Research On Rendering Of Cartoon Style Based On Deep Learning

Non-photorealistic rendering of images is an important research direction in the field of computer graphics.It means to simulate various visual art styles by computer,so as to draw artistic images by computer.Cartoon style is also a kind of visual artistic style.With the rise of new media art and the popularity of social network in recent years,the public demand for personalized cartoon style image is growing,which is of great significance for the research of image cartoon style rendering algorithm.Cartoon style rendering is one of the challenging tasks of Non-realistic rendering.Its purpose is to transform real photos into cartoon style images,while maintaining the semantic content and texture details of the original photos.The existing methods are mainly divided into two categories: methods is based on image processing,and methods based on deep neural network(DNNs).The traditional image processing methods can only deal with the images with simple texture content,because these methods are essentially the combination of image filtering and edge enhancement,the output performance is greatly affected by the image content,and the generalization ability is poor.In addition,some deep neural network(DNNS)based methods are usually difficult to achieve a good balance between the transformation of image global style and the maintenance of image local detail semantic content,which often leads to the lack of stylization or the loss of semantic details in the image,which resulting in artifact.The cartoon style image has the following characteristics: the lines are simplified and enhanced;the color blocks in the image are smooth;the semantic content has a strong corresponding relationship with the natural picture.This is essentially different from natural pictures or other art style pictures.Based on the understanding of the characteristics of cartoon image and the analysis of the defects of the existing methods,this paper proposes a new solution for rendering input natural picture into cartoon style picture based on deep learning.This method can be divided into the following three processing stages: firstly,the natural image is preprocessed,then the processed image is transformed into cartoon style image by image style migration technology,and finally the cartoon style image is post-processed to enhance the cartoon style effects.The pre-processing of natural image mainly includes image enhancement and image filtering.The purpose of preprocessing is to improve the image quality and make it more suitable for subsequent style conversion.In the image style conversion part,this paper proposes a new image cartoon style rendering network based on deep learning,which is composed of three subnetworks: image feature modeling module,feature model alignment module and image re-rendering module.In the post-processing part of cartoon image,this paper uses the coherent line drawing to extract the edge information of the natural image,and then enhance the edge of the cartoon style rendering results according to the extracted edge information.In this paper,we build a dataset for training cartoon style rendering network,which contains 100,000 natural pictures and 145,000 high-quality cartoon pictures.In this paper,full experiments are carried out to fully compare the results of the methods proposed in this paper with those of other existing methods.At the same time,a large number of high-quality cartoon style rendering results are shown,which verifies the excellent performance of the methods proposed in this paper.