| Paper currency identification, a scared samples, nonlinear and high dimensions pattern recognition problem is one of the difficult problems of modern pattern recognition and of specific research significance and practical value.This thesis studies the support vector machine and multi-class classification in the statistical learning theory, and applies the support vector machine into the machine paper currency recognition.The statistical learning theory and support vector machine have been introduced. The model selection, over-learning, nonlinear, dimensions curse and local minimum problems have been researched.The quadratic optimization algorithms of support vector machine composed by chunking algorithm, fix-sample algorithm and sequential minimal optimization algorithm. The sequential minimal optimization algorithm deals with the optimization problem byexplicit method, with high training rate and high identification rate. The sequential minimal optimization algorithm optimizes the standard single threshold error tolerance optimization condition, thus results in the time-consume to seek the second optimization sample. In this thesis, two thresholds, upper and lower are added to judge the optimization conditions. This avoided the original disadvantage, accelerated the training rate and improved identification rates.The compare is made between the one-versus-one, one-versus-rest and directed acyclic graph algorithm of presented support vector machine multi-class classification algorithms. The simulations show the advantage of directed acyclic graph algorithm over the others the identification rate.The support vector machine composed by least sequence and directed acyclic graph algorithm has been used in paper currency identification, shows the advantages of capability in dealing with scared samples, nonlinear and high dimensions. The experiments proved that support vector machine is of higher identification rate and more practical values than BPNN.
|
PDF Full Text Request