Research And Application Of Neural Style Transfer Technique In Image Diversification

Now more and more artificial intelligence technology is used in various industries,image generation technology is becoming more and more mature.The use of artificial intelligence technology to adjust,beautify and transform the style of photos has high research value at present.Popular apps such as Douyin and Bilibili can beautify photos,which shows the commercial value of the research.However,nowadays the mainstream style migration algorithm has the following problems:the semantics of the generated images are easy to be lost,the training process is difficult,and the style is relatively simple.In this paper,a neural style transfer algorithm for face is proposed,which can generate diversified high-quality style transfer images.Specific research work is as follows:(1)A multi-domain neural style transfer method based on star GANv2 is proposed by studying the structure of existing image style transfer algorithms,can be used for face attribute transformation.Modify the generator structure and discriminator structure in the original network structure,and replace the residual module in the generator with a dense convolution module,so as to retain the feature information of the input image more?at the same time,a random noise module is introduced into the down-sampling network structure to improve the diversity and authenticity of the generated images?the frequency spectrum normalization structure is introduced into the discriminator network structure to ensure the 1-Lipschitz constraint in the generated adversariant network.(2)On the basis of the research in the previous chapter,we continue to carry out research,aiming at the conversion of the whole image to multiple different style domains,and improve the generator and discriminator:Firstly,distributed shift convolution is used to replace the convolution process in the generator to reduce the training time and model size?then,the attention mechanism is introduced into the generator network to Ignore the background information in the image and focus on the detailed features of the face?the double discriminator network structure is introduced,and the Markov discriminator is added on the basis of the spectrum normalization discriminator to constrain the image generation process.(3)Finally,a diversified image style transfer system is designed.The system will be the algorithm proposed in this thesis for the actual demonstration,so that users can quickly experience the algorithm provides the image style transfer service,easy to test and improve the algorithm,also reflects the actual value of this thesis.