Smart Rupiah Recognition: A Mobile Machine Learning Approach for Visually Impaired Users

Abstract

Purpose: Despite advances in assistive technology, low-connectivity areas lack reliable solutions for visually impaired individuals, prompting this study to enhance financial autonomy in cash-based economies. This research addresses high fraud risks and the limitations of online tools like Be My Eyes, which fail in areas with only 40% internet access, by developing a 3MB MobileNetV2 model for offline Rupiah denomination recognition on low-end Android devices.

Methods: A MobileNetV2-based Convolutional Neural Network, optimized to 3MB via TensorFlow Lite quantization, was trained on 10,855 augmented images (rotation ±30°, flipping, Gaussian noise, σ=0.1). The Kotlin-based application integrates CameraX for 720p video and Bahasa Indonesia text-to-speech, with a “no object” class. The model was tested on 4–8GB RAM devices, validated through usability evaluations with diverse stakeholders.

Result: The model achieves 90% accuracy (F1-score 0.90) at 1000 lux, 85% at <50 lux, 80% at >60° angles, and 88% for “no object,” with 10ms latency. Self-supervised learning (SimCLR) on 2,000 worn notes improves accuracy by 3% (p < 0.05). Usability evaluations yield 95% session success, with TTS and UI Likert scores of 4.2 and 4.0..

Novelty: The 3MB MobileNetV2 model, with 10ms latency and 15% false positive reduction, outperforms YOLOv5 (500MB, 50ms), Vision Transformer (1GB, 200ms), and YOLOv8 (200MB, 30ms). This model shows potential for cross-currency detection throught preliminary exploration (e.g., USD and euro), which may advance edge AI and financial inclusion in developing nations.