Research On Image Style Transfer Algorithm Based On Deep Learning

Style transfer is the combination of the semantic content of an image and video with the color,texture,and contour information of another image through computer technology processing,so that while retaining the original content,it has another visual style effect.The development of style transfer technology will play an inestimable role in the image processing,computer graphics and vision,film and television production fields.Therefore,it is of great significance to study the theoretical methods of style transfer and further improve the diversification of style transfer.It also has great commercial value and broad application prospects in industry.Although traditional computer imaging methods can achieve the transfer of image textures,the traditional method is combined with low-level semantic information,which makes it unsatisfactory in the final visual sense.This paper uses the method of deep learning to introduce the principles and applications of style transfer with images and videos as research targets,and proposes an adaptive instance-based normalized local area image style transfer research algorithm according to guided image filtering to achieve fast and accurate local area images.Style migration,and further extended to video style migration,the introduction of optical flow information to constrain the time domain,alleviate the artifacts,halos,etc.that appear between adjacent frames of the video,so that the video style conversion has a better visual feeling.The main work content of the paper includes the following points:(1)This paper introduce the background and significance of the research content of style transfer,thoroughly understand the current research status of style transfer technology at home and abroad,and conclude and summarize the traditional computer vision and computer graphics non-photorealistic rendering The method and effect of image information combination with the principles and advantages and disadvantages of the style transfer technology widely used in deep learning.(2)Based on the Adaptive Instance Normalization model proposed by Huang et al.This method improves the flexibility and running speed of style transfer by combining the mean and variance of the characteristics of one image with the mean and variance of the style of another image.The paper adds a guided image filter based on the original network,and proposes an adaptive example based on guided filtering to plan a local style transfer algorithm.The guided filter is used to refine the original mask image edge Method to obtain a mask image with fine edges.The combination of the guided filtering algorithm and the adaptive instantiation style migration algorithm has improved the problem of rough edges and blurred lines on the local target style migration in the network,making the style migration accurate.The image style migration in local areas achieves better results visually and makes the image style migration more diversified.We prove through experiments that this method can effectively and accurately implement the image style transfer of accurate local regions.(3)Adaptive strength normalized style migration has achieved good results in image conversion,extending it to three-dimensional video style migration.In the traditional video style migration methods based on image iteration and forward neural network,respectively,it will take too long,halo artifacts,and inconsistent images.There is a large gap between two frames of information,and the speed is slow.In this paper,an adaptive strength normalized style migration network is introduced to make the converted image have better visual effects.At the same time,the Flow Net optical flow method is introduced to limit the time domain loss function during the network training process.The stylized output at the current moment is compared with the previous one.After stylized output optical flow warp at all times,the time consistency constraint is implemented by pixel-by-pixel operation,which can directly implement the style transfer of different video sequences.The network after training does not need to calculate additional optical flow when performing video style conversion.The overall operating speed,while making the transition between adjacent frames of the video smoother,reducing jumpiness,and making the video coherent,thereby achieving a smooth video style effect,alleviating the problem of halo artifacts that appear on the edges of moving objects faster.Experiments show that the method can improve the smoothness of video,reduce artifacts,and achieve better visual viewing effect while achieving rapidly video style transfer.