Decomposing Relief Maps to Detect Counterfeit Coins Using a Hybrid Deep Learning Method

In this research, we take advantage of deep learning approaches to improve the performance of counterfeit coin detection. As most of the pre-trained networks accept three channels for their input, we propose a new method to represent the relief map (height-map image) in three geometrical forms with Steep, Moderate, and Gentle slopes. This can also make our proposed method more understandable for both coin and AI experts by an eXplainable AI strategy. We compensate for the lack of fake coins by using a proposed Generative Adversarial Network. Then, we proposed a hybrid method using fine-tuning pre-trained deep neural networks to detect fake coins and provide a rejection option to increase the reliability of the system. While a few fake coins available in this research are used in the training process, the model is mostly trained by the images that are generated as fake coins from genuine ones. However, the system produces remarkable results to classify the coins