Deep learning-based aesthetic evaluation of detached housing designs using rendered images

The aesthetic evaluation of architectural computer renderings has traditionally remained subjective and dependent on personal, situational, and cultural factors. Within this research, we investigate if deep learning (DL) can be utilized to provide a scientific data-driven solution for approximating the perceived aesthetics in architecture. Our focus is on standalone house designs and uses a dataset of 1,438 computer-rendered competition entries off the Arcbazar website, assigned a rating by professional architects for visual quality. In this research, "aesthetic evaluation" refers to the numerical scores given to the attractiveness to architectural renderings. Our dataset of renderings was standardized through image preprocessing and paired with averaged expert scores. A supervised convolutional neural network (CNN) regression model was then trained and validated using three-fold cross-validation. Model accuracy was established using standard measures of regression (MAE, MSE, RMSE, and R²). Results indicate that the model was able to predict aesthetic scores with high validity. While the findings demonstrate the validity of DL models to evaluate architectural renderings, the following limitations should be pointed out: The dependence on rendered views, assessment of just one building type, and expertise based on raters from one platform. Future research will have to expand on the aspects of differing building types, cultural contexts, and multimodal inputs. The incorporation of explainable AI methods will further assist in identifying which visual features contribute most to aesthetic prediction. This work establishes a proof-of-concept framework for integrating deep learning into architectural evaluation, supporting an extensible system that allows for design competition and decision-making. Apart from predictive scoring, such models are well-suited to be integrated with generative design frameworks that will enable the generation of novel architectural proposals optimized in aesthetic quality.

Keywords: Automated scoring, aesthetic assessment, deep learning, human scores.