Research On Adaptive Image Style Transfer Algorithm Based On Deep Neural Network And Content Retention

The development of artificial intelligence(or AI,for short)has not only changed people's daily life,but also transformed the models of the production and management.It has penetrated into all aspects of modern society.From a scientific perspective,artificial intelligence is a science that studies and expands human intelligence.The application of artificial intelligence in image processing-image style transfer is also getting deeper and deeper.Image content is presented in another image style,called image style transfer,which is one of the most interesting applications in deep learning.This involves two pictures,a style picture(usually a famous painting)and a content picture(the content we want to draw).We want to transfer the style of the famous paintings of the maestro to ordinary pictures,and machines using artificial intelligence can also paint famous paintings.There are currently several methods to achieve the effect of image style transfer.The first is an iterative optimization method.The speed of the iterative optimization method is very slow,which limits its practical application.The second is to use a feedforward convolution method.The feedforward convolution method can only implement one or a limited variety of styles,which limits the flexibility of the style transfer method.The third method is the transfer method of arbitrary image content and style,but this method shows a blocky effect in the styled image,which affects the quality of the generated image.In this study,an adaptive image style transfer algorithm based on deep neural network and content retention is proposed.The core of our method is a novel adaptive transformation layer by training the adaptive conversion operator A,the second-order statistical conversion of the style image to the content image is completed.The mean and covariance of the content features in this layer are aligned with the mean and covariance of the style features.In addition,we first proposed the MaskShading module,which solves the problems of image content detail loss and blurred edges that occur during the image style transfer process by generating mask patterns.It can also realize the transfer of the specified image area.A mask separates the foreground from the background and applies a style to the specified foreground.Through the experiments,we analyze and compare the image effects of the image style transfer generated by iterative optimization method,feedforward convolution method,arbitrary style transfer method and our method.Experiments show that the proposed algorithm framework achieves a good compromise in speed,flexibility and quality,and is superior to existing methods.This method can solve the problem of image content distortion and blurred contour details,so that the image with rich content can get accurate style transfer.Our method allows flexible user control using a single feedforward neural network,such as content style tradeoffs,style interpolation,color and space control.