SVM And It's Application On Car Plate Recognition

With the development of national economy, the amount of kinds of vehicles has increased rapidly, which not only made in deed our life convenient but also cause a lot of problems. Therefore, to adopt intelligent traffic sruveillance (ITS) has been the main measure and development direction of highway traffic and municipal traffic management today. In addition, car plate automatic recognition technique is the kernel of ITS. And accuracy and speed of recognition of characters are key index of this kind of technique. That is why it is so important to develop a kind of fast and accurate character recognition method.Traditional statistic mode for recognition, such as template matching and neural network, is under the condition of adequate samples, that is, could get sufficient recognition results only if the number of samples runs to infinite. Unfortunately, in practical issues, the number of samples is limited, and then ideal results can hardly be got on the basis of the existing methods. However, Statistical Learning Theory, which is specially designed for the case of small samples, supplied a better theoretical frame to the research of Statistical Pattern Recognition under thecircumstances of limited number of samples, and presented a new pattern recognition method—Support Vector Meachine(SVM).SVM was presented by Vapnik and his fellows in 1995 as a kind of new mechanic learning method, and was on the basis of Vapnik-Chervonenkis Dimension and Structural Risk Minimization theory. It can solve the problem of small samples, non-linear and high dimension pattern recognition. Recently, SVM has been applied in many fields such as face recognition, function approximation and probability density estimation.By the means of SVM, better recognition rate could be acquired in the case that relevantly less training samples has been sampled. At present, many articles about SVM has been published but there are few about how to solve practical problems in the application of projects about SVM. Besides, the selection of parameters of SVM has always been the research direction of many researchers. But until now, there is not a generally acknowledged and qualitative solving scheme. These problems above will be researched in this thesis.First, this thesis implemented the part of training and recognition of the car plate recognition system by SVM. Then on the basis of deep research in SVM, the method of optimized SVM parameter by PSO algorithm was presented, so a big difficulty of SVM parameter selection has been solved. To meet the demand of practical projects, a method of single-sample added training method has been presented, which could made SVM add new kinds of characters that needs to be recognized anytime you want. What's more, a new method to transform SVM output to confidence has been presented and recognition accuracy has been increased further by the set of sutiable rejection threshold.