Classificação de Estilo Arquitetônico através de Rede Neural Convolucional Potencializada por Unidades de Processamento Neural para Dispositivos Móveis

Abstract

The identification and classification of architectural styles are significant problems in sciences, especially in archaeology, art history, and conservation. This task is essential for the preservation of cultural heritage but is often challenging due to the complexity and variation in architectural styles over time and regions. This study addresses this issue by applying the EfficientNetV2B0 model to identify the 5 most present architectural styles in Brazil. These are Baroque, Neoclassical, Neo-Gothic, Eclectic, and Modernist; for this, a dataset was created with 927 images for training, 231 for validation, and 292 for testing. The accuracy achieved was 83%; with the Modernist category reaching a precision of 95%. However, the Eclectic style showed inferior performance, with a precision of only 64%. The model’s performance was tested on different mobile devices, including the A10 Fusion, A14 Bionic, and M2 SoCs; revealing variations in execution time between devices. These results highlight the potential of machine learning in the classification of architectural styles and its feasibility on devices with different hardware capabilities, contributing significantly to the preservation and study of Brazilian architectural heritage.