基于对抗生成网络的纹理合成方法—毕业论文

摘要

纹理合成的概念已经出现很久，随着近年来游戏与影视等行业的迅速发展， 如何实现更快速、效果更好的纹理合成方法也成为了现今的重要研究方向，纹理合成在计算机图形学与计算机视觉等领域，一直是重要的组成部分之一。以往的纹理合成方法合成纹理图像通常是利用特征矩阵，而特征矩阵是对图像中特征信 息的表示。每张图像的特征信息各不相同，这也使得纹理合成方法能够合成出与原图像大致相同的生成图像。不过现有的纹理合成方法在面对现实世界中的具有不规则结构的纹理或非周期性纹理时，无法生成较好的合成结果。深圳大学曾推出一篇使用生成对抗网络实现纹理合成的论文，在面对不规则纹理和非周期性纹理时也有着较好的合成结果，本文便是复现了该论文的方法，以此加深对纹理合成的认识与理解。生成对抗网络作为一种近年来刚出现并且发展迅速的深度网络形式，利用“零和博弈”理论，通过设置生成器与鉴别器，利用生成器与鉴别器之间的博弈来训练网络，使网络能够生成新的数据，且新数据与原数据具有大致相同的分布。基于此，本文复现的纹理合成算法是基于生成对抗网络，以生成器与鉴别器的对抗训练来生成纹理，训练过程中能够高效地同时把图像的内容与对称信息考虑进去，对于具有不规则结构的纹理或非周期性纹理，有着较为优秀的合成结果。

关键词：生成对抗网络；非周期性纹理；真实纹理；不规则结构

·
·

ABSTRACT

The concept of texture synthesis has been around for a long time. With the rapid development of games and film and television industries in recent years, how to achieve faster and better texture synthesis methods has become an important research direction today. Vision and other fields have always been one of the important components. Traditional texture synthesis methods usually use a feature matrix to synthesize a texture image, and the feature matrix is a representation of feature information in the image. The feature information of each image is different, which also allows the texture synthesis method to synthesize the generated image that is approximately the same from the original image. However, existing texture synthesis methods cannot generate good synthesis results when facing irregular or non-periodic textures in the real world. Shenzhen University has launched a paper that uses generative adversarial networks to achieve texture synthesis. It also has good synthesis results in the face of irregular textures and non-periodic textures. This article is to reproduce the method of this paper in order to deepen Understanding and understanding of texture synthesis. Generative adversarial networks, as a form of deep networks that have emerged and developed rapidly in recent years, use the “zero-sum game” theory to set up generators and discriminators and use the game between generators and discriminators to train the network to make New data can be generated, and the new data has approximately the same distribution as the original data. Based on this, the texture synthesis algorithm described in this paper is based on generating an adversarial network, which uses the adversarial training of the generator and the discriminator to generate the texture. During the training process, the image content and the symmetric information can be efficiently taken into account. The texture or non-periodic texture has excellent synthesis results.

Keywords: Generative Adversarial Network; Aperiodic Texture; Real Texture;
Irregular Structure.

·

完整版论文请点这