This study presents a versatile approach integrating various neural network architectures with a focus on classifying architectural works. To address the lack of suitable datasets in the literature, a custom dataset has been created and made publicly available. The study aims to determine the most effective model, taking into account fine details in architectural styles. In this context, a comparative analysis has been conducted on four different convolutional neural network (CNN) architectures, including a baseline model trained from scratch and models using transfer learning methods with VGG, ResNet, and EfficientNet architectures. Through experiments, the EfficientNet architecture was fine-tuned, achieving an accuracy of %84.65 for 3 architects and %74.08 for 16 architects. Additionally, the two obtained models were used as feature extractors to visualize relationships among architects in a 2D space using the t-SNE dimension reduction technique. These promising results indicate that these techniques can significantly contribute to architectural style analysis and serve as valuable tools for creating innovative designs through the use of generative artificial intelligence.
The primary goal is to understand the evolution of architectural styles during the specified historical periods. The study delves into the distinctive styles of architects who played key roles in shaping the First and Second National Architectural Movements.
For more in-depth details, you can read our published paper: Architectural Works of the Early Republic Period from an Artificial Intelligence Perspective, presented at the 2024 32nd Signal Processing and Communications Applications Conference (SIU), Mersin, Turkiye.
- Architects-All Dataset: Hugging Face - Architects-All Dataset
- Architects-Top Dataset: Hugging Face - Architects-Top Dataset
- Trained Models: Hugging Face - Trained Models
- Architects-Top Dataset Accuracy: 84.65%
- Architects-All Dataset Accuracy: 74.08%
- 0.43 silhouette score for the clustering
Watch our video for a concise overview of our study:
All code is available in the Jupyter Notebook classify_arch.ipynb.
- Enhance the Dataset: Integrate more architectural datasets to improve generalization.
- Explore Advanced Neural Network Architectures: Investigate models with Transformer and attention mechanisms for further innovation.
- Generative AI and Practical Application: Apply the trained model to support architects in design processes while exploring connections with generative AI.