基于极限学习机的手写体字符识别方法研究

发布时间：2018-07-25 15:07

【摘要】：随着科技的发展,人们生活和工作产生大量的手写体字符信息,考虑到这些字符所要表达信息的安全性和隐私性,让机器实现快速、准确的手写体字符自动识别技术势在必行。手写体字符识别方法主要是光学字符识别,但因其识别率低、成本高等问题,还未能广泛推广使用。目前包括模板匹配、神经网络和支持向量机等模式识别的方法已经投入到手写体字符的识别研究。本文针对传统的字符识别方法实时性差、成本高等问题,提出采用极限学习机算法实现手写体字符识别。论文首先对模式识别的定义、基本组成系统和基本方法进行了介绍和讨论,引出了利用神经网络进行模式识别的方法,对神经网络的工作原理和特点进行了分析和研究;然后提出用极限学习机实现手写体字符识别方法,针对原始极限学习机存在的结构风险和经验风险不平衡这一问题,提出使用正则极限学习机和傅里叶变换优化极限学习机实现手写体字符识别。设计基于BP神经网络、极限学习机、正则极限学习机和傅里叶变换优化极限学习机四种算法实现手写体字符识别的方法,包括预处理、特征选择和降维等具体过程。手写体字符识别算法仿真的训练样本为MINIST样本库的10000个数字样本,测试样本数量为1000个,除采用BP神经网络、极限学习机、正则极限学习机和傅里叶变换优化极限学习机四种算法实现手写体字符的识别结果外,还设计实验分析隐含层神经元个数对仿真结果的影响。通过对算法仿真结果的对比分析,BP网络作为最经典的神经网络算法,在手写体数字识别结果的精度上达到了较高的水准。极限学习机算法较BP神经网络在训练时间上表现出极大的优势,但是识别精度低于BP神经网络。基于极限学习机的两种优化算法,即正则极限学习机和傅里叶变换优化极限学习机,与原始极限学习机相比提高了算法的泛化能力,提高了手写体数字字符的识别精度。
[Abstract]:With the development of science and technology, people produce a large amount of handwritten character information in life and work. Considering the security and privacy of the information expressed by these characters, it is imperative for the machine to realize rapid and accurate automatic recognition of handwritten characters. The main method of handwritten character recognition is optical character recognition, but because of its low recognition rate and high cost, it has not been widely used. At present, pattern recognition methods, such as template matching, neural network and support vector machine (SVM), have been put into the research of handwritten character recognition. Aiming at the problems of poor real time and high cost of traditional character recognition methods, this paper proposes to use extreme learning machine algorithm to realize handwritten character recognition. Firstly, the definition, basic composition system and basic method of pattern recognition are introduced and discussed, the method of pattern recognition using neural network is introduced, and the working principle and characteristics of neural network are analyzed and studied. Then, a method of handwritten character recognition based on extreme learning machine is put forward, aiming at the imbalance between structural risk and empirical risk of original extreme learning machine. This paper presents a new method to realize handwritten character recognition by using regular limit learning machine and Fourier transform optimization learning machine. Based on BP neural network, extreme learning machine, regular ultimate learning machine and Fourier transform optimization extreme learning machine, this paper designs four algorithms to realize handwritten character recognition, including preprocessing, feature selection and dimensionality reduction. The training sample of handwritten character recognition algorithm simulation is 10, 000 digital samples of MINIST sample database, and the number of test samples is 1000. In addition to the recognition results of handwritten characters, four algorithms of regular limit learning machine and Fourier transform optimized limit learning machine are designed to analyze the effect of the number of hidden layer neurons on the simulation results. Through the comparison and analysis of the simulation results of the algorithm, the BP neural network, as the most classical neural network algorithm, has reached a high level in the accuracy of handwritten digital recognition results. Compared with BP neural network, the algorithm of extreme learning machine shows great superiority in training time, but the recognition accuracy is lower than that of BP neural network. Two optimization algorithms based on LLM, namely regular LLM and Fourier transform LLM, improve the generalization ability of the algorithm and the recognition accuracy of handwritten numeric characters compared with the original LLM.
【学位授予单位】：杭州电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.4;TP18

【参考文献】