Deep learning for ceramic fabric classification: A focus on Levantine pottery

Studying ancient ceramic artifacts plays a significant role in archaeology, as they can give information about social and human activities, providing insights into production process, technological development, uses, and provenance of raw materials, which are fundamental for reconstructing material culture and understanding ancient relationship and interactions. However, studying exchanges and trade routes remains challenging. An innovative approach to ceramic fabric classification is represented by the application of Artificial Intelligence (AI). Recent studies have demonstrated how automated methodologies can significantly contribute to ceramic analysis. Machine learning algorithms, in particular, has proven to be effective in identifying specific compositional, technological, or stylistic patterns. This study focuses on the analysis of Levantine ceramic samples using Deep Learning (DL) techniques. The primary goal is to categorize these samples into their petrographic fabrics through the implementation of Convolutional Neural Networks (CNNs) based on the ResNet18 architecture and Vision Transformers (ViT). A comparison of these models is conducted to evaluate their performance and effectiveness in classification tasks. Additionally, the present research emphasizes the interpretability of DL models by applying saliency maps (heat maps). The objective is to obtain a deeper understanding of the classification process by highlighting the distinctive features in the input data that most influenced the models' predictions .