Nominal Classification of Rupiah Banknotes Images Using Vision Transformer

Indonesian people with visual impairments have difficulty recognizing the nominal value of Rupiah banknotes. They often need to rely on others to help them recognize banknotes. Although blind codes on Rupiah banknotes can assist visually impaired individuals in identifying its nominal value, their effectiveness diminishes as the condition of the banknotes deteriorates. In addition to blind codes, visually impaired individuals often rely on differences in size and color. However, the size differences are minimal, and the similarities in color and imagery across Rupiah banknotes make this method unreliable. Recent advances in computer vision offer potential solutions to this problem. Previous studies have explored image processing and CNN-based models for Rupiah banknotes classification, but none have utilized Vision Transformer (ViT) models. Moreover, existing datasets often fail to represent real-world conditions, such as crumpled banknotes. This study proposes a ViT-based model trained on a mixed dataset that includes both publicly available images and a self-collected dataset of crumpled Rupiah banknotes. Evaluated over five runs with different random seeds, the model achieved average scores of 0.997 on the mixed test set and 0.990 on the crumpled-only test set across accuracy, precision, recall, and F1-score. The results highlight ViT's potential for robust Rupiah banknotes nominal classification, supporting its future application in assistive technologies for the visually impaired.