An Automated Pattern Recognition Based Approach for Classification of Soiled Paper Currency Using Textural and Geometrical Features

Abstract

Research for automatic recognition and identification of paper currency (banknote) has gained popularity in recent years due to its potential applications, e.g., electronic banking, currency monitoring systems, money exchange machines, etc. Existing research work for identification of currency has some constraints that limit their accuracy. We are proposing a pattern recognition-based approach for the classification of Pakistani paper currency. The dataset used for our research work consists of 1750 banknotes, including light variated, torn, worn, dirty, and marked banknotes. The proposed approach was based on extraction of 371 textural features from entire image, as well as from 4 regions of interest. High dimensional feature set was then reduced to most discriminating features. Four classification models, i.e., K*, LogitBoost, PART, and Random Forest were used to evaluate the accuracy of our proposed approach. It was observed that using region of interest with reduced feature set resulted in better performance and lesser computational time as compared to existing approaches. The highest accuracy achieved was 100 % with Kstar classifier. The novelty of our research work lies in the fact that the proposed approach was capable of successfully classifying banknotes, even when the denomination was occluded or completely missing, as compared to existing approaches.