论文总字数:29800字
摘 要
卷积神经网络是一种特殊的深层神经网络模型,作为一种深度学习的架构,已成为当前图像识别和语音分析领域的研究热点。而它的特殊性则体现在两个方面:其一,它的神经元间采用局部连接而非全链接;其二,卷积神经网络同层神经元采用权值共享。以上两点既降低了其网络模型的复杂度,又减少了权值的数量,同时使它的网络结构更类似于生物神经网络。人脸识别技术作为一项视觉领域课题有着重大的实际意义和巨大的商业价值,涉及人脸检测,图像处理,特征提取,人脸识别等领域。卷积神经网络应用于特征提取和人脸识别中,并有着出色表现。
本文将重点介绍卷积神经网络在特征提取及分类过程中的原理,以及如何通过MatConvNet(一种基于matlab的卷积神经网络框架)实现卷积神经网络,并简要介绍人脸图像数据集的预处理及其影响和意义。最后将会介绍一组实验借助MatConvNet实现基于卷积神经网络的人脸识别,并通过分析实验数据,就图像预处理、网络模型等方面对人脸识别的优化进行讨论。
关键词:卷积神经网络,深度学习,特征提取,图像分类,人脸识别,MatConvNet,LFW
Abstract
Convolutional Neural Network is a special deep neural network model, as a architecture of Deep Learning, has become a research hotspot in the field of image recognition and speech analysis. Its specificity can be reflected in two aspects:Firstly, it uses a local connection rather than a full connection between the neurons; Secondly, the neurons in the same layers of the CNN share and take the same weights. The two points above ,which making the CNN much more similar with the network structure of biological neural networks, reduce both the complexity of the network model and the amount of the weights. As a visual subject, face recognition technology involves the face detection, image processing, feature extraction, face recognition fields, with great practical significance and great commercial value.Convolution neural network applied to feature extraction and face recognition fields, has an excellent performance.
The paper will focus on the principle of the CNN how could it work on feature extraction or classification process works, and how theMatConvNet (a CNN framework based on the matlab)implement convolutional neural network. It will also take a brief pre-introduction on the processingof the face image dataset with its impact and significance. Finally, we will introduce a set of experiments by means of MatConvNetto realizetheface recognition based on the convolutional neural network, and talk about the optimization for face recognition on aspects of the image pre-processing and network model design trough the analysis of experimental data.
Keywords: convolution neural network, deep learning, features extraction, image classification, face recognition , MatConvNet , LFW
英文缩略词表
ANN | Artificial Neural Network |
CNN | Convolution Neural Network |
CPU | Central Processing Unit |
GPU | Graphics Processing Unit |
ReLU | Rectified Linear Units |
LFW | Labeled Faces in the Wild |
SVM | Support Vector Machine |
目录
摘要 II
Abstract III
英文缩略词表 IV
目录 V
表格目录 VIi
插图目录 VII
第一章 绪论 9
1.1 研究背景 9
1.2 国内外研究现状 9
1.3 本文研究内容及意义 10
第二章 卷积神经网络 11
2.1 卷积神经网络由来 11
2.2 卷积神经网络核心思想 12
2.2.1 局部感知 12
2.2.2 权值共享 12
2.2.3 多层卷积核 13
2.2.4 多层卷积层 13
2.2.5 池化操作 14
2.3 卷积神经网络层次及模块 15
2.3.1 卷积层(Convolution Layer) 15
2.3.2 池化层(PoolingLayer) 16
2.3.3 Softmax层(SoftmaxLayer) 16
2.3.4 归一化(Normalizemodule) 17
2.3.5 ReLU(ReLUmodule) 18
2.3.6 Dropout(Dropoutmodule) 18
2.4 本章总结 18
第三章 MatConvNet卷积神经网络框架 19
3.1 MatConvNet概述 19
3.2 MatConvNet学习探索历程 19
3.2.1 基于MatConvNet卷积神经网络模型 20
3.2.2 基于MatConvNet卷积神经网络的学习 22
3.3 MatConvNet特性分析 23
3.4 本章总结 23
第四章 实验:MatConvNet下基于CNN人脸识别 24
4.1 实验方案权衡 24
4.1.1 人脸数据集选择 24
4.1.2 图像预处理 25
4.1.3 网络模型选择 26
4.2 实验结果数据分析 27
4.2.1 网络学习方式对实验结果的影响 27
4.2.2 数据集预处理操作对实验结果的影响 29
4.3 本章总结 31
第五章 总结与展望 33
5.1 课题研究总结 33
5.2 研究展望 33
致谢 36
参考文献 37
表格目录
表 1:训练数据顺序打乱概率与人脸识别准确率关系表 27
剩余内容已隐藏,请支付后下载全文,论文总字数:29800字
该课题毕业论文、开题报告、外文翻译、程序设计、图纸设计等资料可联系客服协助查找;