手写体数字识别技术研究

以下是资料介绍,如需要完整的请充值下载. 本资料已审核过,确保内容和网页里介绍一致.  
无需注册登录,支付后按照提示操作即可获取该资料.
资料介绍:

手写体数字识别技术研究(任务书,开题报告,论文12000字)
摘  要
手写体数字识别被广泛应用于金融,税务,普查,邮政等方面,所以对准确率的要求比较高。本文设计的神经网络是以图像的灰度作为输入。
神经网络式是近年的研究热点,因为本身具有高冗余度,非线性等特点,已经变得越来越受到人们的关注。选择卷积神经网络来识别数字图像。卷积神经网络网络能够用于识别移位、缩放和其他的扭曲不变的二维或者三维图像。特征提取参数用于训练和学习可以避免人工提取。本文的卷积神经网络总共分为5层,包含一个输入层,一个输出层和三个隐藏层。
本文采取的算法是BP(BackwardPropagation),在BP的基础上使用权值更新学习效率等超参数,在保证梯度下降速度的同时减少了震荡的概率。在激励函数的选择上,tanh函数均值为0的特性比sigmoid更加适合作为激励函数,这提高了训练的效率。
在正式识别图像之前,图像经过预处理以消除图像中的噪点,并将其转换为更清晰和易于绘制的点线图。包括对图像进行像素亮度、几何、滤波、复原等等操作。然后将字符分解按顺序通过神经网络。
研究结果表明: 本文采取的方案对MNIST数据集的识别具有较高的准确率,从整体来说对数字图片取得了比较好的识别效果。
关键词:手写体数字;卷积神经网络;方向传播;正向传播;反向传播

Abstract
Handwritten digit recognition is widely used in finance, taxation, census, postal, etc., so the requirement for accuracy is relatively high. The neural network designed in this paper takes the gray level of the image as input.
Neural network is a research hotspot in recent years. Because of its high redundancy and nonlinearity, it has become more and more popular. Convolutional neural networks are selected to identify digital images. Convolutional neural network networks can be used to identify shift, scale, and exotic distortion-invariant 2D or 3D images. Feature extraction parameters are used for training and learning to avoid manual extraction. The convolutional neural network in this paper is divided into five layers, including one input layer, three hidden layers, and one output layer.
The algorithm adopted in this paper is BP (Backward Propagation). The use of weights to update learning efficiency and other hyperparameters on the basis of BP reduces the probability of oscillation while ensuring the speed of gradient descent. In the selection of the excitation function, the property that the tanh function has an average value of 0 is more suitable as an excitation function than the sigmoid, which improves the training efficiency.
Before the image is officially recognized, the image is pre-processed to eliminate noise in the image and convert it to a sharper and easier-to-draw dotted line graph. Including the pixel brightness, geometry, filtering, restoration and other operations on the image. The character decomposition is then passed through the neural network in order.
The research results show that: The scheme adopted in this paper has a high accuracy in the identification of MNIST datasets. Overall, digital images have achieved better recognition results.
Key Words:Handwritten digits; Convolutional neural networks; Directional propagation; Forward propagation; Directional propagation.

目录
第1章 绪论    - 1 -
1.1课题研究背景    - 1 -
1.2国内外研究现状    - 2 -
第2章 神经网络    - 4 -
2.1神经元与神经网络    - 4 -
2.1.1神经元    - 4 -
2.1.2激活函数    - 5 -
2.1.3神经网络    - 6 -
2.2神经网络学习方法    - 7 -
2.2.1损失函数    - 7 -
2.2.2前向传播与反向传播    - 8 -
2.2.3过拟合和梯度消失    - 9 -
第3章 优化算法与模型训练    - 11 -
3.1 优化    - 11 -
3.1.1 dropout正则化    - 11 -
3.1.2 momentum梯度下降    - 11 -
3.1.3 RMSprop算法    - 12 -
3.1.4 Adam优化算法    - 12 -
3.2 学习率衰减    - 12 -
3.3 卷积神经网络    - 13 -
3.3.1 单层卷积网络    - 13 -
3.1.2 多层神经网络    - 14 -
3.2基于全连接层的手写数字识别模型的训练    - 15 -
3.2.1 模型的搭建    - 15 -
3.2.2 模型的调参与评估    - 16 -
3.2.3实验结果总结    - 16 -
3.3 基于CNN的手写数字识别模型的训练    - 17 -
3.3.1 模型的搭建    - 17 -
3.3.2 模型的调参与评估    - 20 -
3.3.3 实验结果总结    - 23 -
3.4 检测系统的整体搭建    - 24 -
3.4.1 训练好的模型的选择    - 24 -
3.4.2 搭建输入图片的预处理系统    - 25 -
3.4.3 搭建系统的Inference网络    - 25 -
3.4.4 识别结果    - 26 -
3.4.5 样本数量对CNN的影响    - 26 -
第4章 总结    - 27 -
参考文献    - 28 -
致谢    - 29