Handwritten Number Recognition Method Based On Wavelet And BP Neural Network

In recent years, as the great development of electronic communication equipment and computer terminals, Handwriting-identifying technology begin to appear, and is gradually improved. The Handwriting figures identifying system has been extensively used in Zip code identifying, bank business and so on. Handwriting-identifying means that getting writing information of writer by track getting equipment, then inputting the track to computer. The strokes which people write on writing pad or touch screen are stored in computer in the form of vector graph.The basic principle of Handwriting figures indentifying is that compare the input image to a lot of standard pattern prepared before for deciding which standard pattern match the input image, then consider it to the indentifying result and output it. Different figure has different character. Through research and analyze, people find that these characters could be described by some given measure. These given measure are named character measure. Figure pattern is abstracted to a character value after described by character measure, so kinds of different image pattern could be compared in one character value space. In general, there is learning ability in handwriting indentifying system. This learning ability makes the generation process of standard pattern remembered to every indentifying process, so the rate of identifying increases gradually. In other words, assuming the result of a certain identifying is right, then adds this pattern into standard pattern; if the result is wrong, then input the signal one again.The origin image needs to be denoised before do the handwriting-identifying. Traditional denoising method, for example linear denoising and non-linear denoising, sometimes results in the augment of the entropy value after transforming the image signal, so that the stability and relevance of the image signal can not be described accurately. To conquer the defects above, wavelet analyze is introduced into this field to solve this problem. This paper adopts Q-shift complex wavelet and HGM image denoising method. The principle of this method is the difference between property and statement of image signal and noise signal. Energy of noise distributes all the wavelet coefficients, so the amplitudes of the wavelet coefficients are usually small, but the numbers of them are large. The character of image signal is in direct contradiction to the noise signal. Depending on the characteristic of signal wavelet boundary, noise can be decreased. A threshold is set up to wavelet of the signal first, the wavelet coefficients which are bigger than the threshold are considered as containing both image and noise signal; the wavelet coefficients which are smaller than the threshold are considered as only containing noise, and could be totally abandoned. Noise can be got rid of by this method.Neural network is information processing model that Simulate biological neural system, it has a good autonomic parallel and adaptive study ability that is widely used in pattern recognition. There are lots of classical artificial neural network algorithms. This paper adopts BP algorithm improves it. BP algorithm is a learning algorithm with tutor that is fit for learning and practicing of multi-level neural network, and is based on gradient decreasing algorithms, the main idea is dividing the learning process into two stages: one stage is that the inputting signals are processed by implied level and compute actual output value of every point, the other stage is that recursively compute errors between actual output and expectation output in every level when there is no expectation result in output level, and adjust weight values. By the analysis to the basic principle of BP algorithm, we know that the classical BP algorithm's characteristics are low speed of convergence and the tendency of falling into local minimizing points, this paper adjusts the neural network's output level, implied level and input level connect weight value matrix, and also adjusts error function. This paper programs and does simulation experiment to the BP algorithm and improved one, then compares them. The result showed that the using of these methods is good at improving the shortcomings of convergence and the tendency of falling into local minimizing points. At the same time, compared with the using of improved BP algorithm to handwriting figures indentifying has a great improvement in accurate rate.