Introducing our upgraded scanner: An inside look at the new coin recognition model
We are thrilled to announce that our coin scanning feature has just gotten better with our new and improved coin recognition AI model. The newly developed AI model is specifically designed for coin recognition, and its recognition accuracy has been significantly improved. Our beta testers have reported improved user experience, and we can't wait for all our users to try it out.
We have partnered up with Emblica, a company that specializes in developing state-of-the-art AI systems, to make this possible. Vili, a data scientist at Emblica, will walk us through the development process and share some insights on the challenges they faced and the most rewarding aspects of working on this project.
Developing a custom AI model for coin recognition
The starting point for developing this AI model for coin recognition was good, since the Coiniverse app was already used by tens of thousands of users. The project’s goal was clear: improve the scanner accuracy. As a data scientist, this was a good place to start, since there was a clear KPI (key performance indicator) to measure progress against, and it was clear that our solution would benefit people right away.
When a user takes a photo of their coin using the coin scanning feature, the AI system analyzes the image and compares it to the images in its database. When the coin is identified the system provides the user with relevant information about it, such as its name, country of origin, and other important details.
Emblica's task was to make an AI model specifically designed for coin recognition. The accuracy of the AI model depends on three things: the model architecture, the size and quality of the dataset it was trained on, as well as the quality of the image taken by the user.
The Challenges of Coin Recognition: Lighting and Reflections
Designing a custom AI system requires deciding on the model architecture and then finding a good set of hyperparameters that the model learns to solve the particular task. In this project, designing a good architecture was fairly straightforward, but getting the model to generalize from the vector (database) images to user-taken pictures was not. That required a fair amount of data science magic in enhancing the data.
Images taken by the users are often very different from the catalog images.
The most difficult aspect of this modeling problem was the difference between the image taken when scanning a coin and the image in the coin catalog. The quality of the image, the lighting conditions, and especially reflections from the metal surface can make the images very different. It is easy to take two completely different pictures of the same coin without moving the coin or changing the light: just taking the picture from a slightly different angle will change how reflections and shadows land on the coin's surface. From the model's point of view, these pictures are very different, and regular data augmentations are not enough to solve the problem.
How we can improve the coin scanning feature
The simplest way to improve the model is to have more examples of coins in different lighting conditions scanned with different devices. At the moment, user images are not used to train the model, and they never will be without users' informed consent. Emblica's model learns to recognize everything thrown at it, so gathering more data would be the easiest way to improve the scanning accuracy also on the rare cases.
The accuracy of the new AI model
The AI model's accuracy in identifying unique coins is fairly good, with a validation TOP-5 recall of over 95%. However, it doesn't identify the right coin every single time. For example, there are quite a few coins with Elizabeth II on them. So if a user scans a 50 Pence with Her Majesty on it, the model might give the user the right Elizabeth but on the wrong coin.
Despite the challenges, working on this project was rewarding since we saw step by step how applying some data science magic improved the model’s accuracy. We were able to create a model which learns reliably and generalises well. It's always very satisfying to see the effect of improvements with a short feedback cycle.
The new coin scanner will ease and speed up coin recognition and collection management for our users. We are really excited to hear what you think about the new scanner!
Feedback
We need your help
To understand how to improve the scanner in the future, please fill in this quick survey:
