Abstract

Coins are used in our daily lives for a long time with less depreciation than paper currency. Detecting counterfeit coins visually is a challenging way with lots of errors. This thesis investigates advanced machine-learning techniques to differentiate between counterfeit and genuine coins with a small dataset. It focuses on the implementation of few-shot learning. This study is applied to two different types of datasets. The first dataset contains the images converted to grayscale, and the second dataset contains the four slopes images. As the detection of counterfeit coins is challenging due to their high similarity with genuine coins, more features are required before pre-training the neural network.
For this study, 2,474 labeled images from the CENPARMI dataset belonging to 22 different classes were used. To enable experimentation, the dataset was split into two parts: a Main Dataset (Dm) and a Target Dataset (Dt). We used a pre-trained model, which learns from the Dm and adapted it to Dt. The Inception V3 network was fine-tuned in the main dataset to learn general coin characteristics. This knowledge was transferred to the target dataset to learn new coin types from a few images. FSL using Siamese networks and contrastive loss was used. The algorithm performance was evaluated using the total accuracy with different epochs and different batch sizes to earn the optimum of them, and also the precision and recall and F.score per class.
It is shown that the accuracy of our method in epoch 20 is optimal. At this point, the model achieves a high level of accuracy (92.13% for grayscale images and 94.73% for SMMIG images). the model trained with a batch size of 32 achieves the highest accuracy of 92.13% for the grayscale dataset and 94.73% for the SMMIG dataset, indicating that moderate batch sizes contribute to optimal performance.