| Texture synthesis is one of the fundamental problems in computer graphics,virtual reality and image processing.Its goal is to generate arbitrarily sized textures which maintains global structure and local details of a given texture sample.It has broad application prospects in scene rendering,image restoration,image artistic style conversion,and computer animation.Although the researchers on texture synthesis have made significant progress in the past 20 years,the synthesis of non-stationary textures is still an unsolved question.Many state-of-the-art methods can only handle repetitive and periodic texture elements of stationary textures,but cannot analyze largescale structures and spatial variations for non-stationary textures,which makes it fail in generating satisfactory non-stationary textures.With the development of deep learning techniques in recent years,the researchers have made great achievements in neural-based image synthesis.Among the new progress,generative adversarial network(GAN)is one of the most popular and successful generative model in deep learning research,which greatly promotes the research on image-to-image translation,image synthesis etc.This paper focuses on non-stationary textures synthesis.In particular,the main contents of this paper include:1.Single Texture Expansion.According to the goal of texture synthesis,we propose a new training method based on self-supervised learning,so that the generative adversarial network can learn the mapping from small texture patch to large texture patch containing it.This will enforce the network to have the ability to extend the input texture to an output texture that is twice as large as the input.The experiment proves that the result can maintain the texture pattern and large-scale structure of the input nonstationary texture.Stress testing proves that the network is robust.2.Controllable Multi-Textures Expansion.In order to solve mode collapse and the problem of unable to control the output of the network of single texture expansion,we modify the strategy of extracting training data to mix the training of texture expansion and transfer,add a classifier network to the original network,and add a classification error term.We also modify the generator by adding an inspiration layer and modifying its input.The new generator must have two input images.One provides structure and the other provides texture pattern.By doing so,the network is improved to be capable of expanding multiple textures,and meanwhile the generator network will not suffer from mode collapse.Most importantly,the users can control the output of our network by inputting different style images. |
