Portrait Image Cartoon And Caricature Style Transfer Algorithm Study

The portrait of cartoon-style transfer and caricature-style transfer aims to render the content of an image into a cartoon-style image or caricature-style image,which gains much attention from researchers in the fields of computer vision.However,because of the complex semantic information in the face,existing cartoon portrait generating algorithms often lose semantic clues,which will ruin the image content and structure,leading to a poor cartoon portrait image.Besides,for caricature portrait images,some existing caricature portrait generating algorithms based on unsupervised learning also have issues of image content and structure being ruined.Meanwhile,it does not get excellent face deformation effects,lead to a poor caricature portrait image.We have researched in the field of cartoon-style transfer and caricature style transfer for the above problems.We propose a cartoon style transfer method for portrait images based on semantic information to solve image content and structure loss.We re-define the semantic as the relative depth of field in regional pixel,which guides the style element transfer(color and texture)in the target image while training,thus reducing the color artifact and preserving semantic information completeness.The algorithm also introduces a bilateral filtering operation to denoise images,which avoids noise features while color transfer,reducing color artifacts caused by noise and preserving the contour information of the image.In portrait caricature style transfer,we propose a distributed caricature style transfer method to solve the problem of the content and structure loss and face deformation,which adds a step of exaggeration after the proposed cartoon style transfer approach.The method preserves the semantic information completeness in the image while exaggerating the face-by-face deformation model,thus,improving the quality of generated caricature image.We also reduce the complexity of the model and training process by separating stylization training and face deformation training.Compared with the classic style transfer algorithm and the latest unsupervised image-to-image algorithm,the proposed algorithms' effectiveness is verified.The key contributions are listed as follows:(1)We re-define the semantic as the relative depth of field in regional pixel and reduce disturbing noise while color transfer by bilateral filtering operation,which effectively solves the issue of color artifacts while style transfer.The semantic is used to distinguish the regions of the portrait image and then guide the color and texture synthesis;The bilateral filtering operation is used to denoise the image,preventing noise expansion in our network while preserving the contour information of the image.(2)We divided the portrait caricature style transfer into two parts: cartoon style transfer and face deformation.The proposed method can effectively preserve the image structure information and reduce texture and contour information loss while face deformation;accordingly,the effect and quality of generated caricature images are improved.The main work of this paper includes:(1)A generative adversarial network framework for cartoon-style transfer method is proposed based on semantic information.The algorithm re-defines the semantic as the relative depth of field in regional pixel,which guides the style element transfer(color and texture)in the target image while training,thus reducing the color artifact and preserving the semantic information completeness in the image.(2)On the method of generating cartoon portrait images based on semantic information constraint,we introduce a bilateral filtering operation.The bilateral filtering is used to denoise images,which avoid the influence of noise features while color transfer,reducing color artifacts caused by noise;meanwhile,edge protection characteristics of bilateral filtering ensure the completeness of contour information in the image.(3)we propose a distributed caricature style transfer method to solve the problem of distortion and style loss,which adds a step of face exaggerated deformation by face deformation model after our method of cartoon portrait style transfer based on semantic information constraint to obtain a better effect caricature portrait image.