A deep learning-expert hybrid framework for interpreting interior and architectural design styles in cultural heritage contexts

This study investigates how artificial intelligence (AI) supports the interpretation of interior and architectural design styles (IADS) through a hybrid framework that combines deep learning, expert evaluation, and explainable AI. A diverse set of customized deep learning models, including EfficientNetB3, ResNet50, ConvNeXtTiny, Vision Transformer, Swin Transformer, and Latent Diffusion Model, were evaluated using a newly curated dataset of contemporary, rendered, and heritage images across six and ten design styles. Expert perception was incorporated through a bottom-up survey to compare human- and AI-based style recognition. Experimental results demonstrate that ConvNeXtTiny achieved the highest performance, attaining top-1 and top-3 accuracies of 84.3% and 97.1% for six-style classification and 74.1% and 91.4% for ten-style classification, respectively. Comparative analysis further reveals that AI models and human experts exhibit similar strengths and limitations when styles share overlapping visual characteristics, whereas recognition accuracy improves substantially when distinctive heritage-related architectural and cultural cues are present. In addition, SHAP and Grad-CAM analyses illustrate that the model primarily relies on materials, textures, and architectural elements for style recognition. These findings show the potential of AI-assisted decision-support systems to complement expert judgement in architectural design and heritage-related applications. HighlightsA hybrid deep learning-expert system is proposed for interior and architectural design style classification.An expert-validated dataset facilitates image-based analysis of design styles.Comparative results show complementary behaviours between deep learning models and human experts.The framework improves interpretability and reliability in multi-style recognition tasks.