Turkish Lira Banknote Classification using Transfer Learning and Deep Learning

With the increasing exchange of foreign currencies due to globalization, there is a need for systems that can recognize and validate multiple currencies in real time. Such systems facilitate smooth international transactions and support the finance sector in dealing with diverse currencies. This study focuses on classifying Turkish banknotes using deep learning models. The dataset comprises 6901 images of six different denominations (5 TL, 10 TL, 20 TL, 50 TL, 100 TL, and 200 TL) under various conditions, such as flat, angled, curved, and bent. The proposed model imple ments pre-trained models, including VGG16, VGG19, DenseNet121, DenseNet169, DenseNet201, MobileNet, and MobileNetV2, to classify the images. Different image sizes (50x50, 100x100, 150x150, and 200x200) and optimizers (SGD, RMSprop, Adam, Adamax, etc.) were tested to determine the most effective combinations. The best result was achieved with DenseNet201 with an image size of 200 and the SGDoptimizer, achieving an accuracy of 98.84% in 12 epochs. Smaller image sizes (50x50) resulted in reduced performance for all models. In addition, models such as DenseNet169 and DenseNet121 also demonstrated high performance; however, MobileNetV2 struggled with smaller images.