-
摘要：近年来，模式识别、深度学习成为热门研究课题之一。卷积神经网络作为深度学习的一种模型，在手写字符识别上的运用已经比较成熟。本系统结合了图像处理和神经网络的知识，以26个大写字母作为识别体，对采集的英文手写体文本图像进行识别。首先通过文本定位，字符分割、字符大小归一等操作，得到单张英文字符图像。然后利用卷积神经网络的理论基础，设计搭建并训练了一个具有两层卷积层、一个全连接层的神经网络，网络在测试集上的识别率达到90.04%。系统对125个样本进行了测试，其中利用软件生成的手写图像样本识别率达到87.65%，利用画图软件写的图像样本识别率达到83.46%，利用手机拍摄的手写图像样本识别率达到77.32%。34935
- 上一篇：基于SIP协议即时通信系统设计+源代码
- 下一篇：Android云笔记APP设计与实现+源代码
毕业论文关键字：CNN; 图像处理; 手写英文; 字符识别
The Design and Implementation of English Character Recognition System Based on CNN
Abstract: In recent years,the pattern recognition and the deep learning have become one of the hot research topics.Convolutional Neural Net,as a model of deep learning,has been relatively mature on the application of handwriting character recognition.This system combining the knowledge of image processing with neural network,identifies the English handwriting text image acquired,with 26 capital letters as identifications.Firstly, through text location, character segmentation and size to a first-class operation,I acquired single character image.Then making use of the theoretical basis of CNN, I designed and trained a neural network with two convolutional layers and a fully connected layer.Recognition rate of the neural network reached 90.04% on the test set.Experiments for 125 samples were tested, which recognition rate of handwriting image samples generated by software the reached 87.65%, recognition rate of handwriting image samples written in drawing software reached 83.46%, recognition rate of handwriting image samples taken with my cell phone reached 77.32%.
Keywords: CNN; image processing; handwriting English; character recognition
目录
摘要 1
关键字 1
Abstract 1
Keywords 1
1 绪论 1
1.1 研究背景及意义 1
1.2 国内外研究情况 2
1.2.1 国内研究状况 2
1.2.2 国外研究状况 2
2 开发工具与系统介绍 2
3 图像灰度化 3
4 图像分割 4
4.1 文本定位 4
4.2 图像二值化 5
4.2 字符分割方法 5
4.3 字符大小归一 5
4.3.1 滤波处理 6
4.3.2 字符大小归一 6
5 神经网络 7
5.1 神经网络简介 7
5.2 常见神经网络的分类 8
5.2.1 感知器 8
5.2.2 BP神经网络 9
5.2.3 径向基神经网络 9
6 CNN 9
6.1 CNN的理论基础 9
6.2 CNN的结构 9
6.2.1 典型结构LeNet5 9
6.2.2 系统CNN结构的构建 10
6.3 CNN的训练 11
6.3.1 样本收集 11
6.3.2 激活函数 11
6.3.3 损失函数 13
6.3.4 梯度下降学习最优参数 14
6.3.5 权值更新 17
6.4 CNN的实验结果分析 18
7 系统实现与结果分析 19
7.1 系统实现 19
7.2 结果分析 21