Study On Handwritten Numeral Recognition Based On Dynamic Weight Multi-Classifier Integration

Optical character recognition is an automatization technology step by step developed since the twenties of the 20th century. All kinds of numeral information need to be inputed into computer in daily life, such as large-scale statistics , financial affairs, tax affairs, finance, mail sorting and so on. Therefore, handwritten numeral recognition has a wide prospect for application, and there is a great practical significance to study it.This thesis has studied and discussed the technology of handwritten numeral recognition, and proposed a new handwritten numeral recognition method based on dynamic weighted multi-classifier integration. The recognition method adopts BP neural network as the kernel classifier after the image pre-processing and character feature extraction, then integrates the neural network output vectors as the final recognition result with dynamic weighted multi-classifier integration.For pre-processing, this thesis has been studied emphatically on the uneven illumination characters image's binarization problem, then an edge zero-crossing binarization algorithm was presented based on LoG operator. This algorithm can adapt to the interferes such as uneven illumination, noise, and the binarization images can keep the original shapes of the aims.For feature extraction,three kinds of effective feature vector are extracted.They are original pixels vector features, macroscopical features, macroscopical features and microcosmic features respectively.Through analysing and researching , the different classifier's classification results show a good complementarity by utilizing three kinds of different features vectors to carry on the classification of characters.The classifiers have adopted BP neural network.After researching the trainning algorithm of BP neural network, a synthetical improved trainning method was presented to speed the network trainning and guarantee the network convergence. Using three kinds of different feature vectors as the input of the neural network, three neural network classifiers with different network structure were given.The dynamic weight multi-classifier integration method was proposed to integrate the three classifiers and to improve the system performance and recognition precision.