Image Style Transfer Based On Image Semantic Correspondence

With the rapid development of science and technology,the processing of color images has become more and more popular.Many experts and scholars have done in-depth research in this field.The application of image style migration in daily life is more and more,The processing of people's photos were become more interesting.Traditional image style migration requires more time and effort,and users need to have higher skills in using software to process images.With the continuous development of deep learning.deep learning was also widely used in the style transfer of images.It also makes the style transfer of images easier to implement.It is no longer a skill that a few people have mastered or taken a lot of time.Color transfer technology is one of the branches.Its purpose is to migrate the color distribution information of the reference image to the source image to enhance the visual aesthetics of the image..People can directly use the photos taken by the mobile phone to transfer the color style of the photographic images.However,the color style transfer of the photographic image is particularly important for the relationship between the corresponding objects of the content of the source images and the reference images.This paper focuses on the algorithm of color migration that maintains image content consistency under the deep learning framework.First we propose a photographic image style transfer algorithm guided by semantic correspondence positional relationship to ensure that the semantic information of the source image remains in its original state when the color information of the reference image is migrated to the source image.Firstly,the two-way nearest neighbor method is used to find the corresponding positional relationship between the source image and the reference image in the feature domain of the image.At the same time,the semantic segmentation image is used to constrain the nearest neighbor method to find the range in the feature domain,getting a more accurate local corresponding positional relationship.And based on the corresponding positional relationship,the image synthesis is completed by weighted least squares method,and the image optimization algorithm is applied to the image by using the matting loss function to ensure semantic accuracy and transfer loyalty.In this paper,the VGG-19 neural network is used to extract the feature domain of the source image and the reference image,and extract the feature domain of the conv1 1,conv2 1,conv3 1,conv4 1 and conv5 1 layers,Then establish a semantic correspondence in each layer of the feature domain.After considering some experimental results because the color migration caused by the inaccuracy of the semantic segmentation image is not perfect,this paper proposes two methods for image optimization.The first method is to use the deep learning framework and the extinction loss function to optimization the result image to ensure semantic accuracy and transfer loyalty.This optimization method is also the image optimization method finally adopted in this paper.The second image optimization method is an image optimization method for an image of a result before image imaging using a weighted least squares method,an automatic encoder image optimization algorithm,This optimization algorithm mainly uses the semantic correspondence between the source image and the reference image obtained in the previous step,changes and optimizes in the feature domain of the image,and finally uses the automatic encoder to decode the feature domain into the final result image.In addition,the method proposed in this paper is further extended to study the automatic retrieval of reference images from the database to make the users more friendly when using color migration.The experimental results show that our method can successfully convert the color information of photographic images while maintaining the semantic correspondence between diversity scenes.User research has also shown that our approach is superior to the most advanced photographic style transfer methods.