摘要
采用风景园林学与人工智能的跨学科研究方式,开发了一种将深度学习模型——生成对抗网络(Generative Adversarial Network,GAN)用于风景园林平面图用地识别与图像渲染的新应用场景。以325张细致标注的平面方案图建立用于深度学习的数据集,训练循环生成对抗网络(CycleGAN)实现平面图不同用地类型地块的提取任务,以及平面色块图到色彩肌理图的渲染生成。进一步从图片质量、正确规范性和色彩表达等方面评价模型的识别与渲染结果。该训练模型有潜力被应用于风景园林案例的用地类型分析及平面渲染,帮助设计师提升分析及制图效率。
Integrating interdisciplinary research methods of Landscape Architecture and ArtificialIntelligence,this paper develops a new scenario of applying a deep learning model-Generative Adversarial Networks(GAN)to recognize land uses and render landscape master plans.A landscape plan dataset with 325 images has been established and CycleGAN were trained to complete the work of land use recognition and plan rendering(from RGB color layouts to texture plan drawings).The results of identification and rendering were further evaluated from the aspects of quality,correct-expression and artistic characteristics.The model proposed is of great potential for enhancing the work efficiency of landscape architects,especially when dealing with plan mapping and case analysis.
作者
周怀宇
刘海龙
ZHOU Huaiyu;LIU Hailong
出处
《中国园林》
CSCD
北大核心
2021年第1期56-61,共6页
Chinese Landscape Architecture
关键词
风景园林
生成对抗网络
风景园林平面图
图像识别
图像生成
设计制图
landscape architecture
generative adversarial network(GAN)
landscape plan drawing
image recognition
image synthesis
design mapping
