COMPARATIVE ANALYSIS OF YOLO AND SSD MODELS FOR NEPALI CURRENCY DETECTION

Abstract

Currency detection is one of the critical tasks in numerous applications such as counterfeit detection, automated cash handling systems, banking systems, currency monitoring systems, and money exchange machines. Recent advancements in computer vision and deep learning have introduced various effective object detection models, such as, SSD (Single Shot Detector) and YOLO (You Only Look Once) as promising solutions for automating currency detection. This project aims to perform comparative analysis of SSD and YOLO models for the task of detecting Nepali Currency, encompassing both coins and paper notes along with counterfeit detection. Banknotes in circulation in Nepal include: 5, 10, 20, 50, 100, 500, and 1000 rupee notes. Coins in circulation in Nepal include: 1 and 2 rupee coins. To facilitate this study, we created a comprehensive dataset of annotated images of Nepali denominations which are: 'one', 'two', 'five', 'ten', 'twenty', 'fifty', 'hundred', 'fivehundred', and 'thousand' along with the ‘Fake’ ones. These custom datasets consisting of 10 classes is utilized to train the SSD and YOLO models employing dedicated GPU resources and leveraging pre-trained model. The models are evaluated using various metrics of model’s accuracy, such as mAP (mean Average Precision), FPS (Frames Per Second), and training loss to insight their performance, robustness, and computational efficiency. In the context of our dataset, YOLOv8 outperformed the other models with mAP@.5 of 0.994, mAP@.5-.95 of 0.911, and inference speed was 3.448 FPS. Thus, the YOLOv8 trained model was implemented in Jetson Nano. This project contributes to the field considering the specific challenges and characteristics of Nepali currency in currency detection and counterfeit detection tasks.