Dollar Recognition Based On Improved Principal Component Analysis And Modified Neural Network

With China's accession to the WTO and connection of banking and international finance service industry, foreign deposit and exchange, in particular the dollar counter trading business, have become indispensable. However, the number of dollar checking devices is small, and the reliable ones are even less. In addition, the counterfeiters are more and more rampant, so the current techniques need to be improved urgently, and the dollars' checking research has the dual challenging of theory and practice.After studying the current situation and research methods at home and abroad, dollars' features were analyzed and effective sensors were chosen to obtain useful hidden features. Then, after making preliminary investigations and discussions on the feature extraction and artificial neural networks,the dollar recognition method based on improved principle component analysis (IPCA) and improvably modified learning vector quantization (IMLVQ) neural networks was proposed.In the process of extracting features,the principal component analysis (PCA),which was widely used in the dollar's feature extraction,faced the problem of high computation complexity,inaccurate estimated covariance matrix from training images for dollar recognition and not considering the classified information for dollars' checking.So the improved principal component analysis(IPCA) was proposed.In the algorithm, the subspace basic vector extracted by PCA was substituted by the right singular vectors of training images, so that the transformation from the images to image vectors was avoided. Hence the computation was simplified significantly. Then, after making use of the energy compaction property of PCA and the property of Laplacian matrix, the algorithm almost extracted the dollars' checking information of the training set.In the process of designing classifier, the faults of learning vector quantization neural networks and its improved models,which were widely used in the dollar identification, were pointed out after extensive reading of papers and books. Improvably modified learning vector quantization(IMLVQ) neural network identification algorithms was proposed. It could effectively overcome the traditional LVQ algorithm's sensation to the initial values and improve its generalization ability, and also had better ability of learning.MATLAB simulation and C++ program design were made to compare and analyze results of this kind of model.The experiment indicated that this paper's algorithm gained better outcomes; thereby, it showed that the algorithm had advantage on dollar identification.