Machine learning-based mobile app to recognize card numbers, ensuring fast and accurate data extraction for seamless digital transactions.
Time
Client
Technology
Industry
While developing a face recognition app, we explored the feasibility of building our own character recognition solution.
Our Proof of Concept demonstrated that we successfully trained a machine learning algorithm capable of recognizing diacritical letters. Although many card recognition solutions exist, their capabilities proved insufficient for our needs.
Upon reviewing available options, we found that none met our specific requirements, which included easy integration with iOS, seamless operation on both iOS and Android, support for Polish characters, and adherence to budget constraints.
The application processes images captured by a smartphone camera, analyzing each video frame with a machine learning model that operates efficiently on iOS and Android, both online and offline.
It detects and classifies digits and letters, using probability-based decision-making to optimize accuracy. Within a second, dozens of frames are processed to enhance recognition precision.
To improve efficiency, the system was trained to identify specific card fragments, focusing on the card number and expiration date. The number is then verified using the Luhn algorithm, while the app also detects card providers such as Visa, Mastercard, and American Express and validates the expiry date.
The main feature of our application is the dynamics of the extraction algorithm. The analysis of the image takes place with each video frame provided by an iOS or Android system. The system merges the data coming from the following video frames, thanks to which the recognition process is faster and more accurate as recognised pieces of data complement one another. In the next steps, the application would analyse the image only when it recognises a card in the eye of the camera.
We also plan to improve the accuracy and the performance of our card recognition algorithm using Neural Network API and GPU delegates. 89.7% is the accuracy of our proof of concept solution for card recognition
