Improved Style Transfer Based On Feature Fusion And Semantic Matching

As Gatys et al.Demonstrated the non-negligible role of Convolutional Neural Networks(CNN)in creating artistic images in 2016 by separating and recombining image content and styles,research on the transfer of image styles has become increasingly fierce.Although many phases have been achieved in the field of image style transfer,there are also two important issues that need to be resolved.First of all,the existing image style transfer method has the problem of imbalance between the content image and the style image generated in the transfer result,which results in the transfer result tending to the content image or the style image.Secondly,the semantic information is not considered in the image migration,and there is a problem of mutual interference between different semantics of the content images,which leads to the content heterogeneity in the final migration result.In response to the above problems,this paper focuses on the two existing The problem has been studied,the main research work is as follows:(1)For the balance problem,it is determined through research that deep FCN is used to retain semantic content,and shallow FCN is used to faithfully learn styles.On this basis,an end-to-end dual-stream image style transfer network is proposed for balanced stylization The content and style in the image.(2)In order to balance the contribution balance of the content and style feature layers,an adaptive change balance weighting factor and feature fusion method is proposed,and its output is adaptively subjected to feature fusion and connection,and fed to the decoder to generate styles image.(3)Aiming at the problem of semantic style confusion in migration results,a study was conducted to combine semantic segmentation and image migration algorithms,and a migration network model combining image semantic matching was proposed to improve the existing Some image style transfer algorithms.This method does not make the migrated images have the problem of confusing semantic styles,and improves the image quality,and realizes that a content image is integrated into multiple styles.After studying the two issues of content and style balance and semantic style confusion in the transfer of image styles,simulation experiments were carried out on this basis.The experiment proved that the method in this paper can effectively solve the balance problem and the resulting semantic confusion problem.