Research On Image Style Transfer Based On Deep Hybrid Generative Model

Image style transfer is a task to transform the input image from one region's style to another region's style,which presents a great challenge for current deep learning network structure and algorithms.One of the most effective approaches for image style transfer is using deep hybrid generative model.Specifically,image style transfer is also a kind of image generation,the purpose of which is to input an image and generate another image.For this reason,this paper proposes an image style transfer model based on the deep hybrid generative model,which improves the quality of the translated images.Deep hybrid generative model are mainly from the combination of the auto-encoder and generative adversarial network,this paper,based on the two different basic tasks of image translation,unsupervised and supervised image style transfer,respectively put forward the enhanced cycle-consistent generative adversarial networks processing unsupervised image style transfer and cross-domain auto-encoder network processing supervised image style transfer.Cycle-consistent generative adversarial networks is a general unsupervised image style transfer framework.This paper replaces the basic network block and introduce the VGG network to deepen network structure and capture the image style represented by the high-dimensional feature space,so as to improve the quality of generated images.The cross-domain auto-encoder network improves the quality of generated images by using two auto-encoders to separate and merge the internal features.In this paper,a series of experiments on the standard datasets demonstrate the effectiveness of the enhanced cycle-consistent generative adversarial networks and the cross-domain autoencoder generative adversarial network.The experimental results show that the enhanced cycleconsistent generative adversarial networks is better than the existing model in the quality of generated images and evaluation index,and cross-domain auto-encoder generative adversarial network can effectively solve the task of supervised image translation...