Research And Application Of Style Transfer In Generative Adversarial Networks

Style transfer is a processing method that fuses one kind of image content semantics with another kind of image style semantics,and finally forms a new image.In today's society where people like to express their thoughts and emotions through images,it has certain commercial value.At the same time,properly trained can also complete computer vision tasks such as image generation,semantic segmentation,and extended data-sets,which has research value.The existing style transfer algorithms are traditional style transfer through artificial modeling and deep learning-based style transfer technologies.The traditional style transfer is more suitable for the output of fixed textures,the versatility is not high,and the overlay effect is too rigid.The style transfer technology based on deep learning can separate the content semantic features and style semantic features of images,and the effect is much better than traditional style transfer.But it is still in the process of development.The cycle generative adversarial network combined with capsules proposed in this paper is based on the cycle generative adversarial network and introduces the concept of capsules.By combining the capsule concept and the convolutional layer,the cycle generative adversarial network structure is planned and adjusted to achieve a discriminator of capsule and convolution mixing.The purpose is to clarify the relative relationship between features and obtain stable image style transfer.The cycle generative adversarial network includes two generators and two discriminators,which are responsible for the conversion between the two image domains and the discrimination between true and false.Because the invariance caused by convolutional neural network pooling will cause the module to lose the relative relationship between various features.This paper introduces the capsule structure into the underlying network structure to achieve the same property of the feature.After experiments,a capsule and convolution discriminators are used to retain the relative relationship between the features,and the discrimination's accuracy is greatly improved.This paper improves the loss function.While introducing the capsule structure,the concept of relativity is introduced on the basis of the least square loss function,and the loss functions of the generator and discriminator are obtained after finishing.Then construct the total loss function according to the network structure and optimization purpose.Experiments show that after introducing relativity,the network proposed in this paper can get better results faster in training.Through experiments,compared with the existing methods such as Bi GAN,Co GAN,and Cycle GAN,the cycle generative adversarial network combined with capsules used in this paper is better and closer to the real image.