Bangladeshi Paper Currency Recognition Using Lightweight CNN Architectures

Paper note currency is the most frequently used way of completing a transaction globally, which applies to most developing countries such as Bangladesh. Hence, it is essential to recognize any currency notes within seconds to save time for an uninterrupted transaction. This paper aims to identify different banknotes of Bangladesh using the avant-garde convolutional neural network (CNN) models with transfer learning by utilizing the dataset at hand and updating and augmenting the dataset. To achieve a robust, lightweight, and efficient model for this research problem, we have used various augmentation techniques on a current publicly available dataset and applied custom hyperparameter tuning to various pre-trained CNN models to attain a maximum accuracy of 99.97%.