Classifying polish in use-wear analysis with convolutional neural networks

Abstract

Lithic use-wear analysis examines micro- and macroscopic traces on tool surfaces resulting from human use and post-depositional processes. Polish, formed through surface abrasion with different materials, is a key diagnostic feature that is increasingly analyzed using machine learning to enhance automation and standardization. However, further research is needed to explore whether deep learning approaches, in particular, can be effectively applied to use-wear analysis and to determine the optimal surface area size (e.g., patch size and microscope objectives) and model architecture (custom vs. pre-trained) for achieving the best results. This study employs convolutional neural networks (CNNs) to classify experimental polish based on contact material (wood, hide, bone) and use intensity, while also assessing optimal imaging and analytical parameters. The results of this exploratory study suggest that CNNs may effectively identify polish from bone and hide but perform less effectively with wood. The models also successfully distinguish between polish formed by short- and long-term use. Custom models outperformed pre-trained ones, particularly when using images that captured smaller areas of the tool’s surface, suggesting that bigger surface areas may lack the necessary information for optimal results. These findings underscore the need to expand use-wear datasets in terms of size and variability and optimize CNN architectures and workflows.