Research On Expression Synthesis Algorithm Based On Generative Adversarial Networ

Facial expression synthesis refers to realizing different expression changes through neutral expressions and keeping the identity information unchanged.Facial expressions are one of the important manifestations of people’s daily communication.With the development of the field of computer vision,the research in the direction of expression synthesis technology has received extensive attention,and good results have been achieved.However,due to the complex structure of a human face and the contained rich texture information,the changes in facial expressions are greatly affected by individual differences.At present,it is still a technical difficulty to synthesize facial expression images with a strong sense of reality.Through the research on deep learning and generative adversarial network,this dissertation proposes solutions to improve the image quality of synthetic expressions and improve the change of facial texture features.The work of this dissertation mainly includes the following aspects:(1)An improved expression synthesis method based on CBAM model is proposed.The CBAM model is a kind of model considering attention mechanism,which can use limited information to display the features of the original data.The model combines channel attention and spatial attention to select important features in the input image.In the implementation process,the CBAM model is combined with the residual block of the generator in the Star GAN model for improvement,and the model is added before the residual block is skipped to improve the ability of the network to extract image features.In addition,the original batch normalization algorithm in the CBAM model is changed to instance normalization,so that invalid information can be eliminated during network training,the network training process can be simplified,and the quality of generated images can be improved while improving the training efficiency.(2)An improved expression synthesis method based on self-attention mechanism is proposed.The self-attention mechanism is an improvement of the attention mechanism,which reduces the dependence on external information,improves the ability to capture information,and can mine the correlation between image features more deeply.In the Star GAN model,a two-layer self-attention mechanism is introduced before and after the output of the hidden layer to connect different positions in the image sequence and find the relationship between them,providing key feature information for the direction of network learning and making expression changes more Targeted.The calculation area of this method is not limited to a certain image block,but can share information with long-distance features,greatly utilize the original information in the input data,model the global context,improve the image quality,and improve the expression Synthesized texture details.(3)In the process of network training,generally the discriminator is easy to be proned prematurely and then enter a state where training is more perfect,which makes the generator be unable to obtain useful information through backpropagation,and thus unable to continue training,and the network lacks stability.In view of the above problems,an improved expression synthesis method based on spectral normalization is proposed,and spectral normalization is added to the discriminators of the two improved algorithms normalization,normalizing the parameters in the network,limiting the change of its mapping function to a certain range,and enhancing the stability of the model.The comparative experiments on the Ra FD dataset and the Jaffe dataset show that the expression synthesis algorithm proposed in this paper can synthesize highquality facial expression images and improve the facial texture details of the synthesized images.