A transfer learning approach for real-time detection and classification of Indonesian coins

Abstract

Automated currency recognition plays an important role in banking automation, retail systems, and assistive technologies. While banknote recognition has been extensively studied, coin recognition remains challenging due to small object size, metallic reflectance, visual similarity across denominations, and circulation-induced wear. This study proposes a real-time system for detecting and classifying Indonesian coins using a transfer learning–based deep learning approach. A curated dataset was developed to address the lack of publicly available training data for this domain. The model was initialized with pretrained weights and fine-tuned to adapt to the specific coin classification task. Experimental evaluation on an unseen test set demonstrates high detection accuracy while maintaining real time inference performance. Qualitative analysis under challenging conditions—including glare, low illumination, occlusion, and coin wear— reveals operational limitations and defines robustness boundaries. The findings confirm that frozen-backbone transfer learning provides an effective and computationally efficient strategy for adapting state-of-the-art object detectors to low-resource, domain-specific currency recognition tasks.