Facial Cartoon Style Transfer Based On Deep Learning

Face cartoon style migration refers to the process of re rendering a real face image and converting it into a cartoon image without changing its content characteristics.It plays an important role in enriching people’s social lives and promoting the development of the animation industry.In recent years,many methods for transferring facial cartoon styles based on deep learning have been proposed.However,these methods have some shortcomings and there is room for improvement in the output images.For example,there are artifacts,facial features that are not coordinated,styles that are single,and color losses that are significant.The resolution of these issues contributes to the generation of high-quality cartoon images,providing animators with rich materials and producing realistic cartoon portraits.Therefore,this thesis studies two important issues,namely,the insufficient diversity of output images and the high color loss of output images.The main results obtained are as follows:1.Aiming at the problem of insufficient diversity of output images,this thesis proposes a deep neural network based on the separation of image content and style.In the proposed method based on deep learning,a deep neural network is generally composed of a generation network and a discrimination network,which are used to generate cartoon images.Hidden vectors and real images are simultaneously inputting into the generation network,and the variability of hidden vectors is utilized to generate different stylized results.However,the generation network tends to focus on high-dimensional and structured content images,while ignoring low-latitude hidden vectors,this leads to insufficient diversity.This article encourages the generation network to explore more cartoon styles during training by adding a display diversity constraint to maximize the ratio of the generated image distance to the corresponding hidden vector distance.At the same time,the neural network embeds the variability of hidden vectors into style coding through reconstruction losses,add a circular consistency loss to reduce the content loss of the output cartoon image,add facial recognition losses to enhance facial consistency between cartoon images and real images and use anti loss to learn the style features of cartoon image domains.2.To address the issue of color inconsistency between the output cartoon image and the real image,this paper proposes a deep neural network based on Anime GANv2 that combines SENet in the YUV space,and introduces cosine distance loss in the U and V channels.The neural network consists of a generating network and a discriminating network.Generation networks are used to learn the data distribution of cartoon image domains and map real images to cartoon images.While real images possess cartoon style,their colors are also very prone to change.This article reduces the color changes of output cartoon images by adding cosine distance losses on U channels and V channels.At the same time,by adding channel attention modules to the generation network,important channel feature information is enhanced and secondary channel feature information is weakened,this distinguishes the importance of channel feature information.3.A large number of experiments and analyses have been conducted on the dataset selfe2 anime,which is commonly used for facial cartoon style migration.The results show that the multi style migration network proposed in this thesis can generate more cartoon styles in cartoon images,and the most significant experimental result is that it can learn the data distribution of cartoon styles with glasses in the target domain.It can quickly achieve the diverse output of cartoon images and has a positive role in promoting cross domain transformation and multi domain output issues.On the other hand,the color preserving style migration neural network proposed in this thesis can make the output cartoon image and the input real image basically consistent in color,so that the cartoon image retains more content information of the real image,which is conducive to making a more realistic cartoon portrait.