基于CNN的绘画图片分类实现

论文总字数：27700字

摘 要

图像分类是指提取各个目标在图像信息中所展示的不同特征，并进行类别区分的图像处理方法。大数据时代，内容各异的图像频繁地产生并应用于各个行业，能够高效的按图片内容完成对图像的检索与管理是各个行业的迫切需要。研究具有高准确率的图像分类算法有着极其重要的理论意义与实际意义。

论文以深度残差网络为架构，实现了一个适用于手绘图片分类的卷积神经网络模型ResNet101。本文首先介绍了图像分类技术的发展历史和国内外研究现状，其次介绍了深度学习原理、卷积神经网络原理以及网络的构成部件；在模型实现部分，论文综合考虑了模型准确度、训练成本等因素，选择使用深度学习框架keras实现ResNet101，并针ResNet101存在的问题对网络模型进行了优化，同时还研究了提升模型准确率、加快的模型收敛速度的方法。

本论文设计的模型在谷歌开源数据集quick-draw进行了实验，实验结果为：（1）ResNet101在该数据集上准确率达到92.8%。（2）优化后的ResNet101模型准确率达到93.6%，训练时间减少约25%。模型初步满足了训练时间较短、分类准确率高的要求。

关键词：深度学习，卷积神经网络，深度残差网络，随机深度，准确率，训练时间

Abstract

Image classification refers to an image processing method that extracts different features displayed by each target in image information and classifies them. In the era of big data, images with different contents are frequently generated and applied to various industries. It is an urgent need of various industries to efficiently search and manage images according to the content of pictures. It is of great theoretical and practical significance to study image classification algorithms with high accuracy.

Based on the deep residual network, the paper implements a convolutional neural network model ResNet101 suitable for hand-drawn image classification. This paper first introduces the development history of image classification technology and the research status at home and abroad. Secondly, it introduces the principle of deep learning, the principle of convolutional neural network and the components of the network. In the model implementation part, the paper comprehensively considers the model accuracy, training cost, etc. Factors, choose to use the deep learning framework keras to achieve ResNet101, and optimize the network model by the problem of ResNet101, and also study the method of improving the accuracy of the model and speeding up the convergence of the model.

The model designed in this thesis was tested in the Google open source data set quick-draw. The experimental results are as follows: (1) ResNet101 has an accuracy rate of 92.8% on the data set. (2) The optimized ResNet101 model has an accuracy rate of 93.6% and a training time reduction of about 25%. The model initially meets the requirements of short training time and high classification accuracy.

KEY WORDS: deep learning, convolutional neural network, deep residual network, random depth, accuracy, training time

目 录

摘要Ⅱ

AbstractⅢ

第一章 绪论1

1.1 研究背景及意义1

1.2 图像分类技术的发展2

1.2.1 传统的图像分类算法2

1.2.2 基于CNN的图像分类算法2

1.3 研究内容与要求8

第二章 深度学习与卷积神经网络9

2.1 深度学习9

2.1.1 损失函数10

2.1.2 前馈运算与反向传播11

2.1.3 梯度下降算法13

2.2 卷积神经网络CNN15

2.2.1卷积层15

2.2.2池化层16

2.2.3全连接层17

第三章 手绘图片分类模型的实现及优化18

3.1 CNN模型选择18

3.1.1 深度网络的退化问题18

3.1.2方案论证18

3.2 ResNet101模型的实现20

3.2.1 ResNet101网络搭建20

3.2.2 训练数据预处理21

3.2.3 训练数据增强23

3.2.4 模型的训练23

3.3 ResNet101模型的优化24

3.3.1 ResNet101的缺点24

3.3.2 模型优化实现25

第四章 测试与结果分析28

4.1测试环境28

4.2测试结果28

4.2.1 ResNet101测试结果28

4.2.2 改进模型测试结果29

4.3结果分析30

第五章 总结与展望31

5.1 总结31

5.1.1 提升模型准确率31

5.1.2 加快训练速度31

5.2 展望31

参考文献(References)33

致谢34

剩余内容已隐藏，请支付后下载全文，论文总字数：27700字