Research Of Handwritten Digit Recognition Based On Feature Fusion And Neural Networks

Handwriting numeral recognition (HNR) is a technology which uses the digital computers to recognize the Arabic of optical character automatically. It's a branch of OCR (Optical Character Recognition). HNR can be applied in a very various fields, such as, in postcode, financial statements, bank check, various documents, census information etc. As it often involve in accounting and financial fields, so it request for c which is the major difficulty we are facing. And the system demands for high speed to process the mass data.Handwriting numeral recognition contains two parts, one is offline and the other is online,and the offline handwritten numeral recognition is the most difficult one because it doesn't have any real-time information. Therefore offline handwritten numeral recognition is much more challenging.In this paper, the first chapter describes the present research status, research methods, prospects and the evaluation of the handwritten numeral recognition. Five main categories of the pattern recognition are introduced. By analyzing the requirements of high-dependability and high-accuracy, we show the difficulties and the broad application prospect. The second chapter describes the pretreatment technology of the handwritten numeral recognition. In the digital image smoothing, median filtering Principle is introduced. The third chapter describes the statistic character oriented ways and structure character oriented ways of the HNR. The fourth chapter presents a multilayer feed-forward back-propagation network learning algorithm (BP algorithm), the inherent advantages and unique mechanism.Because of the complexity and limitation of traditional digital distinguish method, Script Digital Distinguish System based on feature fusion and neural network is proposed in this paper. Some different compensated features are extracted from the script digital image. There are coarse grid feature, intersection feature and skeleton structure feature. A multiple BP neural network is applied to classify the pattern with the fusion features. Finally, recognize numeral pattern using the BP network, which is well trained. The results show that compared with the traditional method, the new method can effectively fusion multiple features and have a high recognition rate. The fifth chapter exhibits Matlab software to simulate the system. The experimental results show that compared with the traditional method, the new one can simply using less features and can has better reliability and higher recognition rate, and the recognition rate can reach to 96.5%.