Off-line Handwritten Digit Recognition Based On LVQ Neural Network Research

Off-line handwritten digit recognition technology has become the important researchtopics of researchers in recent years because it’s widely used in the documents, statistics,notes, emails, etc., and the Arabic numerals is the only general character in the world. Withthe development of social finance and other industries, our requirement for off-linehandwritten digit recognition accuracy is becoming higher and higher. But due to differencesin the human factors, handwritten numbers constitute the diverse features and thus have agreat deal of randomness, and these may affect recognition results, so the research ofhandwritten numbers is full of challenges and practical values.Off-line handwriting recognition system is mainly using the following methods:Firstly, this paper first introduces the necessity of off-line handwritten digit recognitiontechnology, the applications and research status quo of handwritten digit recognition, andanalysis of the difficulties of off-line handwritten digit recognition. The prophase processingof off-line handwritten digit recognition consists of image pre-processing and featureextraction. Pretreatment process is the process of image gray, binarization, denoising,normalization, and bone mineralization of the characters. As to the feature extraction aspect,this article uses the characteristics of regional segmentation and reorganization,and that is anew feature extraction method which has good generalization performance, so it can adaptto the characteristics of large handwritten digital changes better;Secondly, after the research and analysis of the basic principle model of neural networkmodel and algorithm, this paper puts forward the self-organizing competitive neural networkmethod is applied to handwritten digit recognition, and designs a kind of off-line handwrittendigit recognition system based on LVQ neural network method. According to the requirementof the handwritten numeral recognition system and the need to achieve, this paper analysesand studies the LVQ neural network input/output node number and determines the number ofhidden layer neurons, and analyzes deeply the weighting coefficients of LVQ neural networkand the way of working;Finally, this paper has realized the LVQ neural network based on Matlab digital imagerecognition, recognized USPS handwritten numerals sample library and five groups ofdifferent artificial handwritten digital respectively recognition process of writing, and compared the experimental results with the traditional BP neural network.In the system’s test results, LVQ neural network for USPS digital sample libraryrecognition rate reached97.4%accuracy rate, and BP neural network is95.7%accuracy rate.LVQ neural network for five handwritten digit recognition accuracy reached100%accuracyrate, while the BP neural network training five handwritten digits, the numbers1,3,8Department have each appeared a false acceptance, its recognition accuracy is less than LVQneural network..The experiment result has proved that the LVQ neural network has the lower deterrentrate when compared that of BP network, and its structure is simpler, has no normalizedrequest to input vector, it does not have the local minimization problem which may occur inthe BP network. Therefore, the LVQ neural network has more obvious advantages than thetraditional BP network, and has a more practical application value.