手写数字识别网络研究
2022-12-20 10:26:33
论文总字数:16366字
摘 要
信息处理的逐渐深入,大家对信息处理的要求持续变高,既要提高可信度,又要加快处理速度。对手写数字识别技术的研讨,在以后的信息发展历程中有着举重若轻的地位,手写字符识别是光学字符识别领域中有重要地位的一部分,在实际生活中也用处很大。另外,手写体字符识别推动了很多学术研究的进步,不仅是最简单的数字,还有复杂的汉字和多种多样的面貌识别。传统的方法对环境要求高,成本高,识别率低,实时性差[1]。
本文研究了手写数字图像识别算法,使用matlab执行程序,首先图像的预处理中包含模板线性去噪,阈值法进二值化等技术,处理出更易让电脑程序识别的图像是进行以下几个步骤的先题条件,接下来是对处理过的图像进行特征点的提取,分割处理过的数字图片,统计分割出来的几个模块的数字的重心、面积等等的数据,最后搭建卷积神经网络,不断地重复训练之后并进行测试其模型,最终检测结果的正确率是否符合预期要求。
本文的卷积神经网络的进行了50次的训练之后,基本保证95%的准确率,达到预期识别准确率85%的要求。
关键词:手写数字;图像识别;神经网络;预处理。
Research on handwritten number recognition network
Abstract
With the gradual deepening of information processing, people's requirements for information processing continue to become higher, which not only needs to improve the credibility, but also needs to speed up the processing speed. The research on the recognition technology of hand writing number, in the later information development process, has the status of light weight, Handwritten character recognition is an important part in the field of optical character recognition. In addition, handwritten character recognition has promoted a lot of academic research, not only the simplest Numbers, but also complex Chinese characters and a variety of face recognition. The traditional method is to use the photoelectric transformation principle of optics to identify Numbers. This method has high environmental requirements, high cost, low recognition rate and poor real-time performance.
This paper mainly introduces the handwritten digital image recognition network research, using matlab to perform, the first image preprocessing includes linear template denoising, threshold method into technology, such as binarization processing out makes it easier for computer programs to identify the image is the first problem conditions for the following steps, the next is carried out on the processed image feature points extraction and segmentation of digital images, statistical split out the focus of several modules of digital data, area and so on, and finally build convolution neural networks, after repeated training and test the model, the final test results accuracy meets the expected requirement.
After 50 times of training, the convolutional neural network in this paper can basically guarantee 95% accuracy and meet the requirement of 85% accuracy.
Key words: handwritten Numbers; Image recognition; Neural network; Pretreatment.
目 录
摘要 I
Abstract II
第一章 引 言 1
1.1 国内外现状 1
1.2 研究目的和意义 2
1.3 论文章节结构 3
2.1 图像平滑去躁 4
2.2 数字图像二值化 6
2.3 数字图像归一化 8
第三章 数字特征提取 10
3.1 图像特征 10
3.2 特征提取的一般原则 10
3.3特征提取的方法 11
第四章 神经网络的训练和数字识别 12
4.1神经网络的原理和训练 12
4.1.1 卷积神经网络的结构 12
4.1.2 神经网络的工作原理 13
4.1.3 神经网络的学习方式 13
4.1.4人工神经网络的特性 13
4.2神经网络的训练 14
第五章 总 结 17
5.1 文章总结 17
5.2 未来展望 17
致 谢 18
参考文献(Reference) 19
第一章 引 言
1.1 国内外现状
当代经济和科学技术水平的逐步进步与发展,走入信息化时代,人们接触到越来越多的文字信息越来越多。数字作为一种文字信息已经渗透进人们生活的更方各面,而阿拉伯数字是世界公用的,所以手写体数字的识别研究公认的由阿拉伯数字领导训练研究。跟随时代的科学技术的发展的趋势,方便人类的日常生活起居,让普通人民的生活出行的麻烦减少,实现对各类文字信息的自动检测、存储和识别等过程快速而准确的执行就显得尤为重要,大量研究者对其进行了深入的研究,从而推动了光学字符识别技术(OCR)的产生和进步[2]。光学字符识别技术首次登上舞台出现在人们的视野和话题中是在1928年。计算机的发明让隶属于模式识别的OCR在学术舞台上大放光彩。
剩余内容已隐藏,请支付后下载全文,论文总字数:16366字