Fossil Shark Tooth Classification Using Deep Learning for Paleontological Heritage

The paradigm of Deep Learning is entering all domains of knowledge, allowing for the creation of data analysis tools that assist subject matter experts. Paleontology is also experiencing this trend. In this study we investigated the ability of Convolutional Neural Networks (CNNs), a particular class of Deep Learning algorithms specifically crafted for computer vision tasks, to classify images of fossil shark teeth collected from various images datasets. We developed and trained from scratch a CNN, named SharkNet, for classification of images containing a single shark tooth, customizing its complexity to our needs. A specific dataset was built in order to train our CNN, composed by more than one thousands images representing 10 species of shark teeth. Images come from online datasets as well as from the paleoichthyological collection of the G.A.M.P.S. Geopaleontological Museum (Italy). SharkNet showed good performance, reaching a mean accuracy of 88%. The goal of the present project is two-fold: on the one hand, we aim to demonstrate how Deep Learning algorithms can be applied to assist in the creation of tools for use in research settings; on the other hand, we hope to stimulate reflection on how this technology can be exploited to find new ways for fruition by paleontological exhibits, for example by developing mobile app to be used by visitors.