|
ИСТИНА |
Войти в систему Регистрация |
ИСТИНА ИНХС РАН |
||
Spatial control for real-time depth-preserving neural style transfer.курсовая работа (Магистр)
- Научный руководитель: Китов В.В.
- Автор: Козловцев Константин Анатольевич
- Тип: Магистр
- Организация, в которой проходила защита: МГУ имени М.В. Ломоносова
- Кафедра: Кафедра математических методов прогнозирования
- Год защиты: 2019
- Курс: 5
- Аннотация: Style transfer is the process of rendering a content image in the style of a another style image. Gatys et al. 2016 solve this task using optimization of stylized image with response to high-feature statistics of style image, but optimization in space of images is inefficient, and also has a problem that stylized image looses it’s perceptual depth. In this paper we introduce a novel approach of style transfer with perceptual image depth preservation. We use architecture of real-time style transfer which can represent content and stylized image in high-feature space and allows to mix non-styled image with stylized one. We conduct experimental comparisons of our method with baseline solutions via user study evaluation and show that our method can provide better quality with higher efficiency.
- Добавил в систему: Китов Виктор Владимирович