Research On Image Stylization For Semantic Preservation

Stylization refers to the abstract expression of specific things with artistic methods,so as to explain things while emphasizing a specific emotional color or ideological tendency.The carrier of specific things can be images,words or sculptures,and the stylized technical methods also vary.With the continuous improvement of people's living standards,image stylization has become an indispensable technical method in family life or advertising design.Compared with realistic photos,stylized images can more strongly convey specific emotions and intentions,making people more Even more impressive.In the past,stylized images were usually hand-painted by artists,such as Picasso,Van Gogh,and Wu Guanzhong of our country.They used rich experience to fuse the "scape" and "meaning" of painting,but this work was also very time-consuming.In recent years,an image stylization method based on a convolutional neural network has emerged,which can quickly and conveniently transfer the specified style to the target content map.Although the stylized image generated by this method is exquisite,it often erases or changes the semantic information to be expressed in the content map,making the generated stylized image lose practical significance.The semantics of an image refers to the main information that the image conveys to the audience.It is generally affected by a variety of factors,such as the salient area of the image: the image content in this area contains the main information that the image is intended to express,distorted or covered such area may cause the image to be incomprehensible or misunderstood.In order to give full play to the advantages of image stylization methods based on neural networks,and to make the stylized results express semantic information consistent with the target content map as much as possible,this paper mainly carried out the following work.This paper proposes a stylization method that can maintain the saliency area of the image.By designing a new loss function to monitor the change of the saliency area in the generated stylized image,it is stylized with the image based on the convolutional neural network.The methods are combined to keep the saliency area of the generated stylized image the same as the target content map.The experimental results show that the stylized image generated by this method not only integrates the specified artistic style well,but also accurately expresses the main semantic information of the ta rget content map,which has higher practical value.In addition,by analyzing and researching a variety of typical image stylization methods that maintain semantic information,and exploring their advantages,disadvantages and commonalities,we find that although these methods perform well on their respective semantic factors,they cannot be integrated.The heads of various households lead to unsatisfactory stylized results in some cases.So we propose an image stylization method that can maintain multiple semantic factors at the same time.This method can flexibly add a new semantic factor estimation method to the system framework,so as to estimate the degree of deviation of the stylized results on this kind of semantic factors,and then through training and optimization,we can get a method that can simultaneously maintain multiple Image stylized model of semantic factors.