Research And Application Of Image Style Transfer Based On Deep Learning

Painting is one of the important visual expressions of artistic creation.Painting style can be used to express the cultural creation background and artistic characteristics of painting.It is the most direct characteristic expression of various art painting factions.In the past,the creation of art painting required trained professionals and a lot of time.With the development of computer technology,the image processing capability of the computer can be used to accomplish the task of image art creation.But the traditional image style migration method is difficult to meet the requirements of practical applications in visual effects,and the emergence of deep learning has largely changed this situation.Among the deep learning models,convolutional neural network is the most powerful and effective model in image processing,which can effectively extract high-level abstract feature representation in images,which greatly enhances the artistic visual effect of image style transfer.However,deep learning is a heavyweight machine learning algorithm with a huge hyperparameter space that is computationally demanding in practical applications.This largely limits the further promotion of image style transfer methods based on deep learning in industrial applications.In addition,the interior of deep learning is still a black box,and the mathematical meaning of its hyperparameter space is still not understood.This makes a lot of difficulty to further enhance the visual effect of image style transfer.In this paper,the main methods of image style migration based on deep learning are deeply studied and discussed,and the improvement methods of related problems are found and proposed.The main work of this paper includes the following three aspects:(1)This paper systematically describes the related technologies of image style transfer based on deep learning,explores the basic structure and core principles of feature extractors for image style transfer,summarizes the current mainstream image style transfer methods based on deep learning into image-based iteration and model-based iteration,and introduces some important representative methods and basic ideas.(2)Image style transfer based on color preservation is a typical application case.Based on the effectiveness analysis of existing methods,this paper proposes a parametric color transfer method based on Lab color space,and introduces the total variation regularization based on L2 norm to improve the spatial smoothness of synthetic images in the process of image style transfer.The experimental results show that the proposed method can effectively solve the problem of unnatural and speckle in synthetic images and obtain considerable visual effects.(3)Image smoothing is a commonly used technique in digital image processing,but in the image style transfer based on deep learning,the effect of the traditional image smoothing method is not ideal.This paper is based on the consideration of global and local spatial smoothness,using semi-supervised learning methods for stylized image smoothing.The method utilizes the similarity of the intrinsic structural features of the content image and the stylized image,and performs smooth transmission between the pixels on the stylized image,thereby realizing smooth processing of the stylized image.Experimental results show that the semi-supervised learning method can effectively improve the quality of stylized images.