摘要
虚拟地形作为三维场景中重要组成部分，在 CG 影视、军事模拟、开放世界游戏、虚拟现实以及三维仿真等领域中被广泛应用。地形建模任务中创作者往往难以高效的复现出真实地形的复杂性与多样性。如何高效通过用户控制创作出具有真实性的虚拟地形一直是计算机图形学领域中的一大挑战。随着深度学习的飞速发展，已经在风格迁移、人脸识别等诸多领域展现出传统算法难以企及的效果。如何基于深度学习技术快速有效地生成用户可控并且具有高度真实感的地形，成为新的研究方向。
在 GAN 模型的各种分支中，条件生成对抗网络提供了对生成结果的约束，增加了对生成图像的控制。pix2pix 是在 CGAN 基础上提出的一种通用的图像转换网络，本文在 pix2pix 网络基础了做出细微调整，设计了适合虚拟地形生成的网络， 实现了由地形特征（山脊线、山谷线等）图生成多风格具有真实感的虚拟地形。该地形生成网络只需根据用户提供的山脊线、山谷线和正负地形等地形特征的描述， 就可以自动生成符合特征约束的虚拟地形。本文的主要工作和贡献如下：
（1）构建了样本数量较大且包含四种风格（平原、丘陵、高原、山地）地形与地形特征组合数据集。
（2）基于 pix2pix 网络架构，设计了适合虚拟地形生成任务的地形生成模型。
（3）设计了具体实验，比较了不同地形特征组合对虚拟地形生成的影响，以及相同地形特征约束下不同类型地形生成效果差异，
（4）将本文中的虚拟地形生成方法进行了实际的应用分析。结果表明本文虚拟地形生成方法不仅能够帮助创作者快速搭建大型地形场景，还能够通过对地形特征图的修改实现局部地形的快速修改。
关键词：地形生成；生成对抗网络；数字高程模型
Abstract
As an important part of 3D scene, virtual terrain is widely used in CG films, military simulation, open world games, virtual reality and simulation. In the task of terrain modeling, it is often difficult for creators to efficiently reproduce the complexity and diversity of real terrain. How to efficiently create realistic virtual terrain through little user control has always been a major challenge in the field of computer graphics. With the rapid development of deep learning, it has shown the effect that traditional algorithms are difficult to achieve in many fields, such as style transfer, face recognition and so on. How to generate user controllable and highly realistic terrain quickly and effectively based on deep learning technology has become a new research direction.
In various branches of GAN model, the conditional Generative Adversarial Network provides constraints on the generation results and increases the control of the generated image. Pix2pix is a Generative Adversarial Network based on CGAN. This paper makes minor adjustments on the basis of pix2pix network, designs a network suitable for virtual terrain generation, and realizes the generation of multi style and realistic virtual terrain from terrain feature (ridge line, valley line, etc.). The terrain generation network can automatically generate virtual terrain conforming to feature constraints only according to the description of terrain features such as ridge line, valley line and positive and negative terrain provided by users. The main work and contributions of this paper are as follows:
(1) The combined data set of terrain and terrain features with a large number of samples and four styles (plain, hill, plateau and mountain) is constructed.
(2) Based on pix2pix network architecture, a terrain generation model suitable for virtual terrain generation task is designed.
(3) Specific experiments are designed to compare the effects of different terrain feature combinations on virtual terrain generation, as well as the differences of different types of terrain generation effects under the same terrain feature constraints,
(4) The virtual terrain generation method in this paper is analyzed in practical application. The results show that the virtual terrain generation method in this paper can not only help the creator to quickly build large-scale terrain scenes, but also realize the rapid modification of local terrain through the modification of terrain feature map.
Keywords: Terrain generation; GAN; DEM
参考文献 .................................................. 40
致谢 ...................................................... 44
