The Research And Implementation Of Image And Video Style Transfer Based On Deep Learning

As the most popular trend in the field of artificial intelligence,deep learning has made a big breakthrough in image recognition,semantic segmentation,natural language processing and other fields in recent years.Because of its strong learning and processing ability,deep learning defeats human in some areas.As a result,some researchers have applied deep learning to image generation tasks and have achieved good results in style transfer.Style transfer is a kind of task which keeps the content of one image while using the style of another.It enables ordinary people to create art-style photos.However,there are still many problems that need to be solved in style transfer,like network training,selfies and videos styling.Aimed at these problems,this thesis puts forward several solutions as below.Firstly,this thesis designs a new style transfer network,called U-Style Net.It uses the method of transfer learning to solve the problem that improper initialization of network parameters may cause highlight or dark blocks in output images.And this training method also speeds up the training of the network.Secondly,this thesis composes four ways to improve selfie styling,which are larger style zoom parameters,global smoothing,local smoothing and fine-grained style weight.Finally,this thesis presents a fast solution of video styling,FV-StyleNet.It has a dual-channel and multi-scale input architecture.This design referred to the existing optical flow network and style transfer network,and combines them together.FV-StyleNet considers the continuity of adjacent frames in the video to achieve a fast and stable video stylization effect.It must be mentioned that the dual-channel and multi-scale input architecture has a very high parallelism,and can fully explore the GPU parallel computing power.FV-StyleNet also utilizes transfer learning to improve the training process.It is firstly trained to an optical flow network.And then with some changes to the architecture,it is fine-tuned to a video style transfer network.U-StyleNet proposed in this thesis successfully solves the problem of style transfer network parameters initialization,and improves the training speed.The four optimization schemes improve the quality of selfie styling.And the fast video styling network,FV-StyleNet,greatly accelerated the styling of videos.It makes it possible to use video styling in real world applications.