Research On Image Style Migration System Based On Deep Convolutional Neural Network

With the rise of deep learning,the cross-collision between artificial intelligence and art represented by image style migration has attracted great attention in the field of graphic image technology and art.Image style migration is an image processing method for rendering image semantic content with different styles.It ensures the original image content structure,transforms the artistic style of the image,and obtains the texture and aesthetic features of the style image,so that the final output is generated.The image presents a perfect combination of different image content and style.However,the current style migration method based on deep learning also has limitations.From the perspective of practical application,the existing network model has the problems of large number of parameters,high storage and computational cost,and huge computational resources,which can only be used under limited platforms.It is difficult to deploy on a limited hardware platform(such as a mobile device).For this part of the actual application scenario,it cannot be effectively applied because of the computational bottleneck.Therefore,based on the deep convolutional neural network,this paper has carried out in-depth research on the problem of mobile-end model compression in style migration.The work done in this article is as follows:(1)Introduce the basic knowledge of convolutional neural networks,and introduce the principle of introducing image feature extraction of convolutional neural networks into image style migration,and visualize the convolutional neural network convolution layer to show more intuitively.The extraction process of image features of convolutional neural networks.(2)Based on the real-time style conversion model of Johoson et al.,based on the network performance of the original algorithm,the unit structure of the original residual network is improved by designing a more efficient network computing method based on the original algorithm network structure.It shows that the improved method overcomes the problem that the image style migration model is expensive and expensive to calculate,consumes huge computing resources,and is difficult to transplant to the mobile terminal.(3)At present,the quality evaluation of the style migration synthetic image is still in its infancy.This paper introduces the CNN-based image evaluation method,and objectively scores the image quality of the improved image migration style model from the perspective of image semantic features and aesthetic perception.Verify the loss of model style conversion performance after compression.(4)Designed and implemented the Android-based image style migration system,and tested a number of indicators on the Tencent excellent measurement platform to further verify the feasibility of the system.