Research On Real-time Image Style Transfer Algorithm Based On Semantic Perception

Image style transfer refers to applying the style of one image to another image.In image style transfer,two images are usually input,namely content image and style image,and the output composite image has both the semantic and structural information of the content image and the color and texture features of the style image.Image style transfer has received extensive attention in the field of computer vision,and has important application value in digital entertainment,social media,and police case detection.The current mainstream style transfer algorithm is based on deep learning,which is fast and can achieve real-time effects,but the generated stylized images have problems such as texture deformation and edge blurring.This paper proposes two new style transfer models through in-depth research and improvements to existing problems.The main research contents include:1.Firstly,the existing image style transfer algorithms at home and abroad are introduced.The introduction is divided into two categories: the first category is the traditional slow style transfer algorithm optimized by image iteration,and the second category is the fast style transfer algorithm based on deep learning,which is optimized through model iteration.Secondly,the related theories of deep learning are introduced,including the composition and principle of convolutional neural networks,and the network architectures commonly used in style transfer—autoencoder networks and generative adversarial networks,which lays the foundation for subsequent research.2.A semantic-aware style transfer model based on a collection of style images is proposed.The Invalid Feature Filtering Module(IFFM)is introduced into the encoder-decoder architecture to filter redundant features that are not related to the content in the original image and the generated image,and solve the problem that the stylized images generated by the existing fast style transfer algorithms are prone to artifacts and distortions.The content retention capability of the model is enhanced by the content consistency loss constraint,which is jointly optimized with the style consistency loss to maintain the distinction of different semantic content in the generated images.Experiments show that the stylized images generated by this method significantly improve the quality of the generated images,significantly improve the problems of deformation and distortion,and can achieve style transfer based on the semantic information of content images.3.Research on the arbitrary style transfer algorithm.The arbitrary style transfer method is not limited by the type of style.When transferring a new style,there is no need to retrain the model.However,the stylized images generated by the existing algorithms are prone to blurred texture and inconsistent color distribution with semantic information.Aiming at the above problems,a semantic-aware style transfer framework based on identity loss is proposed.First,an adaptive attention normalization model is used to comprehensively consider shallow features and deep features to weight style attention;Secondly,an image is input into the network as both a style image and a content image to measure the loss,while considering the semantic local mapping and global statistics between content features and style features;Finally,by learning features between the same content and style,richer and more accurate semantic and stylistic representations can be obtained.The experimental results show that the stylized image texture features obtained by this method are clearer,the color is more vivid,and the effect is better.