Deep Learning for Fine-Grained Classification of Montelupo Majolica: Benchmarking and Explainability

We investigate the use of convolutional neural networks for classifying the majolica from Montelupo Fiorentino, a historically significant ceramic production. A dataset of 67 decorated classes is used to evaluate ResNet-101, DenseNet-121, and EfficientNetV2 under challenging conditions of class imbalance and fragmentary input. ResNet-101 achieves the best overall performance, though each model exhibits unique strengths. We further apply explainability methods - Grad-CAM, LIME, and Integrated Gradients - to interpret model decisions. Our results highlight the potential of deep learning in cultural heritage, and emphasize the value of comparing different architectures and incorporating explainability tools to improve model transparency and support human-centered archaeological interpretation.