Estimation of Human Preference for Architectural Shape using CNN

This study aims to explore the possibility that artificial intelligence can identify human preferences through images using the convolutional neural network (CNN). To determine if people had a consistent preference for form, experiment participants were asked to select the preferred images among 200 images twice, which were automatically generated in dynamo. In the two consecutive image selection processes, ten participants repeatedly selected the same images at a rate of 79 percent. These results confirmed that there is a consistent preference for form. Next, the possibility of identifying the preference for form using CNN was investigated. Data for each experiment participant was divided into two sets. The preferred and non-preferred images were included in each set at a certain percentage. A classification model was produced by conducting supervised learning using CNN with one of the two sets. The classification accuracy was measured by applying this classification model to the other set. As a result of these tests, the classification model created by CNN could classify the preferred and non-preferred images with 82.7 percent accuracy. In random selection, the probability of correctly classifying the preferred and non-preferred images with more than 82.7 percent accuracy was 6.5 × 10-12 percent. Therefore, 82.7 percent reflects a fairly high classification accuracy. Based on this high accuracy, it was possible to identify human preferences for form using CNN

CNN을 이용한 건축적 형상에 대한 선호 추정