Research On Face Generation And Editing Algorithm Based On Generative Adversial Network

With the advent of computer technology boom,face generation and editing based on deep learning has received extensive attention and research.At present,in the field of face generation,the method based on generative adversarial network can generate clear faces and realize face editing,but there are still many problems to be solved in the field of face generation and editing in practical applications.For example,when the training data is insufficient,the training of the generated model is unstable,the pictures generated in complex scenes have obvious flaws,the irrelevant content is incorrectly modified in face editing,the editing attributes are coupled,and the editing effect and real-time performance cannot be achieved at the same time.In order to solve the above problems,this thesis studies face generation and attribute editing based on GAN and proposes a new face generation and editing method.The main work of this thesis is as follows:1.Optimize the face generation model from multiple dimensions.(1)A face generation model that introduces residual self-attention mechanism is proposed to solve the phenomenon of obvious defects in generated pictures.The experimental results show the importance of attention mechanism in the generation model.(2)It is proposed to introduce a data augmentation method into the discriminator to expand the diversity of data without the influence of data transformation,relieve the strong dependence of the generative model on the training data,and realize the training stability under the condition of insufficient data set.(3)Introduce a contrastive loss function to increase the learning of the discriminator’s ability to represent data,reduce the differential impact of data enhancement on the discriminator,and enhance the discriminator’s ability to distinguish between true and false data,thereby promoting the generator to learn more abundantly picture information.2.Optimize the steps of the face editing process one by one.Face editing includes face reconstruction,editing space selection and editing direction establishment.(1)A residual iterative encoder is proposed,which uses a multi-step learning encoder with varying amounts to ensure the authenticity of the face reconstruction effect.Compared with traditional iterative optimization,it can greatly reduce the reconstruction time with a slight loss of accuracy.(2)Determine the optimal editing space to ensure that various attributes have a high degree of decoupling to the greatest extent,and improve the effect of face editing.(3)Propose an unsupervised editing direction establishment scheme.With the help of CLIP’s image-text vector consistency space,the establishment of an adaptive semantic editing direction is realized,and the editing direction is adaptively searched and given specific semantics,without manual identification of the editing direction.Through the optimization of the above two points,the generation of realistic faces under limited data is realized,and the speed and quality of face editing are improved.