基于Caffe的卷积神经网络算法研究

改进MNIST手写数据集的LeNet-5网络模型,使用ReLU激活函数替代了原始函数,并将改进前后的模型训练结果作出了对比,说明了改进后的优势。


摘要: 深度学习(Deep Learning)是人工智能中的重要的一员,其在许多方面(例如语音识别、视频识别、图像分割等)取得了巨大的成果,在传统方法不容易解决的领域也获得了可观的成就。随着需求的不断提升,传统的编程方式已不能满足现今的需要,因此,人们开始寻求更为高效的深度学习开发模式。所以就出现了包含本文所研究的Caffe深度学习框架等许多框架。本文基于 Caffe 的卷积神经网络实行了以下研究:先介绍了关于深度学习的研究现状,并借由一些深度学习框架的介绍逐渐导出 Caffe 深度学习框架。其次,从人工神经网络导出卷积神经网络,对其结构和原理进行说明,叙述Caffe 深度学习框架的特性,搭建环境,再进行仿真实验。实验包含以下部分:对于CIFAR-10神经网络数据集进行仿真训练;对MNIST字符集进行训练,并改进LeNet-5网络。

关键词: Caffe;深度学习;卷积神经网络;机器学习;人工智能

The Study of Convolution Neural Network Based on Caffe

Abstract:  Deep learning is an important part of artificial intelligence. It has made great achievements in many fields (such as speech recognition, video recognition, image segmentation, etc.). Considerable results have also been achieved in areas where traditional methods are not easy to solve. With the increasing demand, the traditional programming method can not meet the needs of today. Therefore, people begin to seek a more efficient development model of deep learning. So there are many kinds of development frameworks including the Caffe depth learning framework studied in this paper. In this paper, convolution neural networks based on Caffe are studied as follows: First introduces the current situation of the research on deep learning, and leads to the Caffe deep learning framework by introducing some deep learning frameworks. Secondly, the convolutional neural network is derived from the artificial neural network, and its structure and composition are described. The characteristics of the Caffe depth learning framework are introduced, the environment is built, and the simulation experiment is carried out. It consists of the following parts: The training and simulation of CIFAR-10 neural network data set; The training of MNIST character set and the improvement LeNet-5 network.

Keywords: Caffe; Deep Learning; Convolution Neural Network; Machine Learning; AI

目录

摘要 i

Abstract i

目录 ii

1 绪论 1

1.1 引言 1

1.2 国内外研究现状 3

1.2.1 深度学习框架 3

1.2.2 图像识别 5

1.2.3 语音识别 5

1.2.4 自然语言处理 5

1.3 结构安排 5

1.3.1 研究目标 5

1.3.2 结构安排 5

2 人工神经网络和卷积神经网络概述 6

2.1 人工神经网络 6

2.1.1 概述 6

2.1.2 单层感知器 6

2.1.3 多层感知器 6

2.1.4 前馈反向传播算法 7

2.2 卷积神经网络 8

2.2.1 概述 8

2.2.2 结构 8

3 Caffe深度学习框架的原理及搭建步骤 12

3.1 概述 12

3.1.1 基本介绍 12

3.1.2 特性 12

3.2 Caffe深度学习框架环境搭建 13

3.2.1 所需条件 13

3.2.2 安装步骤 13

4 基于Caffe深度学习框架的仿真 18

4.1 卷积神经网络模型训练 18

4.1.1 样例CIFAR-10数据集的模型训练 18

4.1.2 训练自己建立的网络模型 20

4.2 MNIST手写字符数据集的LeNet-5卷积神经网络的改进 25

4.2.1 MNIST手写字符数据集 25

4.2.2 LeNet-5网络原理及改进方向 26

4.2.3 仿真结果对比 26

5 总结与展望 28

5.1 总结 28

5.2 展望 28

致谢 29

参考文献 30

绪论

深度学习[1]是机器学习中的一个重要组成部分。其基于人工神经网络[2](Artificial Neural Network),其动机在于模拟人脑神经网络进行分析学习,以仿人脑的机制在不同程度上实现信息处理、记忆、检索和存储等功能。