Research On Image Style Transfer Method Based On Deep Learning

In the technology world,deep learning is used in different ways to achieve specific goals within specific topics.Image style transfer is a hot research field in computer vision.Image style transfer technology is to learn the semantic relationship between content image features and style image features,and then to recombine the content features and style features in space.In the field of image stylization,the research of image stylization based on deep learning has attracted more and more researchers' attention.The development of image stylization research brings more fun and convenience to the public's life,involving life,work and other aspects.Image stylization is a form of artistic expression,because its input and output are pictures,which reproduce the artists' painting techniques and generate a considerable artistic work.Inspired by the CIN style transfer model proposed by Dumoulin et al.,this paper proposes a new stylization model.Firstly,this paper proposes an image stylization generation model that introduces a histogram matching layer.The combination of deep learning and histogram matching aims to provide a method for real-time image style transfer without fixing the style image in the model.Not limited by the predefined style set,it can adapt to more style images.Histogram matching refers to matching the histogram of the characteristics of the content image with the histogram of the characteristics of the style image,so that the content image exhibits the same style characteristics as the style image.Compared with the CINbased style model,the experimental results confirm that the method in this paper can adapt to more style images,and the generated picture quality is better than the based model of CIN.Secondly,the idea of generating adversarial network is introduced into the new model,and a discriminator is added to the new model.The image transfer network which introduces a stylized model of histogram matching layer exists as the generator of GAN.The image style transfer network and the discriminator are alternately trained to improve the generation ability of the image style transfer network,and at the same time,it is also helpful to improve the recognition ability of the discriminator.The goal is to make better quality stylized pictures by using the idea of generative adversarial network.The discriminant network is used to distinguish the "true and false" images.The experimental results prove the effectiveness of the method with high quality visual effects.