An Integration of Transfer Learning in Modern Philippine Banknote Feature Detection
Joel C. De Goma, Marcus Julius Rabano, Dean Marlou Sanson, Kerstein Alyssa Tadena · 2024
This study focuses on the detection of counterfeit currency within the modern Philippine banknotes, specifically, the Php 500 and Php 1000 denominations of the New Generation Currency series. To address this issue, we propose an innovative solution that leverages the combination of MobileNetv2 and YOLOv5s, employing transfer learning for efficient feature detection. Our research involves a comprehensive evaluation of MobileNetv2's performance, comparing it to transfer learned ResNet-18 as a feature extractor, and the optimization of hyperparameters of the former model. We assess seven critical security features, while aiming to enhance counterfeit currency detection efficiency. Our methodology encompasses dataset acquisition, data augmentation, image annotation, and model customization. The results of our study demonstrate that MobileNetv2 achieves a mean Average Precision (mAP) of 92.8% for detecting 500-denomination banknotes and 93% for 1000-denomination banknotes, all while maintaining lightweight computational requirements. In contrast, ResNet-18 achieves a higher mAP of 93.6% and 95%, respectively, but at the cost of heavyweight computational demands. This research offers a promising and practical approach to significantly improve counterfeit currency detection in the Philippine banking industry by employing security feature detection technique.