Abstract
Abstract
Automated architectural-style classification from historical building images can support the documentation, interpretation and preservation of architectural heritage. However, the visual diversity and complexity of historical architectural styles make manual analysis time consuming and difficult to scale across large image collections. This study proposes a Convolutional Neural Network (CNN)- based framework for automated architectural style classification using a Kaggle historical architecture dataset containing images representing 25 architectural styles. For the initial experimental setup, four selected architectural-style classes were used for model training and evaluation. The proposed framework applies image resizing to 256 × 256 pixels, grayscale conversion and pixel normalization to standardize the input data. The CNN architecture incorporates convolutional layers with 32, 64, and 128 filters, ReLU activation, He Normal initialization, batch normalization, max pooling, dense layers, and dropout regularization. Keras Tuner with the Hyperband algorithm was employed to optimize network configurations and hyperparameter values, while the Adam optimizer, sparse categorical cross entropy loss, and early stopping were used during training. The model achieved 99.3% training accuracy and 85.26% validation accuracy, with a macro precision of 0.83, macro recall of 0.81 and macro F1-score of 0.82 on the reported validation evaluation. The evaluation also included a confusion matrix and visual analysis of preprocessed grayscale images and model workflow representations; no Grad-CAM or saliency-map method is reported in the study. The main contribution of the study is an integrated CNN framework combining standardized image preprocessing, architectural feature extraction, regularization and automated hyperparameter optimization for architectural style classification. By enabling scalable categorization of historical building images, the proposed approach can support the organization and interpretation of architectural heritage records and provide a computational aid for heritage documentation and preservation activities. Future work will extend the experiments to a broader range of the 25 architectural styles and investigate more advanced CNN architectures.