Research On Currency Recognition Based On Neural Network

Currency recognition, usually concerning with paper currency or bank note, is involved in recognizing the values and the fake. As a special kind of print, currency is different in size, pattern and material for the various par values. Moreover, currency may be stained or torn in use. These facts cause extreme difficulties in recognition, nevertheless considering about the challenge of high-speed recognition. This thesis, based on extraction and analysis on the feature of currency, presents the feed-forward neural network to recognizing the currency values, combines with the statistical analysis to recognize the fake and forms a set of methods for currency recognition.First, the currency recognition system and its technology situation are summarized; then the methods to extract and how to choose the currency features that best reflected the classified character are discussed; after that algorithms of values classification and fake identification are investigated respectively. The feed-forward neural network is provided to recognizing the currency values, which makes use of the capability to extract featuresautomatically and the error tolerance of neural networks. In this thesis, some kinds of learning algorithms of feed-forward neural network have been analyzed; later a scaled conjugate gradient algorithm is presented to train network, furthermore, modified training methods have been provided to improve the neural network performance. So we can utilize the special sensor data to create the currency feature collection via analyzing currency samples in statistics. The fake currency can be identified rapidly and exactly by this method.Not only the contradiction between exactness and speed has been balanced but also the requirements of practical application been considered, when the designing idea and concrete realization are presented. In this way, the algorithm can be applied in practical recognition system.Finally, some experiments of recognition for current RMB are given to analyze and verify the algorithms investigated in this thesis.