《厦门大学学报（自然科学版）》

图像分类问题包含两个重要的部分:特征提取器和分类器.多年来研究人员一直将精力投入到特征表示中,对于分类器却仅进行局部调参.基于一个性能优异的分类器与特征表示对图像分类系统同等重要的思想,提出了基于卷积特征的栈自编码器(stacked autoencoder on convolutional feature maps,SACF)的分类系统,并在数据集 CUB-200和 VGG-flower上进行了实验,对比了SACF与基于卷积特征和多层感知机的卷积神经网络(CNN)分类系统的分类效果,实验结果表明SACF具有更优的分类效果.

In problems of image classification,there are two important components:a feature extractor and a classifier.Researchers have focused on the former for decades,but only for local parameter adjustmentfor classifiers.According to the idea that a good classifier is just as important as the feature representation for the image classification,a classification system based on stacked autoencoder on convolutional feature maps(SACF)is proposed.Experiments are performed based on the CUB-200 and VGG-flower datasets and the classification results of the SACF classification system and the CNN classification system which based on convolutional feature maps and multilayer perceptron are compared.Results show that the accuracy of classification of SACF is superior to that of CNN.

引言
1 相关工作
2 基于卷积特征的SAE
3 实验
4 结论

图1 自编码器的网络结构<br/>Fig.1 The network structure of AE

图1 自编码器的网络结构
Fig.1 The network structure of AE

图2 SAE的网络结构<br/>Fig.2 The network structure of SAE

图2 SAE的网络结构
Fig.2 The network structure of SAE

图3 神经网络学习到的边缘特征<br/>Fig.3 Edge features learned by the neural network

图3 神经网络学习到的边缘特征
Fig.3 Edge features learned by the neural network

图4 SACF的结构图<br/>Fig.4 The structure diagram of SACF

图4 SACF的结构图
Fig.4 The structure diagram of SACF

表1 在数据集CUB-200上的分类准确率<br/>Tab.1 The classification accuracy obtained on the CUB-200 dataset

表1 在数据集CUB-200上的分类准确率
Tab.1 The classification accuracy obtained on the CUB-200 dataset

表2 在数据集VGG-flower上的分类准确率<br/>Tab.2 The classification accuracy obtained on the VGG-flower dataset

表2 在数据集VGG-flower上的分类准确率
Tab.2 The classification accuracy obtained on the VGG-flower dataset

[1] OJALA T,PIETIKÄINEN M,MÄENPÄÄ T.Multiresolution gray-scale and rotation invariant texture classification with local binary patterns[J].IEEE Transactions on Pattern Analysis & Machine Intelligence,2002,24(7):971-987.
[2] LOWE D G.Distinctive image features from scale-invariant keypoints [J].International Journal of Computer Vision,2004,60(2):91-110.
[3] CALONDER M,LEPETIT V,OZUYSAL M,et al.BRIEF:Computing a local binary descriptor very fast [J].IEEE Transactions on Pattern Analysis and Machine Intelligence,2012,34(7):1281-1298.
[4] RUBLEE E,RABAUD V,KONOLIGE K,et al.ORB:An efficient alternative to SIFT or SURF[C]∥International Conference on Computer Vision.Barcelona:IEEE,2011:2564-2571.
[5] TRZCINSKI T,CHRISTOUDIAS M,FUA P,et al.Boosting binary keypoint descriptors[C]∥Computer Vision and Pattern Recognition.Portland:IEEE,2013:2874-2881.
[6] VAPNIK V.The nature of statistical learning theory [M].New York:Springer-Verlag,2000.
[7] KRIZHEVSKY A,SUTSKEVER I,HINTON G E.Imagenet classification with deep convolutional neural networks[C]∥International Conference on Neural Information Processing Systems.Nevada:NIPS,2012:1097-1105.
[8] ZEILER M D,FERGUS R.Visualizing and understanding convolutional networks[C]∥European Conference on Computer Vision.Cham:Spinger,2014:818-833.
[9] 殷瑞,苏松志,李绍滋.一种卷积神经网络的图像矩正则化策略[J].智能系统学报,2016,11(1):43-48.
[10] WELINDER P,BRANSON S,MITA T,et al.Caltech-UCSD birds 200,CNS-TR-2010-001 [R].California:California Institute of Technology,2010.
[11] NILSBACK M E,ZISSERMAN A.Automated flower classification over a large number of classes[C]∥Conference on Computer Vision,Graphics & Image Processing.Indian:IEEE,2008:722-729.
[12] NAIR V,HINTON G E.Rectified linear units improve restricted boltzmann machines[C]∥International Conference on Machine Learning.Israel:ICML,2010:807-814.
[13] BASTIEN F,LAMBLIN P,PASCANU R,et al.Theano:new features and speed improvements [EB/OL].[2016-01-15].https:∥arxiv.org/pdf/1211.5590v1.pdf.
[14] BAHRAMPOUR S,RAMAKRISHNAN N,SCHOTT L,et al.Comparative study of caffe,neon,theano,and torch for deep learning [EB/OL].[2016-01-15].https:∥arxiv.org/abs/1511.06435v1.
[15] BOTTOU L.Stochastic gradient descent tricks [J].Lecture Notes in Computer Science,2012,7700:421-436.
[16] GIRSHICK R,DONAHUE J,DARRELL T,et al.Rich feature hierarchies for accurate object detection and semantic segmentation[C]∥Computer Vision and Pattern Recognition.Columbus:IEEE,2014:580-587.

备注

引言

1 相关工作

2 基于卷积特征的SAE

3 实验

4 结论

学报简介

备注

引言

1 相关工作

2 基于卷积特征的SAE

3 实 验

4 结 论

学报简介

3 实验

4 结论