Image Style Transfer Based On Deep Learning

Image stylization is the conversion of an ordinary photograph into an image with other artistic effects.In recent years,with the wide popularity of intelligent mobile devices,people have a growing demand for image information processing,a variety of beauty cameras and filters are widely used by people,and the public has higher and higher requirements for the effect of image stylization.In addition,image stylization is also widely used in movie rendering,game rendering,animation design and other fields.Therefore,the study of image stylization has wide application value and commercial use.In the current situation of image stylization research,it is found that there are some problems.In the algorithm based on cross-domain style migration,the global migration is easy to cause the loss of background information.Most style migrations of a network model can only migrate one fixed style.If you want to migrate other styles,you have to re-train the model separately,which is inefficient to use.In view of this problem,this paper has done the related research,finally in the appeal of the two kinds of problems in the more successful image stylization algorithm.The main work of this paper is as follows:1.Study the unsupervised learning cross-domain style transformation model CycleGAN,and propose a attention-based CycleGAN network model to solve the problem that the background information of this model is easy to be lost in the style migration,and the main object contour of the image is indistinct from the background boundary.Experiments were carried out on horse2 zebra,selfie2anime and vangogh2 photo data sets to enhance the detail effect of the stylized image,which was better than the original network.2,research based on any single model style fast stylized migration algorithm,this article USES a set of symmetric VGG network as the encoder and decoder,extract the contents of the encoder features and style respectively through a set of convolution layer and a fully connected to fitting the fusion between the two characteristics of the information,make the network not only retain the original content of the image highlevel semantic information,and can quickly rebuild a stylized the good effect.Through the experiment,it is concluded that the improvement of this paper has some improvement in the effect of stylization.In terms of efficiency,it greatly reduces the time and training cost of style migration.3.In this paper,under the Windows operating system,an application software based on Pyqt interface development tool is built.The system has the input of images of any format.