Title An Analysis on the Evolution of Korean Apartment Unit Plans using Deep Learning
Authors 안의순(Ahn, Euisoon)
DOI https://doi.org/10.5659/JAIK.2021.37.10.13
Page pp.13-22
ISSN 2733-6247
Keywords Deep Learning; Machine Learning; Apartment; Floor Plan; Typology
Abstract Deep learning methods have shown outstanding performance in image recognition on a big data scale, which has been the bottleneck in research on Korean apartments. Several studies have applied deep learning to architectural images, but analyzing Korean apartments required a deep learning model trained on a Korean apartment dataset. We developed an architectural research methodology for floor plan images, which utilizes deep learning, biclustering, and activation mapping methods. The method performs an inductive classification based on the similarity between floor plan images, guided by but not limited to accompanied class labels. We constructed a 50K unit plan image dataset of Korean apartments by collecting and normalizing floor plan images and analyzed the dataset using the developed method. The biclusters of unit plan types, extracted from the learned representation of the model, also showed a closely grouped temporal arrangement. Further examination on the unit plan types using bicluster activation mapping (BAM) showed that the deep learning model could discover areas where new design trend of the era emerged, without any prior knowledge on Korean apartments or architectural design in general.