Research On Scripted Number Recognition Technology Based On NN

The recognition of scripted numbers is an active research area in the automatic pattern recognition field. Aiming at solving the problem of only using one kind of recognition eigenvector in the current scripted number recognition processing technology, this thesis proposes a scripted number recognition method, based on BP NN and weight-function-based multi-classifier combination, which can improve the recognition speed and recognition ratio.Firstly, this thesis summarizes the scripted number recognition technology and processing procedure, introduces BP NN and routine BP algorithms, and improves the weight functions, activation functions, etc., of the algorithms.Secondly, according to the improved algorithms, this thesis designs the system based on BP NN scripted number recognition technology and methods, and aiming to solve the problem of image bi-value marginal noise during the preprocessing of the scripted number recognition, this thesis proposes a marginal zero cross bi-value method based on the LoG arithmetic operators. This method can take advantage of LoG arithmetic operators to decide whether the pixels on the sides of point of zero intersection belong to background or target, and determine the belonging of uniform areas according to their neighbour areas. This method can adapt to the interruptions of uneven lighting and noises, and the bi-valued scripted number image can keep its shape pretty well.Finally, this thesis designs and realizs the BP NN-based scripted number recognition system. Through analyzing the recogniation ratio of this system under the universal evaluation standard, it is proved that the system can improve the precision and speed when recognizing scripted numbers. At the same time, experiment results show that its recognition speed is higher than single-recognizer strategy.