Convolutional Neural Network Models For Automated Art Style Identification: Design Training, And Evaluation

Abstract

Automatic identification of artistic styles using deep learning has become an essential component of modern cultural heritage digitization. This research presents an extensive comparative analysis of Convolutional Neural Network (CNN) architectures-including AlexNet, VGG16, LeNet-5, Inception-v3, EfficientNet-B0, ResNet-50, and a custom lightweight CNN-applied to a multi-class artistic style classification task. A dataset of 35,000 curated images spanning ten artistic schools was constructed from WikiArt and open-access archives. Standardized preprocessing, data augmentation, and transfer-learning protocols were applied to ensure fairness and reproducibility across all models. Evaluation metrics included accuracy, F1-score, confusion matrices, training time, computational cost, and robustness against stylistic overlap. Experimental results demonstrate that deeper models such as VGG16, Inception-v3, and ResNet-50 achieve superior classification performance, while the custom lightweight CNN offers a competitive trade-off between accuracy and efficiency for low-resource deployments. This work contributes to the intersection of digital humanities and artificial intelligence by providing a unified benchmark and design guidelines for automated art analysis systems.