Image Artistic Style Transfer, Neural Doodles & Texture Synthesis
A recent advances in image style transfer allowed incredible end-user applications. At first, Gatys et al. demonstrated that deep neural networks can generate beautiful textures and stylized images from a single example. The core idea of the method was used then to create so-called neural doodles. While the visual quality of both style transfer and neural doodles was astonishing, the methods required a slow and memory-consuming optimization process, which limited their usage. We lately improved the speed of both algorithms significantly, while preserving the quality. This allowed almost real-time stylization using GPU and was used as a core technology in several successful applications. In this talk we overview and discuss the mentioned methods.
Dmitry persuaded his Masters degree at Moscow State University and currently enrolled in PhD studies in Skolkovo Institute of Science and Technologies. His research supervisors are Victor Lempitsky (Skolkovo) and Andrea Vedaldi (Oxford). He is also employed in Yandex Research (#1 search engine in Russia). His research focuses mostly on style transfer and super-resolution. His is also interested in more general research on neural network performance improvement.