Real-time Image Style Transfer Algorithm Based On Distributed Tensorflow And Foreground Background Separation

Entering the twenty-first century,with the continuous development of the Internet,text information is gradually instead by the image,video and other new generation of visual information.And now most mobile devices have a better shooting performance,so people's needs for multimedia information processing become more and more intense.Recently,with the popularity of a software called Prisma,image style transfer began to catch public's attention.This paper mainly concern Justin et al's real-time image style transfer algorithm.The algorithm is very efficient in the testing phase and produces high quality stylized images.However,it takes a long time to model training for a particular style and usually takes several hours.In addition,in the digital printing applications,for the images with whole colored background,the algorithm's effect is not satisfactory.The work of this paper mainly has the following two aspects:(1)Configuring multiple devices as a usable TensorFlow distributed cluster and add new startup parameters to dynamically configure the cluster environment.Constructing the real-time image style transfer algorithm on distributed TensorFlow cluster,through the parallel training of model,we can greatly reduce the model training time.(2)For the image with whole colored background,using image segmentation method to separate its foreground and background,and then respectively applying the algorithm to its foreground and background,finally we can get the image which meets the production requirement.The experimental results show that this paper's method does improve the speed of the model training and can produce the stylized images which meet the production requirements for the images with whole colored background.