Researchers from Skoltech (part of the VEB.RF group) and Central University presented an automated coin grading approach that minimizes the subjectivity of manual assessments and overcomes the limitations of computer vision algorithms typically used for this purpose. The new approach provides more accurate and detailed grading, covering a wide variety of coin types. The solution was presented in the high-ranking ACM Journal on Computing and Cultural Heritage and was further refined in two subsequent studies published in Expert Systems with Applications and IEEE Transactions on Instrumentation and Measurement.
In numismatics and investment, experts manually verify the authenticity and physical condition of coins and assign them a numerical grade on the Sheldon scale. This process is inherently subjective, leading to numerous attempts to automate it.
Although some methods can accurately grade specific types of coins, they have limited practical value. Additionally, grading coins is significantly different from standard image classification. Even if two coins are nearly identical in design, denomination, and mintage date, slight differences in condition and visual appeal may be enough to assign them to different grades. This is fundamentally different from classifying animals, such as cats or dogs. Furthermore, previous computer vision methods are insufficient for processing a dataset with different coin types.
“We built a dataset of several coin types and proposed a dual deep learning model with two varieties of ResNet50, as well as a domain-specific algorithm that analyzes the probability distribution of the models’ outputs using domain knowledge. The algorithm considers adjacent classes in the ordinal structure, such as coins that are uncirculated and those that have been minimally circulated, which have overlapping visual features,” says Melaku Getahun, a PhD student from the Skoltech “Engineering Systems” program.
In the industry, classifying coins as one of the basic categories, such as “Uncirculated” or “Minimally Circulated,” does not accurately determine their value. Therefore, the team undertook a second study focusing on valuing “Mint State” coins.
The novelty of the study lies in its vision-language model that simulates the real-world practice of simultaneously evaluating both sides of a coin using a dual-input image encoder. The model’s language component is highly useful for grading features that lend themselves more readily to textual description. Since this is particularly true for uncirculated coins, the developers of the algorithm added a grading prompt and trained a text encoder, which was then used to calculate the cosine similarity between the textual and visual representations.
In their third study, the team improved efficiency further by integrating the vision-language model and a hierarchical transformer. This resulted in higher evaluation performance.
Compared to previous research, the approach presented by the Skoltech and Central University team has significant advantages in terms of quantitative grading precision.
“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 minimizing computational costs will help create a transparent and reliable coin grading system that will be easily accessible to collectors, investors, and museum experts,” says Andrey Somov, the team leader and a professor at the Skoltech Center for Engineering Systems and Sciences and Central University.
Going forward, the team plans to reduce the computational complexity of their approach in order to lower its hardware requirements.
