points of the face images.The detected key feature points are used to segment the face in domain source and target domain into four zones:left eye,right eye,mouth and the rest of the face.The improved CycleGan model is used to train each region separately,and finally the training results are weighted and fused into the final new expression image.The zonal training further enhances the authenticity of expression synthesis.The experimental data comes from the SAVEE database,and the experimental results are presented with python 3.4 software under the Tensorflow framework.Experiments show that the new method can directly generate real and natu-ral new expression sequences in real time on the original facial expression sequence without data source driver.Furthermore,for the voice video,it can effectively ensure the synchronization between the generated facial expression sequence and the source audio.

Key words: Facial expression synthesis, Region segmentation, CycleGan, Covariance constraint, Deep learning