用于图像分类的两级分层特征学习_自然图像的分层结构资源-CSDN文库

30 浏览量 2021-02-25 14:28:07 上传评论收藏 695KB PDF 举报

资源推荐

资源详情

资源评论

Song et al. / Front Inform Technol Electron Eng 2016 17(9):897-906 897

Frontiers of Information Technology & Electronic Engineering

www.zju.edu.cn/jzus; engineering.cae.cn; www.springerlink.com

ISSN 2095-9184 (print); ISSN 2095-9230 (online)

E-mail: jzus@zju.edu.cn

Two-level hierarchical feature learning for

image classiﬁcation

∗

Guang-hui SONG

1,2

, Xiao-gang JIN

†‡1

,Gen-langCHEN

,YanNIE

(

College of Computer Science and Technology, Zhejiang University, Hangzhou 310027, China)

(

Ningbo Institute of Technology, Zhejiang University, Ningb o 315100, China)

(

College of Science and Technology, Ningbo University, Ningb o 315100, China)

†

E-mail: xiaogangj@cise.zju.edu.cn

Received Oct. 20, 2015; Revision accepted Apr. 10, 2016; Crosschecked Aug. 8, 2016

Abstract: In some image classiﬁcation tasks, similarities among diﬀerent categories are diﬀerent and the samples

are usually misclassiﬁed as highly similar categories. To distinguish highly similar categories, more speciﬁc features

are required so that the classiﬁer can improve the classiﬁcation performance. In this paper, we propose a novel

two-level hierarchical feature learning framework based on the deep convolutional neural network (CNN), which is

simple and eﬀective. First, the deep feature extractors of diﬀerent levels are trained using the transfer learning

method that ﬁne-tunes the pre-trained deep CNN model toward the new target dataset. Second, the general feature

extracted from all the categories and the speciﬁc feature extracted from highly similar categories are fused into a

feature vector. Then the ﬁnal feature representation is fed into a linear classiﬁer. Finally, experiments using the

Caltech-256, Oxford Flower-102, and Tasmania Coral Point Count (CPC) datasets demonstrate that the expression

ability of the deep features resulting from two-level hierarchical feature learning is powerful. Our proposed method

eﬀectively increases the classiﬁcation accuracy in comparison with ﬂat multiple classiﬁcation methods.

Key words: Transfer learning, Feature learning, Deep convolutional neural network, Hierarchical classiﬁcation,

Spectral clustering

http://dx.doi.org/10.1631/FITEE.1500346 CLC number: TP391.4

1 Introduction

The deep convolutional neural network (CNN)

has achieved impressive classiﬁcation performance in

the ImageNet benchmark (Krizhevsky et al., 2012).

Surprisingly, transfer learning methods based on the

deep convolutional feature trained on a generic recog-

nition task are also successful in various computer

vision tasks, such as object classiﬁcation, domain

adaptation, and scene recognition. They achieve

results superior to those of the previous meth-

‡

Corresponding author

Project supported by the National Natural Science Foundation

of China (No. 61379074) and the Zhejiang Provincial Natural Sci-

ence Foundation of China (Nos. LZ12F02003 and LY15F020035)

OR CID: Xiao-gang JIN, http://orcid.org/0000-0002-7787-7228



Zhejiang University and Springer-Verlag Berlin Heidelberg 2016

ods (Donahue et al., 2014; Zeiler and Fergus, 2014;

Cai et al., 2015). Therefore, the feature learning

ability of deep CNN has received considerable at-

tention. In previous studies, deep CNN models were

used as feature extractors but not as classiﬁers, and

they provided a way to obtain more speciﬁc visual

features (Yosinski et al., 2014).

At present, most deep CNN models serve as

ﬂat end-to-end classiﬁers for image recognition tasks.

These deep models take the raw image as the network

input, extract image features using back-propagation

through layers of convolutional ﬁlters, and ﬁnally

output the categorized results using a softmax out-

put layer. However, the reality is that image datasets

have a growing sample size and image category. Simi-

larities are diﬀerent among diﬀerent categories, with

898 Song et al. / Front Inform Technol Electron Eng 2016 17(9):897-906

some categories being more diﬃcult to distinguish

than others. For example, let us classify three fruit

varieties, i.e., apples, oranges, and bananas. The

ﬁrst two varieties are clearly diﬃcult to distinguish,

whereas the bananas are easy to separate from the

other two. A similar situation exists in many image

datasets. Fig. 1 shows that the same border colors

and line types represent the highly similar categories

in the Oxford Flower-102 dataset.

