The Application Of Feature Extraction And Selection In Handwritten Digit Recongnition

Handwritten digit recognition is widely applied in a lot of fields. In most of the fields, a handwritten digit recognition system with high rate of correctness and a low rate of errors is requested. At the same time, to process a lot of data, a system with high speed for recognition is requested. Many ways which are perfect in theory are not make it, because the recognition speed of them is too low. Therefore it is still a very important research direction to study simple and high-efficiency handwritten digit recognition.A handwritten digit recognition system with feature extraction, feature selection and BP neural network is probed. Input the data handled by feature extraction and selection into the BP NN which would be trained time after time. At last, a handwritten digit recognition system with high rate of correctness is obtained. To improve the correctness rate and recognition speed, the following is done:1. Based on the Kirsch feature and the Fourier feature, a relationship map between the sample number and the decreased feature dimension is obtained. It shows that there is a maximum for the decreased feature dimension for every feature. It is enough to present every sample and ensure that the sample number used for K-L transform is minimal.2. The defect on analyzing the BP algorithm is pointed. And a method by using matrix analysis is put forward, which makes us clearly and globally see the ingredients that influence the modification of weights and makes for the programming and improvement on BP algorithm.3. An improved BP algorithm by modifying the error function is represented. The experiment shows that not only does it boost up the robustness of BP algorithm, but also it improves the correctness ratio on recognizing the sample.4. Implement a handwritten digit recognition system with feature extraction, feature selection and BP neural network. The experiment shows that the system with feature extraction and feature selection has much shorter training time and an equivalent correctness rate compared with the original system.