Research And Implementation Of Image Style Transfer Algorithm Based On Deep Learning

Image style transfer is one of the research hotspots in the field of computer vision.Its main purpose is to apply the style of one image to another,resulting in an image with a new style.This technology has broad application prospects in image processing,computer vision,computer art and other fields.Traditional style transfer techniques work well in artistic image style transfer,but perform poorly in realistic image style and animation image style.In terms of realistic image style,images generated by traditional algorithms often have problems such as image texture distortion,distortion,and style overflow.In terms of animation image style,traditional algorithms are often difficult to preserve the characteristics of the original image,and there are problems such as overfitting and artifacts in the generated image.Therefore,this thesis has made the following work to address the problems existing in the above two style transfers:(1)Aiming at the problems of texture distortion,distortion and style overflow in images generated after migrating based on realistic style images,this thesis proposes a style transfer algorithm based on regularization loss constraints and semantic segmentation.First,by introducing regular loss constraints to solve the problem of texture distortion and distortion in the generated image,only the color information of the original image is changed.In this way,the generated stylized image will not be affected by texture distortion and retain the original image.Content features and color features;then,by introducing a semantic segmentation algorithm,the image is divided into different regions,and each region is stylized separately,so as to control the local style of the image;finally,on the COCO and Wiki Art datasets The experiments show that the PSNR value of the effect image based on the improved model transfer is generally increased by 2.7%,and the SSIM value is generally increased by3.4%.This method can achieve fine control of the local style of the image,thereby generating a more realistic and natural stylized image.(2)Aiming at the problems of semantic information loss,overfitting and artifacts in images generated after migration based on animation style images,this thesis proposes a style transfer algorithm based on residual block and adaptive point layer instance normalization.First of all,this thesis introduces a residual module on the basis of the Cartoon GAN model to solve the degradation phenomenon that is prone to occur in training,so as to extract more detailed features;secondly,introduce Ada Po LIN adaptive point layer instance normalization to replace the original BN normalization One,to solve the problem that BN normalization cannot transfer the color,texture information and shape information of the local area from the real image to the generated image at the same time,so that it is more conducive to the transformation of color,texture style and local shape;then,by introducing Perceptual loss improves the realism of generated images and avoids the problems of traditional pixel-level loss functions;finally,experiments on the train2014 data set show that the PSNR value of the image after migration based on the improved model is generally increased by 2.5%.The SSIM value has been increased by 3.3% overall,so the data shows that the above improved method can well improve the migration effect of the image.(3)Finally,combined with the method proposed in this thesis,an image style transfer system is designed and implemented.Users can choose different types of image style transfer according to their own needs,which greatly facilitates the user’s operation.