Towards Reliable Currency Recognition: A Hybrid CNN-KNN Framework for Indian Banknotes

Abstract

This paper introduces a classification model developed for Indian currency note identification and classification of five denominations, namely Rs. 10, Rs. 50, Rs. 100, Rs. 500, and Rs. 2000. The training dataset contains 12,050 images, and the overall accuracy achieved is 95.02%. The model shows high precision, recall, and F1 scores across all denominations. For Rs. 10, precision was 94.83%, recall 93.22%, and F1-score was 94.02%. For Rs. 50, there was precision at 95.14%, recall at 96.71%, and the F1-score at 95.92%. For Rs. 100, precision and recall were 95.18% and F1-score was 95.18%. For Rs. 500 and Rs. 2000, precision, recall, and F1-score was at 94.94% and 95.00%, respectively. All denominations had achieved 98% accuracy. With coverage of all the currency classes with a robust set of training data, the developed model has assured its use for real-life applications, especially in the domains of automated recognition and classification of currencies. Given that the system performed with almost 98% accuracy across every denomination, this proposed model ensures reliability and an efficient solution for a very important task related to currency identification.