Development of an Automated Coin Grading System: Integrating Image Preprocessing, Feature Extraction, and ML Modeling

Abstract

For more than 70 years, the Sheldon Coin Grading Scale has been essential in quantifying the value of coins within the coin collecting industry. Traditionally, coin grading has relied on human graders who may deliver inconsistent results. This inconsistency leads to variations in coin values. In this thesis, we present an automated coin grading system that uses image preprocessing, feature extraction, and advanced machine learning techniques to predict the grade across different coin types. Our system employs synthetic reference masks to identify "expected" regions, like the contours of reliefs, and "unexpected" regions, such as surface non-uniformities. All detected significant elements and tiny elements, extracted from these regions, will serve as one of the feature sets. Additionally, we extract color histograms as another feature set to analyze color and texture in detail. Both feature sets from the obverse and reverse sides of the coins are processed using a multi-layer perceptron (MLP) model and a random forest model. The best-performing model is then selected to grade the coins by analyzing their overall wear patterns and color characteristics. Our grading system has demonstrated an accuracy of up to 91.3% in predicting the Sheldon Grading Scale across five coin types, allowing for a grading tolerance of ±4. For a single coin type (Franklin Half Dollar), it has achieved an accuracy of up to 95.1% with a tolerance of ±1.