Research Of Photorealistic Style Transfer Algorithm Based On Deep Learning

Photorealistic style transfer is a new research hotspot in the field of computer vision.Its main task is to transfer the style of one image to another.The stylized image retains the content structure of the original image and has the style of the new image.Photorealistic style transfer is widely used in image processing,video processing,data annotation and other fields.It is an important research.While the existing methods have achieved promising results,they are prone to generating either structural distortions or inconsistent style due to the lack of effective style representation.Based on deep learning,this paper studies the photorealistic style transfer to solve these problems.The main contents and innovations of this paper are as follows:(1)Photorealistic style transfer algorithm based on Multi-order Image Statistics Network(MISNet).In order to effectively model inherent style information and improve computational efficiency,this paper designs a novel dual-branch structure that subtly combines first-order and second-order image statistics to decorate content features with style information while maintaining the authenticity of images.The first branch captures visual style using a covariance matrix to ensure consistent style colors for stylized images,while introducing a learnable linear transformation matrix to learn the covariance of feature maps to improve computational efficiency.Another branch uses first-order statistics to align features of content and style for accurate image reconstruction.In addition,a lightweight and efficient adaptive aggregation mechanism,namely Triplet Feature Fusion(Tri FF),is proposed.Tri FF can dynamically select effective information from each branch while retaining their unique complementarity feature.Qualitative and quantitative experiments show that the photos transferred by this method have better realism and computational efficiency.(2)Photorealistic style transfer algorithm based on Multi-scale Feature Fusion Network(MFFNet).In order to retain more structural details,a multi-scale photorealistic style transfer network is designed on the basis of MISNet,which improves the problem of incomplete content structure retention in existing algorithms.In the proposed MFFNet network,the information of different levels of features in VGG19 is fully considered.The accurate image reconstruction is achieved through skip connections and adaptive instance normalization between features at different scales.In addition,the network introduces a parallel attention mechanism and Laplacian loss function to reduce artifacts.Experimental results show that the proposed algorithm can effectively eliminate accidental artifacts,preserve more content structure details,and correctly transfer important style information.This thesis contains 31 figures,9 tables and 86 references.