Transfer Learning-Based Bhutanese Currency Recognition

Abstract

Currency recognition is an important field in the area of pattern recognition. Recent improvements in the field of deep learning have increased its ability to recognize complex features from images. The goal of this research was to create a first-ever Bhutanese dataset and train using transfer learning techniques. This paper proposes three different pre-trained models to train the Bhutanese currency datasets. A parameter fine-tuning was applied to get better accuracy from the custom dataset. The currency images were gathered using crowdsourcing techniques and then the data augmentation was performed to generate 1000 images per class. The datasets were split into the train, test and validation sets with the ratio of 70:20:10. The train and test sets were used while training the model and after the training, it was validated using the validation set. After training the model for certain epochs, VGG16 architecture outperformed other models with a training accuracy of 99.82%, a test accuracy of 99.12% and a validation accuracy of 95.5%. In the future, a greater number of images need to be included in the datasets and trained using other pre-trained models.

Keywords

• Ngultrum
• transfer learning
• fine-tuning
• deep learning
• Bhutanese currency