Analyzing Architectural Styles: The Impact of Visual Variability on Recognition Performance

Abstract

This study investigates how variations in visual conditions and structural characteristics shape the perception of architectural form, using deep learning not as an end in itself but as an analytical lens. Rather than treating recognition as a purely computational task, the study interprets it as a process of reading architectural composition through patterns of form, detail, and visual context. Experiments are conducted across both homogeneous and heterogeneous image environments, including a merged configuration that reflects the diversity of real-world architectural expressions. By separating feature extraction from classification, the analysis examines how different stages of visual processing contribute to the interpretation of architectural characteristics. The results reveal that recognition breaks down when visual conditions shift, suggesting that architectural style cannot be understood as a fixed visual identity tied to a single data distribution. Instead, performance improves significantly when models are exposed to greater visual diversity, indicating that style emerges through a range of variations rather than stable appearances. Furthermore, the findings show that the ability to capture structural and compositional features plays a more critical role than the final decision mechanism, emphasizing the importance of how architectural form is visually structured. Overall, the study proposes that generalization should be understood as an emergent property of architectural composition, where visual variability and structural diversity jointly define how form can be recognized across contexts.