基于BP神经网络的手写数字识别研究

摘要

手写数字识别目前已经广泛应用于各种空间领域。现在许多手写数字信息都需要自动输入工业生产，商业以及经济活动信息，比如纸质分数，纸币，财务报表，统计报告以及邮政编码等，所有的这些信息都需要通过手写数字识别技术来进行相关的处理，但是如果这些都可以通过使用手写数字识别技术来完成，那么可以显著的提高生产能力和效率。

本文介绍了神经网络、图像的处理和识别，为了研究手写体数字的识别，将系统分为三部分：预处理，特征提取和分类器。用于图像预处理的算法主要是一些基本的图像处理算法：将256色位图转换为灰度、灰度二值化等等。其次就是对特征的提取，最后进行学习训练并通过识别预测。用matlab软件的rands函数来实现网络权值的初始化，网络结构为输入层35，潜藏层25，输出层10，学习速率为0.1，同时潜藏层激励函数为Sigmoid函数形式。随机选取4500个图像进行提取，根据公式计算潜藏层和输出层输出误差，更新网络权值。

训练好神经网络之后，用随机选取的500个数字字符预测网络，输入特征向量，计算潜藏层和输出层输出，得到最终的预测的数据，并计算出每个数的正确率和总体精度。

该论文有图28幅，表格2个，参考文献47篇。

关键词：BP算法神经网络预处理手写数字数字识别

Research on Handwritten Digital Recognition Based on BP Neural Network

Abstract

Handwritten digit recognition has been widely used in various space fields. Now many handwritten digital information needs to be automatically input industrial production, commercial and economic activity information, such as paper grades, notes, and financial statements, statistical reports, and post code, etc., all the information needed by handwritten numeral recognition technology for the processing of related, but if these can be through the use of handwritten numeral recognition technology to complete, so can significantly improve the productivity and efficiency.

This paper introduces the processing and recognition of neural network and image. In order to study the recognition of handwritten numeral, the system is pided into three parts: pretreatment, feature extraction and classifier. The algorithms used for image preprocessing are mainly some basic image processing algorithms: the 256 color bitmap is converted to grayscale, grayscale binarization and so on. Then it is the extraction of the features, and finally the learning training and the recognition prediction. Using the rands function of matlab software, the initialization of network weight value is realized. The network structure is the input layer 35, the hidden layer 25, the output layer 10, the learning rate is 0.1, and the underlying excitation function is the form of Sigmoid function. At random, 4500 images were selected to extract, and the output error of hidden layer and output layer was calculated according to the formula, and the network weight value was updated.

Good training the neural network, with 500 randomly selected digital character to predict network, the input feature vector, the calculation of hidden layer and output layer output, obtains the final prediction data, and calculate the number of each accuracy and overall accuracy.

The paper has 28 pictures, 2 tables and 47 references.

Key Words: BP algorithm neural network preprocessing pretreatment handwritten numeral numeral recognition

摘要 I

Abstract II

目录 III

1 绪论 1

1.1 手写数字识别研究目的及意义基于BP神经网络的手写数字识别研究:http://www.chuibin.com/jisuanji/lunwen_206214.html