Research And Implementation Of Image Style Transfer Algorithm Based On Generative Adversarial Networ

The purpose of image style transfer is to render image contents in the art style of another image.The basic concept of neural style transfer is to transform styles into distributions in the convolutional neural network feature space,thus achieving the desired style by matching its feature distributions.Generative Adversarial Networks also show good potential in image-to-image transformation.Artistic style transfer is to synthesize an image sharing structure similarity of the content image and reflecting the style of the artistic style.Here,artistic style implies the genre of paintings by the artist,and the artistic images refer to a set of images created by the same artist,and each image has a unique character.This article comprehensively summarizes style transfer algorithms based on NST and GANs,and makes a comparison study of these algorithms.Finally,the whole paper is summarized and some deficiencies in future research are pointed out.These methods can not only be applied in the field of art,but also be extended to real life,with broad application prospects.(1)A method for solving the style transfer problem of salient regions in non-paired datasets is proposed in this paper.Traditional style transfer methods fail to capture the geometric or structural patterns in the salient regions,leading to the loss of details in structured regions and smoothing of regions.To address this issue,this paper proposes a salient feature generation adversarial network,which introduces a salient feature network to be trained simultaneously with the generator to increase the suppression of salient regions and generate consistent results.The salient feature network has two functions: to provide constraints for content loss to increase the suppression of salient regions;to provide salient features to the generator to generate consistent results.In addition,two new loss functions are proposed to optimize the generator and the salient feature network.This method retains details in the salient regions,with an IS score improvement of approximately 1.2%,and a FID score reduction of nearly 0.9%.(2)In this paper,a "style flow" processing unit is also designed to prevent content leakage during general style conversion.It is demonstrated through experiments that these methods can generate high-quality style transferring results and perform excellently in avoiding content leakage.The style flow consists of a reversible neural flow and a biasfree feature transmission module,which supports both forward and backward transmission and runs within the mapping-transferring-restoring system.The forward transmission maps the input image to deep features while the backward transmission remaps the deep features back to the input image in a lossless and bias-free manner.Through qualitative and quantitative comparison with several leading methods,the proposed method generates suitable style transferring results with an SSIM score improved by about 6.5% and a PSNR score improved by about 0.1%,thus demonstrating the superiority of the proposed method.