| Handwriting numeral recognition is a key step in information input. It is widely used in various domains, such as public security, revenue, traffic, finance. Recently none of the handwriting numeral recognition technology could reach 100% recognition rate though there were various technologies. For the purpose of using this technology in our daily work, in this paper, we not only do a lot of work to improve the recognition rate, but also design an information input and processing system based on handwriting numeral recognition technology.The difficulties in handwriting numeral recognition research are: 1) handwriting numeral figure is very small so that it contains little information, 2) different people has different habit in writing so that the numeral figure is daedal. What' s more, for some application, the request of exact handwriting numeral recognize is even stricter. In this paper, BP network arithmetic was selected as handwriting numeral recognition arithmetic after fully comparison in various recognition technologies. In fact, BP network maps input to output. In theory, it has the ability to implement any nonlinear mapping no matter how complex it is. Further more, it is good at handle problem which was very complex on its inner mechanism.The effect of handwriting numeral recognition is also depend on the quality of pattern set. In this paper, to get a high recognition rate, low mis-recognition rate classify neural network, about 50000 representative handwriting numeral image was collected from different people. After the collection, a list of pretreatment operation such as two-valued, removes noise, thinning, locate and cut, size normalization, was performed to build a input pattern for neural network. All numeral patterns are split as train set and test set, then train the neural network repeatedly, finally, we get a good classify neural network. It' s discovered in this paper when training that some patterns do 'harm' to neural network. We try to take away the 'bad' patterns and train and test again, then the identify rate and reject rate increase, while misidentify rate decrease.It' s our final goal that to use handwriting numeral recognition technology in actual application. To reach the goal we designed a information input and processing system based on handwriting numeral recognition technology. After the content of information card was scanned and saved in computer system using scanner, program The result of recognition would display on the screen of computer, and the numeral character which couldn' t be recognized would proofread by operators. Because of the existence of misidentify, a batch proofread method was designed to deal with those characters which were misidentified in this paper. The result of the test which using BP neural network as handwriting numeral recognition arithmetic indicate that the identify rate is over 96. 8%, the reject rate is below 2. 7%, and the misidentify rate is below 0. 5%. In the actual application, the batch proofread method could multiple the rates at handwriting numeral misidentify proofread. A skilled operator could process 100 characters one second or even more.The test in this paper indicates that: the recognition rate of BP neural network classifier is at less 96. 8 percent, refuse rate is less than 2. 7 percent; mis-recognition rate is less than 0.5 percent. When work in practical application, batch proofread could speed up handwriting numeral processing, and reduce the harm caused by misidentify efficiently. A skill operator could check more than 100 characters per second, the misidentify rate will less than 1 10000 after manual intervention.In this paper, first, an algorithm for numeral recognition and training using BP neural network is implemented and used in practical application; second, it' s discovered that 'bad' patterns affect the effect of handwriting numeral recognition, and a method to find out the 'bad' patterns given by author; third, a recognize result batch proofread software is designed, which increase the efficiency of proofread.In the future, I will make great efforts on quick neural network arithmetic and image pretreatment. I will do my possible on neural network and handwriting numeral recognition' s application and promotion. |
