Adaboost Algorithm-based Digital Recognition Technology Research And Application

With the development of computer science and the improvement of pattern recognition, it is possible to process a mass of notes and forms using computer, which is known as Optical Character Recognition (OCR) technology. Digit Recognition is a special problem in OCR, and it has a wild use in many fields .The basic theory of Aadaboost algorithm is to select several weak learners which has just slightly better accurate than random prediction and combines them into a strong learner. Adaboost arises from PAC learning model, which has been proved that a series of weak learners can be combined into a strong learner with highest accurate if training data is it adequate. The thesis researched on Adaboost makes some improvement to satisfy the need of digit recognition. A mayor improvement is that we construct a classifier structure of two layers, with which the classical two-class Adaboost algorithm mustn't be modified heavily to be a multi-class algorithm. In the first layer of classifier structure, we convert the output range from {0> 1} to [0,1], and raised the concept of "grade of membership". This thesis also add a new feature form to rectangle features and determine the adaptive features for digit recognition(positions, numbers, and forms) in the means of experiment. The digit recognition system based on adaboost of this thesis achieves low error rate and has advantage in the aspects of training time and hardware needs.This thesis also implements a evaluate system for medical record which successfully used the digit recognition algorithm based on adaboost in a real application. The system establishes a note rules to make the position determination of information blocks easier. This system also designs a high efficiency algorithm for orientation detection based on the tool of integral image. For the practibility of the digit recognition system, this thesis implement another digit recognition algorithm based on "open or close" to compare the results with Adaboost algorithm, which could reduce the risk of recognition error.