Abstract

Abstract
Local Image Patterns for Counterfeit Coin Detection and Automatic Coin Grading
Coins are an essential part of our life, and we still use them for everyday transactions. We have always faced the issue of the counterfeiting of the coins, but it has become worse with time due to the innovation in the technology of counterfeiting, making it more difficult for detection. Through this thesis, we propose a counterfeit coin detection method that is robust and applicable to all types of coins, whether they have letters on them or just images or both of these characteristics. We use two different types of feature extraction methods. The first one is SIFT (Scale Invariant Feature transform) features, and the second one is RFR (Rotation and Flipping invariant Regional Binary Patterns) features to make our system complete in all aspects and very generic at the same time. The feature extraction methods used here are scale, rotation, illumination, and flipping invariant. We concatenate both our feature sets and use them to train our classifiers. Our feature sets highly complement each other in a way that SIFT provides us with most discriminative features that are scale and rotation invariant but do not consider the spatial value when we cluster them, and here our second set of features comes into play as it considers the spatial structure of each coin image. We train SVM classifiers with two different sets of features from each image. The method has an accuracy of 99.61% with both high and low-resolution images. We also took pictures of the coins at 90˚ and 45˚ angles using the mobile phone camera, to check the robustness of our proposed method, and we achieved promising results even with these low-resolution pictures.
Also, we work on the problem of Coin Grading, which is another issue in the field of numismatic studies. Our algorithm proposed above is customized according to the coin grading problem and calculates the coin wear and assigns a grade to it. We can use this grade to remove low-quality coins from the system, which are otherwise sold to coin collectors online for a considerable price. Coin grading is currently done by coin experts manually and is a time consuming and expensive process. We use digital images and apply computer vision and machine learning algorithms to calculate the wear on the coin and then assign it a grade based on its quality level. Our method calculates the amount of wear on coins and assign them a label and achieve an accuracy of 98.5%.