Evaluation of Deep Learning Models for Detection of Indonesian Rupiah

Abstract

This study compares the performance of current object detection models, namely YOLOv7-tiny, YOLOv8n, and EfficientDetd0, using YOLOv5n as the baseline model in addressing the challenge of Rupiah banknote detection. The challenge involves recognizing unique features on the banknotes, which may have higher complexity compared to common objects in object detection tasks. The dataset used covers 2022 Emission Year Rupiah banknotes, is manually created, and covers various real-world scenarios for comprehensive evaluation. This research also explores the impact of data augmentation to optimize model performance. Results show that YOLOv8 is the top-performing model, with mAP@0.5 scoring 0.995 and mAP@0.5:0.95 scoring 0.994 on the test data, also consistently maintaining high performance even without augmentation. YOLOv5 also showed impressive mAP scores of 0.995 and 0.973 with augmentation. YOLOv7, although it did not surpass YOLOv8 and YOLOv5 in accuracy, achieved good results, especially with data augmentation. In terms of inference time, YOLOv5 excels with 6.7 ms without augmentation and 6.5 ms with augmentation, emphasizing its efficiency. YOLOv8, although slightly less efficient, with inference time of 7.8 ms without augmentation and 8.2 ms with augmentation, provides higher accuracy. The choice between the two depends on the balance between accuracy and efficiency. This research also highlights the positive impact of data augmentation, especially in YOLOv5’s responsiveness to additional data. While EfficientDet is efficient in inference time and resource usage, it suffers in performance, especially without augmentation. This study attempts to develop a dependable method for identifying banknotes. By achieving this, the aim was to improve accessibility in financial tasks and everyday life, particularly benefiting those with visual impairments or other disabilities.