A generic, reference-based approach for the detection of counterfeit banknotes using orb feature matching, color histograms, and aspect-ratio analysis

This work addresses the prospective counterfeit currency detection in the future with image processing mechanism and presents a prototype, cost effective, reference-image based, visual software application adopting an interpretable, threshold- based decision model. Importance of such work comes from the fact that quality of counterfeit banknotes is rising rapidly due to evolution of printing and scanning technology and specialized detection equipment are scarcely available for small commerce to install. As a result, a software-only, open-source app solution is a straightforward and achievable alternative. The study is an applied experimental research, based on which the model was developed in Python with OpenCV along with GUI in Tkinter and applied on image samples of two banknote types (10,000 and 250 dinars) which includes original and counterfeit/out of sync notes, undergoes experimental tuning of decision thresholds before final validation.The numerical results of Tables indicate that for the 10,000 dinar value genuine notes obtained very high feature_score scores (e.g. 592 and 510) with?virtually zero aspect_diff (0.000 and 0.004) while counterfeit notes reached significantly smaller feature_score scores (47 and 39) and larger aspect_diffs (0.063 and 0.058). Yet, an equally crucial false-alarm case was noted, in which genuine banknote with feature_score = 592 and aspect_diff = 0.000 was classified as Likely FAKE or MISMATCHED because a color_similarity score of 0.765 had fallen short of the set threshold – this false alarm despite that color indicator being highly sensitive to illumination and imaging conditions. For the 250 dinar value, genuine samples also had feature_score scores between 110 and?135 and high color_similarity scores (0.93–0.96) with relatively low aspect_diff (0.012–0.018), while counterfeit samples had lower feature_score (28–40) and color_similarity (0.74–0.80) and higher aspect_diff (0.065–0.070). The model resulted in surprising production results, with a 91% accuracy for the 10,000 dinar (41 samples were correctly classified?among 45) and 93% for the 250 dinar (42 out of 45). The confusion matrix for the?10,000 dinar class (TP = 28, FN = 2, FP = 1, TN = 14) reveals that the misclassified samples were few, and the model showed a relatively conservative decision rule (i.e., a slight bias toward rejection in uncertainty). In addition to its numerical performance, the model can be interpreted at two levels: (i) visual output of feature matches and (ii) a?textual report describing the values of the numerical indicators and the justification of the final classification decision. These features make the proposed system a good educational and experimental tool as well as a powerful base for?further extensions with bigger databases, more precise threshold tuning, or the integration of machine learning and deep learning classifiers while preserving the analytically driven core of the proposed model.