Banknote Image Analysis Based On Manifold Learning

Banknote analysis technology is consisted of value recognition, counterfeitverification and quality assessment. The purpose of technology is to increase theefficiency and accuracy of banknote classification. Banknote analysis technologymaintains the neat and safety of the currency, normal order of the financial andcurrency circulation by detecting counterfeits and stained banknotes andassessing the quality of the banknotes. Banknote validation includes banknotesverification and quality assessment, which plays an important role in thebanknote analysis technology.Existing methods of banknote verification and quality assessment have somelimitation: the relationship between banknote verification and quality assessmentis ignored, and there is no suitable model for this relationship. The lack ofcounterfeits makes trouble for the training of the robust classification interface.Because of the limitation, the accuracy of the classification can’t be increaseradically. Manifold learning is introduced in banknote verification for the firsttime, and two novel methods are proposed to solve the above problems.The first method is based on isometric mapping(ISOMAP), which usingISOMAP to obtain the low dimension of the features of banknote images,arranging the samples with different qualities on the manifold and making iteasily to classify the samples on the manifold. Then, the method apply self-organized mapping(SOM) to the low dimension of the features, SOM maintainsthe topological structure of data and works well on the clustering of features.The second method is based on locally linear embedding(LLE), it judgeswhether a sample is counterfeit by the distance between the projection of thesample and the sample itself, and it assesses the quality of a sample by the length of the geodesic line between the sample and quality-clustering centers. themethod don’t compute the low dimension of the features, it is just a efficiencydistance classifier.To solve the problem of counterfeits shortages, a novel method is proposedusing image processing technology to produce counterfeits. The method canobtain enough counterfeits to training classification interface and solve the aboveproblem to some extent. The method also can adjust the levels how a counterfeitlooks like a real banknote, this makes it possible for a comprehensivecomparisons between different algorithms using different counterfeits.This article uses RMB and US dollar to do counterfeit verificationexperiment and quality assessment experiment, the result shows that the methodsproposed by this article can distinguish the genuine and the counterfeit withaccuracy above98%, and the results of the quality assessment coincides withthose judged by human’s subjective feelings.