A novel deep neural model for efficient and scalable historical place image classification.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Cultural Heritage Preservation

Background:

• Current AI classification systems for historical sites lack generalizability and interpretability.
• High computational demands and specialized expertise limit AI adoption in heritage conservation.

Purpose of the Study:

• Introduce HistoNet, a hybrid deep learning framework for efficient and accessible historical place image classification.
• Improve generalizability, interpretability, and computational efficiency in AI for cultural heritage.

Main Methods:

• HistoNet combines Convolutional Neural Networks (CNNs), Transformers, and Mamba state-space models for feature extraction and global context.
• Incorporates Squeeze-and-Excitation attention and SHapley Additive exPlanations (SHAP) for enhanced accuracy and interpretability.
• Features a user-friendly web interface for broad accessibility, even in resource-constrained environments.

Main Results:

• Achieved 95.66% accuracy on the Architectural Heritage Elements dataset and 96.46% on the Historical Building Recognition dataset.
• Demonstrated superior performance compared to VGG16 and ResNet50, with significantly improved computational efficiency.
• Showcased robust generalization across diverse conditions (lighting, weather) and architectural styles in Malacca.

Conclusions:

• HistoNet offers an efficient, accessible, and robust solution for historical place image classification.
• The framework democratizes advanced AI tools for heritage professionals, supporting global cultural heritage conservation efforts.
• Addresses digital divides and promotes sustainable development goals in the heritage sector.