Discrimination of wheel-thrown pottery surface treatment by Deep Learning

The study of pottery surface treatment is essential to understand techniques used by ancient potters, in order to explore the cultural and economic organisation of past societies. Pottery is one of the most abundant materials found in archaeological excavation, yet classification of pottery surface treatments remains challenging. The goal of this study is to propose a workflow to classify pottery surface treatments automatically, based on the extraction of images depicting surface geometry, calculated from 3D models. These images are then classified by Deep Learning. Three Convolutional Neural Network algorithms (VGG16 and VGG19 transfer learning, and a custom network) are quantitatively evaluated on an experimental dataset of 48 wheel-thrown vessels, created by a professional potter specifically for this study. To demonstrate workflow feasibility, six different surface treatments were applied to each vessel. Results obtained for all three classifiers (accuracy of 93 to 95%) surpass other state-of-the-art quantitative approaches proposed for pottery classification. The workflow is able to take into account the entire surface of the pottery, not only a pre-selected spatially limited area.