Camellia

Mallow

Coltsfoot

Barberton daisy

Fig. 1 The same border colors and line types rep-

resenting the highly similar categories in the Oxford

Flower-102 dataset. Distinguishing between camellia

and mallow is diﬃcult. The situation is similar be-

tween coltsfoot and barberton daisy. References to

color refer to the online version of this ﬁgure

Inspired by the idea of transfer learning, we have

come up with two questions: ﬁrst, whether the gen-

eral feature extractor trained on all fruit varieties can

be adequate for distinguishing the categories with

high similarity, and second, how to extract more

speciﬁc features using the feature learning method

to improve the classiﬁcation performance. Adopt-

ing the hierarchical classiﬁcation and feature learn-

ing methods to solve the above problem is natural.

Hierarchical classiﬁcation methods usually organize

categories into a tree hierarchy (Deng et al., 2011;

Srivastava and Salakhutdinov, 2013). In this way

the parent node can select diﬀerent hand-designed

feature descriptors according to the concrete con-

ditions of child nodes for the classiﬁer. In trans-

fer learning with feature learning ability, the gen-

eral features are ﬁrst learned on the root node, and

the speciﬁc features are further obtained from the

parent nodes for the highly similar categories. Hier-

archical classiﬁcation methods have the advantages

of high eﬃciency and scalability, but they can only

achieve a trade-oﬀ between accuracy and eﬃciency

(Zhao et al., 2011; Liu et al., 2013). In particular,

too many levels produce error accumulation, which

seriously aﬀects the classiﬁcation accuracy. We in-

tend to ﬁnd a way to use the hierarchical feature

learning ability while preventing the accumulation of

errors, leading to a decrease in accuracy. However, so

far, only limited studies have focused on combining

deep feature learning with hierarchical classiﬁcation

to improve classiﬁcation accuracy (Yan et al., 2015).

In this paper we propose a novel two-level hier-

archical feature learning framework, which involves

three steps. First, the general feature extractor is

trained by ﬁne-tuning a pre-trained deep CNN model

on the target dataset. At the same time, the similar-

ity among diﬀerent categories is obtained by process-

ing the confusion matrix. Second, the highly similar

categories are divided into the same parent node,

and the speciﬁc feature extractor is trained sepa-

rately on each parent node. Finally, the general and

speciﬁc deep features are fused into a feature vector

used for the ﬁnal linear classiﬁer. The experiments

using the Caltech-256, Oxford Flower-102, and Tas-

mania Coral Point Count (CPC) datasets demon-

strate that transfer learning methods gradually en-

hance the expression ability of deep features. Our

proposed method can eﬀectively improve the classi-

ﬁcation accuracy based on the deep CNN models on

these three datasets.

2 Related work

2.1 Deep convolutional neural network and

feature learning

The deep CNN model has been widely used

because of its better classiﬁcation performance. It

is composed of ﬁve convolutional layers (conv1–

conv5), two fully connected layers (fc6 and fc7),

and a softmax output layer (Krizhevsky et al., 2012).

This network model has strong feature learning

ability and can generate feature representations by

learning low-level features in the early convolu-

tional layers and transforming them to high-level

semantic features in the latter convolutional layers

剩余9页未读，继续阅读

评论收藏

内容反馈

weixin_38614377

粉丝: 2
资源: 945

用于图像分类的两级分层特征学习

基于分层稀疏表示特征学习的高光谱图像分类研究

基于散列的分层特征表示的高光谱 图像分类方法

用于大型图像分类和新颖类别检测的大利润率树分类器的成本敏感分层学习

基于分层稀疏表示的多实例半监督学习及其在图像分类中的应用

基于多层卷积神经网络的SAR图像分类方法.pdf

digital image 图像黑白化 图像直方图 分层

组织图像肿瘤分类数据集

MNIST分类的两级RCN模型参考实现

基于matlab图像预处理的相关论文

一种用于高光谱图像分类的新型分层半监督支持向量机

数字图像处理知识点总结 (3).pdf

通过类别结构的两阶段分层学习推荐

图像处理分层块搜索方法HBMA得到运动矢量matlab代码

数字图像处理知识点总结.docx

一维卷积神经网络用于雷达高分辨率距离像识别.pdf

DBN.rar_dbn 分类_deep classifier_immediatelyeyg_分类器_深度信念

具有分层多特征学习的文本本地化

数字图像处理知识结构

图像平滑伪彩色变换阈值分割图像的层级表示.zip

数字图像处理学习笔记（十一）——用Python代码实现图像增强之线性变换、对数变换、幂律变换、分段线性变换、灰度级分层、直方图均衡化、平滑滤波器、锐化滤波器

2016-2017《数字图像处理》试卷A答案1

图像压缩的MATLAB程序

遥感数字图像处理：图象显示和拉伸.ppt

深度学习算法教程(Deeplearning Algorithms Tutorial) 完整版PDF

遥感图像中居民区语义分割的分层弱监督学习

基于超像素的Graph_Based图像分割算法.rar_Graph-Based 分割_graph-based_图像分割_基于超像

(完整版)数字图像处理知识点总结.pdf

最新资源

基于散列的分层特征表示的高光谱图像分类方法

digital image 图像黑白化图像直方图分层