Research On Image Style Transfer Methods Between Real Scenes Based On Generative Adversarial Network

With the rapid development of electronic technology and equipment,image or video technology has also been rapidly developed.And with the pursuit of diversity in entertainment life,images are everywhere in normal social life.This has promoted the rapid development of related technologies in the field of computer vision,and deep learning technology has also risen.As a new hot research direction in the field of deep learning,Generative Adversarial Network(GAN)has also quickly found a research direction in the computer field.GAN has made outstanding research results in many image tasks,such as: image generation,attribute editing,image style transfer.This paper focuses on the application of GAN in the task of image style transfer.The task of image style transfer is mainly to extract the style features of the target style domain image,and then perform style rendering on the input image to obtain a new image in the target style domain.Image style transfer has many applications in face attribute editing,pose transformation,and scene style transformation.This paper mainly uses image style transfer to transfer the style of real scenes(such as weather,seasons).The specific work is as follows:First,this paper proposes an unsupervised dual-domain image-to-image translation method-SA-UNIT.UNIT is a GAN network based on a variational autoencoder.It compresses the input image through the encoder to obtain the implicit variables with the characteristics of the input image,thereby improving the generation ability of GAN.In this paper,the self-attention model is introduced into the UNIT model,and the module is added to the convolutional layer of the generator to redistribute the weights so that the next layer of neurons can learn more useful features.For the normalization technology,this paper introduces the adaptive layer normalization technology,and uses the appropriate normalization technology in different network layers by optimizing the updated parameters.Experiments have proved that SA-UNIT has achieved good results in the image style transfer task,and the quality of the generated image has also been improved to a certain extent.Secondly,this paper proposes an unsupervised multi-modal image-to-image translation method-MMUNIT.The style in each image domain is not completely unified,and it shows different styles due to factors such as light and weather.In response to this problem,this article improves on the MUNIT network and introduces a Mapping Network to further decouple implicit variables.In addition,the image is added to the domain label code,and the multi-task discriminator is used to discriminate the domain label,so that the network can realize the translation of multi-domain image to image.Experiments prove that MMUIT can realize the style transfer of multi-domain images,and generate different modal style images in the same domain,which verifies the effectiveness of MMUIT.