A Case Study on Neural Networks as Classifiers to Design Banknote Recognition Systems

Abstract

We proposed a banknote recognition method based on the adoption of CNN for feature extraction and FFANN as a classification. Banknote images are normalized to have the same size and fed into trained CNN models corresponding to the pre-classified size classes. When passing through the neural network, banknote features are extracted by the convolutional layers and feed into the FFANN classifier to identify its respective denomination as well as to verify its originality. Our experimental results using two-fold cross-validation on the Ethiopian currency dataset show that the proposed CNN-based banknote recognition method yields better accuracies than the method in the previous study. Although CNN-based classification has been used in various fields due to its high performance, it has the disadvantage of requiring intensive training with a lot of training data. However, it is often the case to have difficulty in collecting a lot of training data in actual experimental environments. Therefore, we have applied a data augmentation to increase the size of the datasets. Further studies are required to conduct on the CNN model by with better architecture like Google Net and ResNet with a very large dataset to classify and recognize Ethiopia banknote.