Research On Facial Expression Synthesis For Facial Reenactment

Facial expression synthesis has always been a relatively active topic.The thesis attempts to solve the task of facial expression synthesis for facial reenactment,and edits the expression of the current character on the basis of retaining the facial actions of the person in the image,maintain semantic relevance to the original image.The image generation stability under multiple scenes has put forward higher requirements.In the thesis,based on the study of the existing algorithms of facial expression synthesis,a method of facial expression synthesis based on the generation adversarial network is proposed,and it is trained and tested on Emo Vox Celeb facial expression database and Vox Celeb2 facial video database respectively to verify the effectiveness of the algorithm.Firstly,the thesis discusses the feasibility of using structural information as a condition,and proposes a network structure of dual path generator.The dual path generator has two encoding and decoding paths,which correspond to structural information and image information respectively,and adaptive learning target conditional variables to guide the synthesis process of image information.The experimental results show the effectiveness of the application of dual path generator and structural information.Secondly,the thesis discusses the defects of the conventional training structure on the unbalanced data,and puts forward separated discriminators and the interactive self-supervised learning strategy between dual paths.the strategy makes full use of the structure information which is easier to learn in the path to assist the learning process of the complicated image information.The strategy promotes the convergence of the network and improves the learning ability of the network through internal parameter check and external explicit supervision.Finally,the method uses the residual mask network to optimize the generated results by introducing the attention mechanism,combines the original input and the current synthesis results to get a higher quality synthetic image,then the whole process of facial expression synthesis is completed.