目录
摘要Ⅲ
关键词Ⅲ
AbstractⅣ
引言
引言1
一、研究背景及意义1
（一）研究背景1
1.深度学习图像分类技术发展背景1
2.学术图片研究背景2
（二）研究意义 3
二、深度学习图像分类模型介绍3
（一）CNN3
（二）VGGNet4
（三）GoogLeNet5
（四）ResNet6
三、实验过程及结果分析 7
（一）数据来源及预处理7
（二）模型选择及参数设置8
（三）分类规则及评价指标9
（四）实验流程及结果9
四、结论与总结 12
致谢13
参考文献13
图21 VGG16结构示意图5
图22 最原始的Inception结构示意图5
图23 GoogLeNet结构信息图6
图24 ResNet残差模块结构图7
图31 batch_size为4时实验结果最大值9
图32 batch_size为8时实验结果最大值10
图33 batch_size为32时实验结果最大值10
图34 使用Adam优化器时实验结果最大值11
图35 模型可视化结果11
图36 准确率及损失值折线图12
表41 参数效果对比12
面向学术图片的深度学习分类模型构建研究
摘要
自科学诞生以来，科学可视化一直是一种直观而简洁的手段，被广泛的应用于各种类别的学术研究中。学术图片是科学可视化传播的重要媒介之一。随着科技的不断发展，大量的科研成果以学术论文的形式体现出来。在这些学术论文中，涌现出了大量的学术图片。例如，学术论文中的概念图，数据展示图等。由于深度学习具有适合于针对图片等复杂信息学习的特点，其可以使机器拥有和人类相似甚至超越人类的图像分类能力，所以深度学习模型同样可以适用于学术图片的分类。因此，本课题主要集中研究基于深度学习模型的学术 *51今日免费论文网|www.51jrft.com +Q: ^351916072* 
图片分类问题。本文从中国知网图片库获取了1600张学术图片，将图片按照训练集和测试集分别按照总数据比例的80%和20%进行分割。另外，本文利用TensorFlow和Keras搭建卷积神经网络模型，采用两种不同的优化器对模型进行优化，并对图像利用ImageDataGenerator和Data Argumentation进行了图像增强，通过不断的参数调整，最终得到的模型分类准确率最高达到95%以上。由此，本文验证了卷积神经网络对于学术图片的自动分类的可行性。
RESEARCH ON THE CONSTRUCTION OF DEEP LEARNING CLASSIFICATION MODEL FOR ACADEMIC PICTURES
ABSTRACT
Since the science emerged into our world, scientific visualization has been an intuitive and concise method, which was widely used in various types of academic research. Academic picture is one of the important media of visual communication of Science. With the continuous development of science and technology, plenty of scientific research achievements are embodied in the form of academic papers. In these academic papers, a great number of academic pictures have emerged. For example, concept map, data plot, etc. in the academic paper. Because deep learning is suitable for learning complex information such as pictures, which means that deep learning can make the machine have the ability of image classification similar to or even beyond human beings, so deep learning models can also be applied to the classification of academic pictures. Therefore, this paper focuses on the problem of academic image classification based on deep learning model. In this paper, 1600 academic pictures were obtained from CNKI image database, and those pictures were divided into 80% and 20% of the total data according to the training set and test set respectively. What’s more, this paper used TensorFlow and Keras to build the convolutional neural networks model, optimized the model via two different optimizers, and ImageDataGenerator and Data Argumentation were activated to enhance the images, and through constant parameter adjustment, in the end, the accuracy of classification of the final model is more than 95%. Therefore, this paper verifies the feasibility of convolutional neural networks for automatic classification of academic pictures.

原文链接：http://www.jxszl.com/jsj/xxaq/607034.html