The Research On Handwritten Numeral Recognition Based On SVM

With the rapid development of science and technology,the processing speed of computer is faster and faster.Using computers to solve real life problems is a hot research direction. Handwritten numeral recognition exists in many areas, such as the e-mail classification,the document retrieval and the bank notes,so, handwritten numeral recognition is a more important research direction of pattern recognition. how to improve the recognition's accuracy and efficiency is a goal of us.In this article,there are four steps for researching handwritten numeral recognition.The first step, relative to the face and other objects, the vector's dimension of handwritten numerals is smaller, Therefore,we can train and recognize the samples by using support vector machine and multi-classification algorithm directly. In this paper, we choose four multi-classification algorithms which are the 1-(k-1) algorithm, the binary tree algorithm, the dichotomy and the voting method. Through the analysis of experimental results, we can see that the dichotomy and the voting method are superior relatively. However, the imperfection is that, If we want to get the high recognition rate, we have to need a lot of training samples, So,we must take a long time for training samples.The second step, we can reduce the dimension by using PCA,which can reduce the time of training samples. However, by reducing the dimension, the recognition's accuracy do not improve, This result may be related to samples and the data of sample vectors.The third step,we may begin with reducing the number of training samples, considering similar shape, we can choose the samples by the idea of similar shape-pairing. At last 639 training samples are selected,and recognize 100 samples by using the 639 training samples,the result is 87%,this result has a significant improvement relativing to the first step.The fourth step, we extend the idea of the third step,but new training samples are obtained by clustering analysis. At last we see,if we chooce the similarity distance+the middle distance algorithm of hierarchical clustering+the voting method,the he recognition's accuracy will arrive at 97% when the number of training samples is 1800. This result ensures the high recognition's accuracy, and greatly reduce the time of training samples.