Research Image Style Transfer Algorithm Based On Generative Adversarial Networks

Image style transfer is a research hotspot in the field of image processing.Image style transfer is to transfer a common image through the technology of style transfer,make its style change under the premise of the content remains unchanged,and generate an image with another style.Image style transfer methods are mainly divided into the traditional image style transfer method and deep learning image style transfer method.The traditional image style transfer method is not effective and can not be applied to people's life.As computer vision,digital image processing,the rapid development of machine learning,in-depth study of the image style migration method received extensive attention of the researchers,most of these methods is on the premise of paired data sets,using convolution migration of neural network to realize image style,but pairs of data set is more difficult to obtain,and the high cost.Therefore,it is of great significance to study the style migration algorithm of unpaired image implementation.This paper is mainly based on the image style migration algorithm realized by Cycle GAN for improvement.The main contents of the improvement include the following four aspects.(1)The auto-encoder and variational auto-encoder network models are added to the network model in this paper.The coding part of the autocoder is used to extract the image content features and the variational autocoder is used to extract the image style features.By combining and modifying the extracted style features and content features,the model is guided to combine and modify the image style and content,and a more realistic stylized image is achieved by adjusting the weight of content loss function and style loss function.(2)The discriminant network in this paper used multi-scale discriminator.The multi-scale discriminator is used to force the sample generated by the generator to be more realistic and approximate to the target image,thus improving the effect of image style migration.(3)The training of the network model in this paper adds an adaptive instance normalization method.The adaptive instance normalization method can accelerate the stylization of the image and make the generated image retain strong style.(4)In the generation network of the network model in this paper,the traditional deconvolution operation is replaced by image resization and convolution operation.When the image is generated by generation network,deconvolution will produce checkerboard effect and the generated image is not clear.Using the nearest neighbor interpolation method to enlarge the image and then carry out convolution operation can improve the quality of image style migration and avoid the checkerboard effect.In order to verify the rationality and effectiveness of the improved image style migrationalgorithm in this paper,the results of style migration generated by different methods are compared and evaluated by combining subjective evaluation method with objective evaluation method.The final experimental renderings and objective evaluation indexes verify the rationality and effectiveness of the image style migration algorithm in this paper.