Deep Learning Based Bangladeshi Currency Coin Recognition
Serajum Monira, Nurun Nahar Fiha, Md. Golam Moazzam, Md Musfique Anwar · 2025
In today's dynamic world, accurate recognition of currency coins is crucial for various financial activities, especially for visually impaired individuals who face significant challenges in identifying coins for daily transactions in Bangladesh as most of the Bangladeshi coins are similar in size, shape, and texture. This study aims to address these issues by developing and evaluating deep learning-based models for Bangladeshi currency coin recognition by utilizing convolutional neural networks, specifically the Faster Region-based Convolutional Neural Network (Faster R-CNN) model and standard CNN model using PyTorch. The pretrained ResNet-50 backbone in the Faster R-CNN extracts high-level features such as edges, textures, shapes, and patterns to classify coin images into 14 distinct categories, while the CNN model provides a simpler yet competitive approach. These coin images were captured under diverse lighting and environmental conditions simulating real-world scenarios. Evaluations are conducted on the trained model using a variety of standard and own-recorded datasets, including shifted, rotated, and translated images. We found an average accuracy of 91.18% for Testing dataset and 90.99% for Validation dataset using Faster R-CNN, whereas the CNN model achieves 84.04% and 81.07% accuracy on the Test and Validation sets, respectively.