基于优先度排序神经网络的手写数字识别

发布时间：2018-05-10 06:41

本文选题：手写数字识别 + 特征提取　；参考：《青岛大学》2017年硕士论文

【摘要】：手写数字识别是模式识别和图像处理领域一个重要的研究方向,虽然数字的种类只有十种,笔画也特别简单,但由于书写习惯的不同,字形的随意性很大,使得手写数字识别一直达不到理想的结果。手写数字识别系统设计包含识别器(分类器)设计和使用识别器(分类器)方法。王守觉院士提出的优先度排序神经网络是目前有可控拒识功能且识别率较高的识别器。本文借鉴王守觉院士提出的优先度排序神经网络思想,着重设计了一种基于优先度排序神经网络的手写数字识别器。在该识别器的设计过程中,一方面通过对样本加不同噪声、做不同的6种线性变换、7种Lie导数变换把MNIST数据库扩充为600000个训练样本、100000个测试样本的样本库。丰富了MNIST数据库,为设计高识别率手写数字分类器奠定了基础。另一方面,细化和优化了王守觉院士提出的优先度排序神经网络设计算法,改进了王守觉院士提出的优先度排序神经网络。用扩充MNIST数据库对多种识别器进行了实验。实验表明,优化的基于优先度排序神经网络的手写数字识别器的识别率明显高于传统的BP网络的识别率,也高于王守觉院士提出的优先度排序神经网络的识别率。
[Abstract]:Handwritten numeral recognition is an important research direction in the field of pattern recognition and image processing. Although there are only ten kinds of numbers and strokes are especially simple, due to the different writing habits, the character is very random. So that handwritten numeral recognition has been unable to achieve the desired results. The design of handwritten numeral recognition system includes the design of recognizer (classifier) and the use of recognizer (classifier) method. Academician Wang Shoujue's priority ranking neural network is a recognizer with controllable rejection function and high recognition rate. Based on the thought of priority ranking neural network proposed by academician Wang Shoujue, a handwritten numeral recognizer based on priority ranking neural network is designed in this paper. In the design of the recognizer, on the one hand, by adding different noises to the samples, 6 different linear transformations and 7 kinds of Lie derivative transformations are made to expand the MNIST database into a sample base of 600000 training samples and 100, 000 test samples. It enriches the MNIST database and lays a foundation for the design of handwritten number classifier with high recognition rate. On the other hand, it refines and optimizes the priority ranking neural network design algorithm proposed by academician Wang Shoujue, and improves the priority ranking neural network proposed by academician Wang Shoujue. Experiments on various recognizers are carried out with extended MNIST database. The experimental results show that the recognition rate of the optimized handwritten digital recognizer based on priority ranking neural network is obviously higher than that of the traditional BP neural network and the recognition rate of the priority ranking neural network proposed by academician Wang Shoujue.
【学位授予单位】：青岛大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【参考文献】