Data Augmentation of Iberian Pottery Collections for Deep Learning Based Classification

Abstract

One of the most frequently found archaeological artifacts are ceramics. They can be used to distinguish between chronological and ethnic groups and to reconstruct economic history, showing trading routes and cultural relationships, especially in the case of ceramic vessels (Orton, Tyers & Vinci 1993; Kampel & Sablatnig 2003). Unfortunately, ceramics are fragile and therefore most of the vessels recovered from archaeological sites are broken. For this reason, the vast majority of the available material is in the form of fragments. Therefore, the available databases of complete vessels are too small to be analysed with deep learning (DL) techniques, especially in the case of wheel made vessels, which are all different. Therefore, we need mechanisms that allow us to generate sets of profile vessels that mimic real ones, taking the existing databases as reference. In this work, we propose the use of a Generative Adversarial Network (GAN), capable of modelling the probabilistic space of a given dataset and generating new samples, allowing us to increase the size of training datasets.