| 摘要: |
| 平面图生成是生成设计研究中的核心部分，然而数
据集的匮乏制约了生成设计研究的发展。为了低成本地解决
数据瓶颈问题，提出并验证了一个基于算法驱动的数据增强框
架。首先，结合深度学习和生成设计任务特征，提出外环境切
割、分层训练和曲线优化3个关键步骤，基于此构建了一套完
整有效的风景园林平面图数据增强框架。其次，基于测试集对
框架各个部分进行消融实验，验证该框架的有效性。最后，应
用该框架生成一套高质量的数据集，并将数据应用于风景园林
生成设计的“图像分割”“布局生成”“平面渲染”三大主流
任务，在实践中验证该框架的有效性。 |
| 关键词: 风景园林 深度学习 数据增强 平面生成设计 语义分割 |
| DOI：10.19775/j.cla.2024.09.0036 |
| 投稿时间：2023-04-27修订日期：2023-11-08 |
| 基金项目:国家自然科学基金项目(52208041) |
|
| Research on Enhancement Methods for GeneratingDesign Datasets for Landscape Architecture Plans |
| CHEN Ran,LUO Xiaomin,LING Xiao,ZHAO Jing |
| Abstract: |
| Plane graph generation is a core part of generative design
research, yet the lack of datasets constrains the development of
generative design research. In order to solve the data bottleneck
problem cost-effectively, this study proposes and validates
an algorithm-driven data enhancement framework based on
algorithms. Firstly, combining deep learning and generative design
task characteristics, three key steps of outer environment cutting,
hierarchical training, and curve optimization are proposed, based
on which a complete and effective data enhancement framework
for landscape plan drawings is constructed. Secondly, ablation
experiments are conducted on each part of the framework based on
the test set to verify the effectiveness of the framework. Finally, the
framework is applied to generate a set of high-quality datasets, and
the data are applied to the three main mainstream tasks of "image
segmentation", "layout generation" and "planar rendering" in the
design of landscape garden generation to verify the framework in
practice. |
| Key words: landscape architecture deep learning data enhancement plane generation design semantic segmentation |
