A geometric harmony evaluator for facades' design using deep learning
Imene KeskasPurpose
This paper presents a design harmony evaluator tool designed to assist architects during early design stages in optimizing urban landscape aesthetics, which directly influence human psychology and behavior. By exploring the critical interplay between visual complexity and compositional rules, the research leverages deep learning (DL) to automate and classify building facade harmony based on established design laws.
Design/methodology/approach
This paper presents a three-step approach to evaluating urban façade design harmony. First, images are annotated into harmonious and non-harmonious categories based on a geometric harmony scale. Second, visual complexity is quantified using Canny edge detection alongside Python-based fractal dimension analysis to investigate the correlation between complexity metrics and geometric harmony. Finally, pre-trained DL models are implemented in Python to automate façade harmony classification and assessment.
Findings
DL architectures specifically InceptionResNetV2, InceptionV3, DenseNet121 and ResNeXt50 successfully automate the classification of façade harmony, achieving an 80% precision rate. Rather than replacing human intuition, this automated evaluator serves as a predictive design-support tool. Integrated into the early design stages, it allows architects to iteratively test facade design harmony.
Research limitations/implications
The facade harmony evaluator could be used by all architects. However, researchers are encouraged to explore the development of similar tools based on the use of artificial intelligence (AI) to evaluate specific architectural styles or elements.
Originality/value
This paper fulfills an identified need to study how AI and DL can help to create more harmonious facades and urban landscape. Unlike traditional workflows that evaluate visual complexity only after a facade layout is completed, this study inverts the design sequence by targeting core geometric composition at the earliest sketch stage. To achieve this, a dedicated geometric harmony scale was established to define the rules of compositional order. Finally, AI and DL architectures were leveraged to successfully translate this theoretical scale into a fully automated computational tool.