Automated Honduran Banknote Image Classification using Machine Learning

Abstract

A new L200 banknote arrived to commemorate the bicentennial of Honduras. This made necessary to update automated methods for banknote image classification. The goal of this work was to develop algorithms that take centered images of banknotes and determine their denominations and visible sides. Two classification methods are presented. The first one uses local descriptors such as ORB or SIFT to match keypoints between the input image and templates to create feature vectors, and the images are then classified using Support Vector Machines or Random Forests. The second method is a Convolutional Neural Network (CNN) called LempiraNet, where transfer learning was used to deal with the limited data available. In both methods, image preprocessing can be used to locate the banknote to make its classification easier. To evaluate the effectiveness of these methods, two sets of 412 and 265 images were used for training and testing respectively. Multiple configurations were considered per method, and each one was evaluated in terms of recall, precision, F1 and run time. It was found that both methods reached at least 98% F1 score when using image preprocessing for locating the banknote in the input image. Also, it is observed that SIFT has better performance than ORB. In terms of run time, LempiraNet was at least 20 times faster than the other method, making it usable in real applications.