An Effective and Accurate Ensemble Model with Transfer Learning for Automatic Currency Detection
Senthil Pandi S, S. Shamili Shanmugapriya, U Surya, Gayathri S S · 2024
Systems of automation, from money changers down to computerized banking, are supposed to be integral in living today. One of the astounding features of automated currency identification systems is their ability to decrease human effort to ascertain the cash value. This is important, considering the challenges that illegible or deteriorated currency notes can cause through errors. The choice of appropriate algorithms and methods to be used for feature extraction and selection is a critical factor that determines the level at which these systems achieve effectiveness. Development of such systems should make them easy to comprehend, user-friendly, and fast to implement, guaranteeing general adoption and utility for any number of operational situations. The importance of currency in facilitating seamless financial transactions and improving operational efficiency in different situations emphasizes that effective identification and processing of currency can be done with precision. In these regards, the authors have proposed in the present study a new approach for automatic currency detection using an ensemble model for which training is already done. This methodology is based on transfer learning to improve current models, hence fine-tuning them to tailor very well to the work and, at the same time, avoid the issue of overfitting. It uses data augmentation techniques such as image enhancement, rotating, scaling, and translation to improve the robustness and model performance. This paper contributes— with the presentation of comprehensive taxonomy of multiple pre-trained neural networks—that assesses their effectiveness in detecting currency. Furthermore, it evaluates the effectiveness with which an ensemble of these models works in this application area. The results of the performance evaluation for the suggested model make it very clear that it is extremely effective in cash detection, outperforming other models with a phenomenal accuracy rate of 97.93% than other models. This accomplishment demonstrates that their technique is both superior and practically viable, and it holds the promise of considerable breakthroughs in automated currency identification systems.