Style Transfer Method Based On Deep Learning And Significant Region Reservation

With the rise of 5G networks,creating an image with rich artistic styles and sharing it on online social platform accounts has become increasingly popular.The style migration method mainly involves adding stylized textures and colors to the entire image.The style transfer method based on deep learning can apply different artistic images to content images,ultimately generating stylized images,greatly enhancing people’s visual senses.However,there is a problem with these methods,which is that it is easy to ignore the semantic information of the original image during style migration,resulting in distortion and deformation of the content image.Therefore,this paper proposes a new improved method that can pay more attention to the prominent areas of content images,and generate stylized images that retain the original semantic structure while producing better stylized effects.The research results of this thesis are as follows:(1)Aiming at the lack of semantic features in the style transfer network,an arbitrary style transfer method based on deep learning is designed and implemented.The VGG-19 model is used to obtain image data,and the structure of the network model is optimized to make the network more suitable for extracting image feature information.The loss function of content image and style image is calculated through the established network.(2)Aiming at the problem that the visual effect of the final generated image in the style transfer network is not ideal,the attention mechanism is added to the style transfer network to effectively reduce the training time cost of the network,improve the performance of the style transfer model,and make the trained style transfer model pay more attention to the key areas of the image to improve the quality of image generation.(3)Aiming at the problem of semantic loss in style images,semantic segmentation networks are studied and used to divide different regions of content images.Firstly,by constructing an improved Deeplab V3+ network structure,we extract useful information from the image,obtain semantic mask maps of each part,and achieve image semantic segmentation tasks.Then,we use the mask maps of each region to achieve multi region style migration.(4)Design a style transfer method based on deep learning and significant region retention.Through rigorous experiments,the model is compared with other classic style transfer methods to verify the advantages of the style transfer model in this article,and the score of style transfer effect is obtained by using multiple subjective visual perception evaluation.