基于pix2pix的类平面图生成及评价方法研究  

Research on Pix2pix Based Architectural Plan-like Generation and Evaluation Methods

作　　者：崔哲[1] 郭昱 李华[1] CUI Zhe;GUO Yu;LI Hua

机构地区：[1]同济大学建筑与城市规划学院

出　　处：《住宅科技》2023年第9期27-35,共9页Housing Science

摘　　要：文章定义“类平面图”的概念,通过统一规则生成4556套不重复的类平面图样本,以此解决样本收集和量化评价难的问题;并基于pix2pix算法,设置单一GAN、堆叠GAN两种构架的生成模型,从生成图与真实图的像素相似度和主观评分两方面评价模型的学习效果。结果显示:①单一GAN模型的学习效果随样本量和学习回合的增加而稳步提升,而堆叠GAN模型的学习效果与训练参数的相关性较弱;②从像素相似度的平均值来看,单一GAN模型优于堆叠GAN模型;③从像素相似度的最高值和主观评分来看,堆叠GAN模型优于单一GAN模型;④整体而言,单一GAN模型生成的平面图空间分割更清晰,对规则的学习效果也优于堆叠GAN模型。研究结果为后续复杂建筑的训练、生成图中的断线优化、损失函数梯度的降低提供研究思路,也为大规模生成数字建成环境提供技术和理论基础。The paper defined the concept of“architectural plan-like”and generated 4,556 sets of non-repeated architectural plan-like samples through unified rules in order to solve the problem of difficult to collect sample and quantitative evaluation;and set two types of structure generating models,i.e.single GAN and stacked GAN,on the basis of pix2pix algorithm for evaluating model learning results from two aspects of pixel similarity between generated plan and actual plan and subjective scoring.The results show:(1)The learning effect of the single GAN model increases steadily with the increase of sample size and learning rounds,while the correlation between the learning effect of the stacked GAN model and the training parameters is weak;(2)From the average value of pixel similarity,the single GAN model is better than the stacked GAN model;(3)From the highest value of pixel similarity and subjective scoring,the stacked GAN model is better than the single GAN model;(4)Generally,the spatial segmentation of the plan generated by the single GAN model is clearer,and the effect of regular learning is better than the stacked GAN model.The research results provide train of thought for subsequent training of complex buildings,optimization of broken lines in generated plans as well as reduction of loss function gradients,and it also provides technical and theoretical basis for large-scale generation of digital built environments.

关 键 词：pix2pix 平面图生成 类平面图 训练方法 评价方法 

分 类 号：TU204.2[建筑科学—建筑设计及理论]