Computational aesthetics in architecture: a framework for quantifying preferences using computer vision and artificial neural networks

• Ano de defesa: 2024
• Autor(a) principal: Sardenberg, Victor Carrilho
• Orientador(a): Becker, Mirco
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Universidade Presbiteriana Mackenzie
• Palavras-chave em Português: computational aesthetics; artificial neural networks; aesthetic measure
• Link de acesso: https://dspace.mackenzie.br/handle/10899/39572

This research develops a computational aesthetics framework to predict the hedonic response of groups of people to architectural images. The theoretical basis relies on classical aesthetic theories of parts to whole, such as Alberti´s, combined with early 20th-century quantitative aesthetic metrics by G. D. Birkhoff and digitally-enabled contemporary technologies, such as Computer Vision (CV) and Artificial Neural Networks (ANN). This work focuses on the visual perception of architecture through perspectival images. CV is applied to identify parts in images, such as walls, doors, and windows. These parts are reorganized in diagrams to analyze the number of parts (DSP) and quantify their relations to the whole (DCG). The quantities derived from the diagrams inform two methods for quantifying and predicting the hedonic response of other images: 1. Birkhoff's Aesthetic Measure (AM) formula is adopted to reduce the complicated aesthetic experience into numbers. CV is applied to automate it, speeding up its application and making it unambiguous. The formula is calibrated to fit the audience's preferences better, producing a Calibrated Aesthetic Measure (cAM). 2. ANNs are trained because of their ability to find patterns in data. The numerical output from the DCG and DSP, the AM, and the cAM are used as inputs to train the model. This model is named the Predicted Hedonic Response (PHR) model. The described framework requires surveying specific audiences to incorporate their bias. Therefore, this research does not aim to develop a universal model of aesthetic evaluation but to embrace the specificity of each group of individuals. The framework is applied to navigate design spaces for parametric models and generative adversarial networks. The thesis discusses the implications of quantification in architectural evaluation, parts to whole relationship paradigms, and the role of images and playing in architecture. Finally, it concludes that the computational aesthetics framework build is a heuristic for predicting the aesthetic preferences of groups.