License Plate Recognition System Architecture - Tigerwong

The license plate recognition system adopts two-layer feedforward neural network. According to the self text analysis, license plate character recognition is a small classification problem. The neural network with one hidden layer can approach any nonlinear number. Therefore, only two-layer neural network is needed to recognize characters. Because the characters on the license plate are divided into Chinese characters, letters and numbers, according to the specific situation, this paper designs four neural networks, namely Chinese character neural network, letter neural network, digital neural network and letter and number discharge neural network. Each pixel in the normalized character h image is used as an input of the neural network. The unified size of the character is 32 * 16, so there are 512 inputs in total. The number of output neural nodes is determined by the number of categories of the problem to be classified. For the Chinese character network, there are 31 Chinese characters for the abbreviations of provinces, municipalities directly under the central government and Baizhi District, 13 Chinese characters for military vehicles, and 51 Chinese characters for embassies, consulates, temporary cars and coach cars. The number of output nodes of the Chinese character network is 51; The number is from. To 9, there are ten digital networks. The problem solution in this paper refers to the weight and threshold of the network, that is, the purpose of network training is to find a set of optimal combination of weight and threshold. Generally, the license plate recognition system adopts the = binary coding method, but the license plate recognition system involves many parameters and requires high accuracy. The accuracy of binary coding is limited by the chromosome length. In this paper, real number coding is adopted, on the one hand, it meets the accuracy requirements, on the other hand, the problem coding significance is clear. 03 it is necessary to select excellent parents for various genetic operations, There needs to be a standard. Genetic algorithm is based on fitness function. In this subject, error is the fundamental standard to measure individual good and bad, that is, the smaller the error, the higher the individual fitness. On the contrary, the larger the error, the worse the individual fitness.