Comparative analysis of machine learning approaches in Kazakh banknote classification

Abstract

Nowadays, smartphones seamlessly blend into every aspect of our lives, including as handheld assistants for individuals with disabilities. Therefore, this research addresses the need for a robust system that can classify Kazakh banknotes. By capitalizing on the availability of smartphones and the ability to integrate detectors with classifiers this study introduces classifiers of Kazakh banknote images specifically designed for banknotes ranging from 500 KZT to 20,000 KZT. It compares traditional and hybrid machine learning (ML) approaches, utilizing a dataset of diverse banknote images, aiming for both lightweight and high accuracy. Competitive performance is demonstrated by the traditional approach, enhanced by thoughtful feature engineering. The hybrid approach, utilizing features from a pre-trained ResNet-18 model, showcases remarkable accuracy and robustness. Evaluation metrics reveal significant achievements, with the traditional approach attaining 94.00% accuracy and the hybrid approach excelling at 99.11%. Model stacking, combining classifiers from both approaches, outperforms individual classifiers, achieving 95.00% and 99.55% accuracy for the traditional and hybrid ML approaches, respectively. Our methodology’s comparable outcome in classifying Thai banknotes and coffee beans roasting levels demonstrates their versatility in image classification tasks that rely on color differentiation, showcasing the potential beyond banknote recognition.