Title: Decoding Architectural Heritage: Deep Transfer Learning for Style Classification with ResNet50 and MobileNetV3
Authors: Nguyen Huynh Ng?c Tam, Nguyen Thu Ky, Nguyen Minh Duc
Volume: 10
Issue: 7
Pages: 300-307
Publication Date: 2026/07/28
Abstract:
Architecture records the technology and social arrangements of the era that produced it, yet classifying and documenting it still depends largely on manual work: field surveys, archival research, and cross-referencing of morphological features. That work is slow, and two experts may disagree - particularly when a building is hybrid or transitional. ArchiStyles classifies heritage architectural styles automatically from real-world photographs, using computer vision and transfer learning. The pipeline was built around two CNN backbones chosen for opposing priorities: ResNet50 for accuracy, MobileNetV3 for deployment on mobile devices. Both were fine-tuned on 7,526 curated images covering 18 styles, from antiquity to the present. Outdoor photographs vary in viewpoint and lighting and often contain background clutter, so preprocessing and augmentation were applied to compensate. Classification reports and confusion matrices then served to trace the morphological relationships behind inter-class similarity. On the validation set ResNet50 reaches approximately 85% accuracy and MobileNetV3 approximately 81%, at lower computational cost. Both are deployed in a visual application for heritage documentation and teaching.