Convolutional Neural Network for Classification and Value Estimation of Selected Gemstones in Sri Lanka

Abstract

Gemstone classification and value estimation are considered to be tedious tasks encountered in the gem industry all over the world. This happens due to colour variations found in the same gem type which is often difficult to detect with the naked eye. This paper presents a machine learning approach to automatically classify the gem type by using an image and also to estimate the value of the stone using a few measurements. The proposed technique uses a microscopic image of a gemstone which is taken using a gemological microscope. A Convolutional Neural Network (CNN) is trained to classify gem type while features such as type, colour palette, shape and weight are used to estimate the value of the. This work creates a system that is capable of classifying and estimating the value of four types of gemstones, namely, Blue Sapphire, Yellow Sapphire, Amethyst and Cat’s eye. The results indicate that the proposed technique managed to classify the gemstones with the highest accuracy of 87% for yellow sapphires and 77% for blue sapphires. The yellow sapphires produced the highest accuracy in colour categorization which can be attributed to the high contrast of the images vailable. As such, it can be concluded that the quality of the original image is important in correctly identifying the exact colour of a gemstone.