Design Of Image Style Tansfer System Based On Embedded Platform

Digital images have become the main mode of information exchange and emotional exchange among people.With the help of innovations in the field of deep learning,computers use convolutional neural networks to separate and combine image content and style to achieve artistic creation.Although the image style migration has already achieved very good results on the embedded side such as Prisma,there are still many bottlenecks:(1)Cannot implement style transitions offline;(2)Parts of the content image will lose content and grid problems after style migration;(3)Model training takes a long time.In order to solve these problems,this paper implements an offline image style migration system on an embedded system based on Tiny4412.And use a monochrome light background content image to verify its performance.After verifying that the system can perform quick image style migration for the images taken by the camera in the off-line state,the indicators meet the actual requirements.(1)An instance normalization residual network is proposed.The existing residual network is improved by an instance normalization algorithm.Replace the batch normalization algorithm in the original network and changing the stride of a convolution operation in a deconvolution network.Let the size of the convolution kernel be divisible by the stride.Verification through light-colored monochrome backgrounds images,instance normalization residual network is superior to batch normalization residual network.Experimental results show:The instance normalization residual network can preserve the contrast information of the content image,overcome content information loss and grid problems of existing algorithms.(2)A training acceleration algorithm based on specific data sets.Training on instance normalization residual networks using specific data sets.Experimental results show that,in the task of performing style transfer for a specific content image,the learning speed of a particular data set is doubled compared to an ordinary data set.(3)Offline image style migration embedded system.Quantify the networks trained on PC and migration to Tiny4412 embedded development board with NDK technology.After systemtesting,the embedded system can quickly complete the image style migration task in the offline state.