Bangladeshi Paper Currency Recognition with Improved Dataset Using Vision Transformer

In this study, our primary objective is to present an innovative dataset designed to enhance the recognition of Bangladeshi paper currency through the utilization of the state-of-the-art Vision Transformer (ViT) model. Notably, the application of ViT in the realm of Bangladeshi paper currency recognition has not been explored by any prior researchers. Moreover, the limitations of previous datasets, characterized by simplicity, pose challenges as training models on such datasets lead to poor performance on real-world, noisy images. Visually impaired individuals face currency transaction problems in their daily lives. Addressing this, our study aims to have a meaningful impact on the lives of such individuals in Bangladesh. It is imperative to note that, with the exception of one dataset, all existing Bangladeshi paper currency datasets lack images of the newly introduced 200 taka notes, and the backgrounds of the available images are overly simplistic. To overcome the limitations inherent in publicly available datasets, we have created a novel dataset. This dataset incorporates images captured on complex backgrounds, various orientations, and diverse lighting conditions, rendering it more suitable for real-world applications. The efficacy of our approach is evaluated by assessing the performance of the Vision Transformer model on our dataset and the publicly accessible Augmented Bangla Money dataset, the only open-access dataset featuring images of 200 taka notes, totaling 10000 images. The outcome of our study reveals a remarkable maximum test accuracy of 99.93% on the Augmented Bangla Money dataset.