Classification of unexposed potsherd cavities by using deep learning

Abstract

The aim of the Deep Research for Akkon project is to use machine-learning techniques, specifically convolutional neural networks (CNNs), to analyze the large corpus of archaeological pottery fragments belonging to the Late Jomon period (ca. 4000 to 3200 aBP) in a non-invasive and non-destructive manner. In Japan, recent research conducted by using the ”impression method” revealed that ceramic vessels belonging to that era contain impressions of plant seeds. Studying these impressions allows archaeologists to better understand the cultural practices of the Jomon people and the path of rice diffusion in Japan. Most of the analysis was conducted by visually inspecting X-ray images. However, results based on using only X-rays are often inconclusive, leading researchers to use other invasive techniques that damage the potsherds. In this paper, we present a method that classifies X-ray images of potsherds by using deep learning. A dataset composed of 1036 images with seven classes was used to evaluate our approach. We generated different models using different CNN architectures and parameters and analyzed them separately. Then, we applied techniques such as ensemble and test-time augmentation to evaluate how such techniques impact our model. The best result was obtained by combining EfficientNetB7 trained with different parameters, achieving a classification rate of 90%.