Host: The Gemmological Society of Japan
Pages 1-2
Fossa Magna Museum in Itoigawa City has been struggling to keep up with the increasing demand for its stone identification service. To address this issue, a study was conducted to develop a jade identification device using photographs of shore pebbles. The stones used for learning were pebbles collected along the coast of Itoigawa, with jade pebbles collected with the cooperation of museum staff. The study made use of the TensorFlow machine learning software library developed by Google and the image classification and object recognition architecture NASNet for transfer learning.
After allowing NASNet to study around 13,000 photographs of stones, the software achieved a successful identification rate of 96% for jade and non-jade stones when working with the teacher data. The engine achieved a 95% success rate when identifying 20 photographs of jade and a 100% success rate when identifying 13 photos of non-jade stones. The study shows that the identification of jade and non-jade stones is possible using artificial intelligence and deep learning. Future research will focus on improving identification rates through the provision of additional teacher data, teaching the engine to recognize each rock type, and trials in using smartphone photography to provide image samples.
