STYLE TRANSFER What is style transfer Style transfer

STYLE TRANSFER 风格迁移 李克钹

What is style transfer? Style transfer is the technique of recomposing images in the style of other images. What can it do?

3

4

History of Style Transfer • There exists some non-parametric algorithms, but they are limited to inform the texture transfer with only low-level image features of the target image. • Gatys introduced a way to use Convolutional Neural Network (CNN) to separate and recombine the image content and style of natural images by extracting image representations from response layers in VGG networks. • Later Johnson proposed a way to train a feed-forward network to approximate the operation. This technique was used by Prisma, an app that let users render any image with the chosen artistic style.

How to realize Style Transfer ▪ Visualize data ▪ Basic Preprocessing/preparing our data ▪ Set up loss functions ▪ Create model ▪ Optimize for loss function

Algorithms of Style Transfer • Slow Neural Method • Parametric Slow Neural Method • Non-parametric Slow Neural Method • Fast Neural Method • • • Per-Style-Per-Model(PSPM) Multiple-Style-Per-Model(PSPM) Arbitrary-Style-Per-Model(ASPM)

Some interesting application • Text Style Transfer + • Face Aging • Video Style Transfer(《Come Swim 》) • 2 D-to-3 D Style Transfer

reference • 陈淑環, 韦玉科, 徐乐, 董晓华, 温坤哲. 基于深度学习的图像风格迁移研究综述[J/OL]. 2019, 36(8). [2018 -07 -23]. http: //www. arocmag. com/article/02 -2019 -08 -063. html. • Palsson S , Agustsson E , Timofte R , et al. Generative Adversarial Style Transfer Networks for Face Aging[C]// 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). IEEE Computer Society, 2018. • Kato H , Ushiku Y , Harada T. Neural 3 D Mesh Renderer[J]. 2017. • Gatys L A , Ecker A S , Bethge M. Image Style Transfer Using Convolutional Neural Networks[C]// 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2016. • https: //golden. com/wiki/Image_Style_Transfer • https: //www. jiqizhixin. com/articles/2018 -05 -15 -5

Thanks 李克钹