A new deep-learning approach could make the assessment of collectible coins more accurate, detailed and objective
Researchers from the Skolkovo Institute of Science and Technology, a partner of TV BRICS, and Central University have developed an automated method for assessing the condition of collectible coins. The approach is designed to reduce subjectivity in expert evaluations and overcome limitations of conventional computer vision systems when working with different types of coins, as reported by the institute's website.
The research covers the automated assessment of coins with different designs, denominations and dates of minting.
Coin grading traditionally relies on specialists who assess authenticity and condition before assigning a grade according to the Sheldon scale. Unlike standard image-classification tasks, coin grading requires the system to distinguish between objects that may look almost identical but differ slightly in preservation and visual appeal.
The researchers created a dataset covering several types of coins and developed a dual deep-learning model. The system analyses the probability distribution produced by the models while taking into account the ordered nature of coin grades. This allows it to recognise that neighbouring grades can share similar visual characteristics.
The team subsequently focused on coins belonging to the highest grading categories. A vision-language model was introduced to assess both sides of a coin simultaneously. Its text component helps describe characteristics that can be difficult to distinguish through visual analysis alone, particularly when evaluating coins that have not been in circulation.
In a third study, the researchers further improved the method by modifying the visual encoder and introducing a hierarchical transformer. This increased the accuracy of the assessment compared with the previous version of the system.
"Our approach can contribute to the automation and objectivity of coin grading within the numismatic community. We are pioneering this area of research in the hope that scientists from around the world will build on our findings. Efficiently achieving the desired level of accuracy while minimising computational costs will help create a transparent and reliable coin grading system that will be easily accessible to collectors, investors, and museum experts," said Andrey Somov, the team leader and an associate professor at the Skoltech Center for Engineering Systems and Sciences and a professor at Central University.
The researchers plan to further reduce the computational requirements of the technology. This could help make automated coin grading more accessible while maintaining the accuracy needed by collectors, investors and museum specialists.