Research On Image Real-time Style Sransfer Based On Group Normalization

The traditional image style migration method needs manual modeling,and the commonly used normalization method is batch normalization,which results in slow convergence speed of training loss function.At the same time,the image texture is unstable in the migration process,and the texture shape changes.To solve these problems,this paper proposes a real-time style transfer method based on group normalization.Firstly,based on the real-time style transfer model of perceptual loss function,a more concise and efficient image generation network is designed.Secondly,in the normalization layer of image generation network,group normalization is used instead of batch normalization to speed up the convergence of loss function.Finally,histogram loss is added to the style loss function as a constraint.Through the above methods,the aim is to train an efficient image real-time style transfer model.The experimental results show that the convergence speed of the loss function of the proposed method is faster than that of the batch normalization method,and the performance of the model will not be affected by the change of batch size.Adding histogram loss to style loss function can effectively solve the problem of unstable texture migration.Through subjective comparison and investigation,the proposed method can produce satisfactory visual effects.In terms of migration speed,it can meet the needs of real-time migration,and has been applied in the video style migration.There are 44 figures,2 tables and 50 references in this paper.