Генерация текстур с использованием свёрточных нейронных сетей

Abstract

Nowadays textures are used in a wide range of tasks such as games or environment emulation software. However, creation of detailed textures is time-consuming. In such cases automatical means of texture synthesize could be used. Here we introduce an approach to synthesize textures based on feature representation of an original image in convolution neural network which takes into account the complex data dependency. An important aspect of the proposed approach is that, if we change a white noise image on the start of a texture synthesize process, the corresponding appearance of the generated textures will also be changed.