A Multistage Neural Network System For Handwritten Numeral Recognition

The handwritten numeral recognition is a technology, which auto-recognizes the handwriting Arabian numerals via machines or computers, and a special field in the Optical Character Recognition technology. The handwritten numeral recognition research is greatly general-purpose and significative, because of the universal Arabic numerals. On the same score, the handwritten numeral recognition technologies are playing an important role in a number of automatization systems.In this thesis, a handwritten numeral recognition simulated system based on the multiply neural network combination is proposed to recognize isolated handwritten numerals. Some correlative technologies faced on the off-line recognition is investigated and researched. The whole simulated system is combined with the following four modules. First of all, the input of whole system is depend on two modules such as image preprocessing module and characters feature extraction module. And then the Back-Propagation Neural Networks is designed as classifiers for numeral recognition, namely classification module. Finally, the decision module, namely multi-classifiers fusion module, will product final result of numeral recognition through fusing the output of each classifier. The main works is focused on the characters feature extraction and fusion of multiply classifiers. In the feature extraction module, some potential feature extraction method were investigated and test via simulated experimentation.According to these experimental result of different feature sets (or vectors), one set of global feature and two sets of local features are proposed as the main parts of feature set, respectively. While some lower dimension of features are used to the assistant parts of feature set. The experimentation proved that the proposed three feature vectors is produce with the predominance of the single features for classification. In the multi-classifiers fusion module, a new multi-stage neural networks fusion approach was developed and compared with some existed fusion approaches via simulated experimentation. The experiment of recognizing handwritten numerals illuminated that the multi-classifiers fusion scheme is better than each isolate classifier, and the fusion method proposed by the thesis is better than others.In the chapter 1 of this thesis, the state of the art of handwritten numeral recognition and its application and potential application in future are presented. Some of the preprocessing technologies for character recognition are presented, respectively in the chapter 2. The chapter 3 mainly contains feature extraction approaches using to the simulated system and experimentation. In the chapter 4, neural network theory and designed method of network structure for handwritten numeral classification are introduced and experimented. In the chapter 5, some useful classifiers fusion methods are presented and compared by experimentation. Finally, the numerical experiments expose a competitive result that the correct recognition rate is 98.63% in the MNIST.