Abstract

Analyzing 3-D height-map images leads to the discovery of a new set of features that cannot be extracted or even seen in 2-D images. To the best of our knowledge, there was no research in the literature analyzing height-map images to detect counterfeit coins or to classify coins. The main goal of this thesis is to propose a new comprehensive method for analyzing 3D height-map images to detect counterfeit of any type of coins regardless of their country of origin, language, shape, and quality. Therefore, we applied a precise 3-D scanner to produce coin height-map images, since detecting a counterfeit coin using 2D image processing is nearly impossible in some cases, especially when the coin is damaged, corroded or worn out. In this research, we propose some 3-D approaches to model and analyze several large datasets. In our first and second methods, we aimed to solve the degradation problem of shiny coin images due to the scanning process. To solve this problem, first, the characters of the coin images were straightened by a proposed straightening algorithm. The height-map image, then, was decomposed row-wise to a set of 1-D signals, which were analyzed separately and restored by two different proposed methods. These approaches produced remarkable results.
We also proposed a 3-D approach to detect and analyze the precipice borders from the coin surface and extract significant features that ignored the degradation problem. To extract the features, we also proposed Binned Borders in Spherical Coordinates (BBSC) to analyze different parts of precipice borders at different polar and azimuthal angles. We also took advantage of stack generalization to classify the coins and add a reject option to increase the reliability of the system. The results illustrate that the proposed method outperforms other counterfeit coin detectors.
Since there are traces of deep learning in most recent research related to image processing, it is worthwhile to benefit from deep learning approaches in our study. In another proposed method of this thesis, we applied deep learning algorithms in two steps to detect counterfeit coins. As Generative Adversarial Network is being used for generating fake images in image processing applications, we proposed a novel method based on this network to augment our fake coin class and compensate for the lack of fake coins for training the classifier. We also decomposed the coin height-map image into three types of Steep, Moderate, and Gentle slopes. Therefore, the grayscale height-map image is turned to the proposed SMG height-map channel. Then, we proposed a hybrid CNN-based deep neural network to train and classify these new SMG images. The results illustrated that a deep neural network trained with the proposed SMG images outperforms the system trained by the grayscale images. In this research, the proposed methods were trained and tested with four types of Danish and two types of Chinese coins with encouraging results.