The Research And Application Of Real-time Style Transfer Web Technology Based On Convolutional Neural Network

With the further maturity of machine learning and deep learning technology,research and practical application of related technologies are becoming more and more important.Whether it is PC,mobile or embedded,it is the ultimate goal of technology development to practice and apply mature technologies.Web applications are based on the B/S model,which has a global market share of more than 50% due to its convenience and cross-platform nature.Due to the application of image style transfer technology in the practical application,the establishment of deep convolutional neural networks requires a large number of data sets for learning and training.The computational complexity is huge and it is difficult to transfer real-time image styles.Therefore,research on real-time image style transfer and its application technology is of great significance.Research significance and market application valueIn this paper,the insufficiency of image transfer in traditional non-convolution neural networks is studied in depth.The improvement of image style transfer using traditional convolutional neural networks is analyzed.It is found that the convolutional network structure in deep learning can input images into users.The content and style information features are effectively segmented.Through the idea of migrating learning,the information feature extraction of the middle content and the style image of the image is completed.Based on the combination of Django framework and TensorFlow platform,a Web prototype system was established to realize and verify the real-time image style transfer technology based on convolutional neural network.The main work of this paper is as follows:(1)In this paper,for the traditional image style transfer method,there are problems such as slow image generation time,style transfer effect to be improved,and style of image style transfer.A method based on image information feature histogram adaptive matching(HdaIN)is proposed.The method improves the different statistical methods in the image information feature matching process and improves the original histogram based on the mean and variance matching method.The histogram of the image information features is adaptively matched to obtain the target image information features.The experimental results show that the method can realize image style transfer in real time and improve the image quality of target stylization.(2)Based on the proposed HdaIN method,this paper designs and implements a real-time image style transfer network structure based on convolutional neural network.Aiming at the problems existing in real-time image style transfer,the real-time style transfer Web prototype system is realized and verified and performance analysis.The system has the ability to realize style transfer in real time,span multiple servers,and guarantee any style.