Modelling the perception of visual design principles on façades through fuzzy sets: towards building an automated architectural data generation and labelling tool

Recent studies showed that deep learning techniques and image processing can identify the distinguishing design principles in architectural façades. However, predicting the strength of a principle is still a challenging task, as it requires a huge amount of annotated design variations. The difficulties in both searching such big numbers of data – and its labelling by experts – slow down the research. This paper proposes a computation approach for obtaining this type of data faster. With the help of parametric modelling and evolutionary algorithms, we could manipulate the design elements, and thereby generate different solutions. An integrated fuzzy logic decision mechanism could enable to carry human knowledge in the judging and labelling of alternatives automatically. The final synthetic data developed from real building images could be used for machine learning applications to enhance our understanding of artistic expression.

现有研究表明，深度学习技术与图像处理手段可识别建筑外立面（architectural façades）的差异化设计原则。然而，评估某一设计原则的影响力强度仍是一项极具挑战性的工作，因为该任务需要海量带有标注的设计变体样本。既难以获取如此规模的数据集，又需依赖专家进行标注，这两大难题拖慢了相关研究的进展。本文提出了一种可快速获取此类数据的计算方法。借助参数化建模与进化算法，我们能够对设计元素进行调控，进而生成多样化的设计方案。集成化模糊逻辑决策机制可实现将人类知识自动融入对备选方案的评判与标注流程中。基于真实建筑图像生成的最终合成数据集，可应用于机器学习任务，以深化我们对建筑艺术表达的理解。