Research On Image Style Transfer Based On Deep Learning

Image style transfer is a new research field of multidisciplinary integration,which has high academic and commercial value.The early methods mostly adopt modeling,which takes a long time and has poor effect.With the rise of artificial intelligence,people find the potential of neural network in dealing with this problem,and gradually focus on the method based on deep learning.There are many style transfer algorithms based on deep learning.This paper mainly studies the algorithm of feature fusion.In the course of the study,we found that there were deficiencies,In order to better complete the task of transfer,this paper has proposed two new improved algorithms based on the analysis of the existing style transfer algorithms.The specific work is as follows:(1)The style transfer algorithms based on deep learning in recent years have been sorted out,and several classic algorithms have been studied with emphasis,and they have been divided into traditional algorithms and general algorithms.Then the principles of these classical algorithms have been discussed,and then the experimental simulation has been carried out to compare the advantages and disadvantages of the algorithms through qualitative and quantitative indicators.(2)Aiming at the shortcomings of the traditional class algorithm,an improved fast two-stream residual algorithm has been proposed.Based on the feedforward model,a two-stream residual structure was added,and a weight adjustment mechanism and a super-resolution mechanism were added to make it have the functions of real-time,multi-style transfer,fine-tuning weight and super-resolution reconstruction under the premise of guaranteeing the rendering effect.(3)An improved nonparametric projection algorithm has been proposed to overcome the deficiency of the general algorithm.By using image codec,style feature sequence was rearranged according to content feature sequence,Kullback-Leibler divergence and cosine similarity loss was introduced to make it have better rendering effect of texture and style under the premise of ensuring universality.