ABSTRACT
Currently, the license plate recognition technology has been used in the parking systems and traffic monitoring systems in some developed countries. In allweather conditions the recognition accuracy is more than 95%. The license plate recognition technology development is slower in China.Although we achieved some results, but still remain in the experimental stage. Although the current laboratory recognition accuracy of 90%, in all-weather conditions recognition accuracy is less than 85%, far less than the actual application requirements. At present, the license plate recognition mainly has the following types of identification methods: template matching method, the statistical feature matching method and the neural network identification method. Template matching method of regular character recognition rate is relatively high, but in the case of the character deformation, the identification capability is limited. The feature statistical matching method in practical applications, when the characters appeared missing or fuzzy, the recognition effect is not ideal. The neural network recognition can effectively identify high resolution license plate. It has a strong ability of classification and fault tolerance. A lot of character recognitions in vehicle license are based on neural networks. The BP neural network is the most widely used network algorithm. BP neural network algorithm has some shortcomings, such as the slow convergence, local optimal, it is difficult
to determine the number of hidden layer nodes and the training process oscillation. For the above, this paper has made some improvements in the selection of the momentum factor, the learning rate improvements and modifications of the network weights. It can effectively prevent network may fall into the local minimum, and can adaptively adjust the learning rate, accelerate the convergence speed, avoid oscillation.
