Rupiah Banknote Denomination Classification Using CNN with MobileNetV2 Architecture

Abstract

The automatic identification of Rupiah banknote denominations using digital image processing and deep learning is a pivotal focus in financial technology, offering potential support for visually impaired individuals. While deep learning models achieve high accuracy, standard approaches often overlook computational variations and the challenges posed by the physical condition of banknotes in real-world scenarios. This study implements a Convolutional Neural Network (CNN) with the MobileNetV2 architecture to classify seven Rupiah banknote denominations. To enhance scientific novelty and address the limitations of basic replication studies, this research evaluates model performance through a multi-resolution framework (64×64, 128×128, 224×224, and 256×256 pixels) and validates the system against an external, non-ideal dataset. The dataset, comprising 2,398 images, was partitioned using a robust three-way split strategy: 64% for training, 16% for validation, and 20% for testing. Evaluations using a detailed confusion matrix indicate high multi-class classification stability. Furthermore, robustness tests demonstrate the model’s viability under realistic deployment constraints, such as folded, wrinkled, tilted, and unevenly illuminated banknotes. These findings establish a solid technical foundation for future mobile-based assistive applications designed to support the financial independence of visually impaired individuals.