The Research Of Methods About Character Recognition

Character recognition, a branch of pattern recognition, could increase the speed of collecting and inputting information greatly, meanwhile the work intensity can be lightened. With the rapid development of computer technology, character recognition technology was also improved and wide applied in many fields, the result of which is electronization of information disposal, that is a great deal of document information can be input into computer quickly, conveniently and automatically in time. Up to now, though the achievements in the study of character recognition are fruitful, they are still not enough to meet our daily needs. Therefore, it is of high importance to study the character recognition technology and improve recognition rate.Aiming at the methods related to character recognition, this paper did some research on the following aspects:1. Before recognizing, the picture should be pretreated. This paper studied some pretreating methods, such as graying, binaryzation, noise processing, Eigenvalue Extraction. Through comparing performance of different methods in graying and binaryzation, aiming at the bad binaryzation effect problem of vehicle license plate caused by uneven light, this paper bring up an improved binaryzation called local threshold method, by which we can got better binary image.2. In the image pretreatment, this paper introduced image edge-detection method first. Edge detection is very important to character outline extraction, license plate location, character eigenvalue extraction, etc. Based on the deliberate research of traditional Canny operator, this paper proved that the double threshold is extreme point, and put forward an improved Canny operator. Through carrying out a lot of experiment for picture increased salt and pepper noise, the method proposed in this paper can perform better than the traditional Canny.3. Shape Context is a shape description method. This paper proposed a improved shape context, using for recognizing the adhered and complicated CAPTCHA. The traditional method of extracting pixel points one by one and template matching, can only recognize simple CAPTCHA, while there is no efficient method to recognize the adhered and complicated CAPTCHA. Aiming at the problem that single character feature can't be extracted efficiently because of conglutination, this paper proposed improved shape context to extract character feature, combined with character global recognition, and realized the complicated CAPTCHA recognition.4. As one of the most popular neural network algorithm, BP neural network was widely used in the recognition of vehicle license plate. For there are some defects existed in BP, such as low convergence rate, easy to trap into local minimum, this paper improved it through introducing momentum factor and self-adaptive learning rate, improving activation function, and using LM algorithm. Through a great deal of vehicle plate experiments, this paper compared the performance of all kinds of improved BP algorithm, and also compare the method proposed in this paper with the traditional template matching method, proved that the algorithm proposed in this paper is better than the traditional template matching method.