Research On Real Time Fast Style Transfer Algorithm Based On Convolutional Neural Networks

With the development and improvement of deep learning theory,image style transfer algorithms have attracted widespread attention.Today,due to the rapid rise of the Internet industry,more and more visual materials such as photos,art drawings,and comics are stored in digital form on the cloud.These diverse and diverse content image materials have important values for conveying ideas,entertaining the public,and carrying culture.However,manually beautifying images is time-consuming and labor-intensive,and it is not possible to complete the processing of large amounts of data in a short time.Therefore,in the context of increasingly rich image style migration algorithms,visual expression for image stylization operations requires more innovative research.This article takes image style transfer as the research object,focusing on the issues of single stroke stylization effect and insufficient flexibility of stylized models.The research content of this article is mainly divided into the following aspects.First of all,we proposes an improved defocus adaptive style transfer method based on the stroke pyramid to solve the problems of the existing methods,such as the single stroke of stylized effect,insufficient ability to learn large strokes,and weak sense of spatial structure.This method introduces the concept of stroke feature fusion of defocus estimation to realize the distribution of multiple strokes in the spatial dimension of the generated image,and uses hole convolution to expand the branch Receptive field to improve the network’s ability to learn large strokes,And introducing the mean standard deviation loss term to increase the spatial structure of the style image and its impact on the generated image.Then,we proposes a framework for arbitrary style multi stroke style transfer based on meta networks,which addresses the issues of insufficient flexibility and stylization effect of multi stroke style transfer methods.This method introduces the concept of meta learning to achieve the function of receiving any style,and uses Gram matrix to replace the original mean standard deviation as a feature vector and pass it into the meta network to enhance the style consistency of the generated image in terms of color,shape,and other features,Use multi-scale feature fusion to enhance the stroke differences of each branch synthesis.Finally,we designs a style transfer rendering system that controls stroke size in real-time.The system mainly has functions such as real-time image rendering style switching,real-time control of stroke size for image rendering style,and smooth transition between multiple styles,providing a new solution for controllable style transfer of strokes.This article has successively completed the research of an improved defocus adaptive style transfer algorithm based on the stroke pyramid,the research of an arbitrary style transfer algorithm based on multi stroke meta learning,and the design and implementation of a real-time style transfer rendering system with controllable strokes.This paper not only extends the learning ability of multi stroke networks,but also uses convolutional neural network technology to achieve feedforward defocus estimation.In terms of running code,a general style transfer model training framework has been established,laying a foundation for the future application of in-depth learning in the field of style transfer.