Artistic Style Transfer Based On Deep Learning

With the porpularization of smart mobile phone and the pursuit of artistic beauty,the related applications of stylized image are popular.The stylized image could be not only used in application,but also in game rendering,animation production,advertising design,film production and other fields.Because the stylized non-photorealistic image has a strong visual appeal and cultural connotation,it is of great significance for research and development.The artistic style transfer based on deep learning makes the non-photorealistic algorithm intelligent,so it can also quickly generate a variety of artistic stylized images.There are some problems in the existing algorithms of artistic style transfer based on deep learning,so we come up with some solutions.When we transfer the texture of style image into a content image,the color of style image will be also transmitted,which resulting in losing the original color of content image.When the style is transmitted to the portrait,the region of stylized face will be dim.It can only produce artistic image with one kind of style by using previous algorithms.In order to solve this problems,this paper combines all the solutions into a program framework to achieve a more comprehensive artistic style transfer algorithm.The solved problems and realized functions are as follows:(1)It achieves the different degree of stylized images by adjusting the weight of the content loss function and the weight of the style loss function,and it can generate corresponding stylized images according to needs.(2)It solves the problem of transmitting the style but changing the original color of content image,so that you could choose to retain the original color in the style transfer.First,we divide the content and style image into color channel and luminance channel,and we only transfer the style in the luminance channel.At the end,add the color channel into the stylized image to implement the reservation of content original color.(3)Because the effect of the style transfer on the portrait is undesirable,the method of local style transfer is put forward.First,the DeepLab2 segmentation algorithm is used to divide the portrait image into foreground image and background image.Then we could solve the blurry stylized portrait problem by transferring the style into background image.In addition,we could transfer the style into foreground image,or we could transmit different style into background image and foreground image.(4)It realizes a multiple style transfer algorithm,which could transfer various styles of different painters to only one content image.This method is realized by changing style loss function to a weighted sum of multiple style loss functions.(5)Dynamic image stylization is realized by combining the DeepFlow algorithm and convolutional neural network.The key of this algorithm is to build the dynamic stylized loss function.And ultimately to achieve the stylization of the extracted frames which could be serialized to get the final stylized results.(6)By using the TKinter GUI library of Python,the final program interface is built.Command line is replaced by the buttons in the interface,which could invoke neural networks quickly and conveniently.