Work

Rupiahvision: Introduction to Mobilenetv2-based Rupiah for the Visually Impaired and Real-world Performance Gaps

Year: 2026
Type: article
Abstract: Visually impaired people often cannot distinguish Rupiah banknote denominations, and classifiers that score well on internal test data may still fail on everyday photographs. This study built RupiahVision, a mobile Rupiah recognizer, and measured the gap between internal and real-world performance. A pretrained MobileNetV2 with a frozen backbone and an eight-class head was trained on 3,173 images combining a public Kaggle dataset with the team's photographs, then served through FastAPI to a React Native application. Evaluation used 485 internal test images and 259 separately prepared real-world photographs. Internal accuracy reached 92.99%, whereas real-world accuracy was only 48.65%. Errors concentrated on the Rp2,000 prediction (91 of 259 predictions) and Rp20,000, while the non-banknote class kept a recall of 0.95. Internal accuracy alone therefore cannot establish readiness; training data diversity and consistent preprocessing at serving time should be prioritized.
Language: English
Open Access status: diamond
OpenAlex ID: W7235357093
FWCI:
Cites: 16
Cited by:
Related to: 10