Handwritten Character Recognition Based On Support Vector Machines

Optical character recognition (OCR) means reading text from paper and translating the images into a form that the computer can manipulate (for example, into ASCII codes). Handprint character recognition is a branch of OCR and plays an important role in many fields such as post office system, bank system and the input of cells and PDA. Because of its changeability, handprint character is more difficult to be recognized than machine print character. SVM (support vector machine) which was born in last century 90s does a good job in many fields including character recognition. But training a SVM needs a lot of time and memory especially when the database of training patterns is large. This paper proposes a new algorithm: select those patterns that are most likely to be the support vector and train on them and then we can save a lot of time.In this paper our main work is as follows.1. The relationship of distance of training patterns between input space and feature space has changed because of the kernel function. So we analyze the relationship of distance between two spaces.2. Based on the analysis of relationship of distance between the two spaces, this paper proposes a training pattern selection algorithm with the property of KNN (K Nearest Neighbors). This paper analyzes how many patterns will be selected. A lot of time will be saved because in the proposed algorithm just a small part of training patterns will be selected while most of the support vectors also will be selected.3. The experiments done on the database of MNIST and UCI show that the time consumption of train SVM on selected patterns decrease a lot while the error rate of the recognition is nearly the same as we train SVM on all the patterns using SMO algorithm.