基于深度哈希学习的商标图像检索研究

袁培森; 张勇; 李美玲; 顾兴健

doi:10.3969/j.issn.1000-5641.2018.05.015

基于深度哈希学习的商标图像检索研究

doi: 10.3969/j.issn.1000-5641.2018.05.015

1.
南京农业大学信息科学技术学院, 南京 210095
2.
南京工程学院基础部, 南京 211167

基金项目:

国家自然科学基金 61502236

中央高校基本科研业务费专项资金资助 KYZ201752

详细信息

作者简介:
袁培森, 男, 博士, 讲师, 主要从事数据挖掘、海量数据处理与分析研究.E-mail:peiseny@njau.edu.cn

中图分类号: TP391
计量
- 文章访问数: 249
- HTML全文浏览量: 105
- PDF下载量: 271
- 被引次数: 0
出版历程
- 收稿日期: 2018-06-27
- 刊出日期: 2018-09-25

Research on trademark image retrieval based on deep Hashing

1.
College of Information Science and Technology, Nanjing Agricultural University, Nanjing 210095, China
2.
Department of Mathematics and Physics, Nanjing Institute of Technology, Nanjing 211167, China

摘要

摘要: 大规模图像检索具有广泛的应用前景，其核心在于图像特征提取和高效相似性计算.深度学习技术在图像特征提取具有较强的特征表示能力，同时哈希技术在高维数据近似计算方面具有较好的性能.目前，基于哈希学习的技术在大规模图像检索及相似性查询方面获得了广泛的研究和应用.本文结合卷积神经网络和哈希技术实现商标图像检索，通过深度学习技术提取商标图像特征，使用位哈希对数据对象编码，在海明空间折中查询的质量和效率.基于卷积神经网络模型，提出了深度哈希算法，并研究了损失函数和该数据集上的优化器选择，通过获取符合哈希编码规范的位编码实现对在二元空间对商标图像数据快速检索，该方法分为离线深度哈希学习和在线查询两个阶段.在真实商标数据集上进行实验，实验结果表明，本文方法能够在商标数据集上获得较高质量的位编码，并具有较高的检索精确度和在线查询效率.
- 深度学习 /
- 哈希学习 /
- 商标检索 /
- 卷积神经网络 /
- 位编码
Abstract: Large-scale image retrieval has great potential for a vast number of applications. The fundamentals of image retrieval lie in image feature extraction and high-efficiency similarity evaluation. Deep learning has great capability for feature representation in image objects, while the Hashing technique has better efficiency for high-dimensional data approximation queries. At present, hash learning technology has been widely researched and applied in large-scale image retrieval for the similarity query. This paper extracts trademark image features using convolutional neural network techniques; the data objects are then encoded with bit codes and an approximate query is applied in Hamming space with high efficiency. In this paper, the convolutional neural network is employed and a deep learning based Hash algorithm is proposed; in addition, the loss function and optimizer for the trademark dataset are studied. By obtaining the bit codes that satisfy the Hash coding criterion, the retrieval of trademark data is efficient. Our method can be divided into offline deep Hash learning and online query stages. Experiments are conducted on real trademark data sets, and the results show that our method can obtain high-quality bit code, which has high retrieval accuracy and online query efficiency.
- deep learning /
- Hash learning /
- trademark image retrieval /
- convolutional neural networks /
- bit encoding

HTML全文

图 1 HTR哈希学习基本过程

Fig. 1 Illustration of hash processing with HTR

下载: 全尺寸图片幻灯片

图 2 卷积神经网络的基本架构

Fig. 2 Basic framework of convolutional neural network

下载: 全尺寸图片幻灯片

图 3 基于CNN网络的HTR哈希架构图

Fig. 3 Hash learning framework of HTR based on CNN

下载: 全尺寸图片幻灯片

图 4 FlickrLogos商标数据示例

Fig. 4 Illustration of FlickrLogos dataset

下载: 全尺寸图片幻灯片

图 5 优化器的损失曲线对比

Fig. 5 Loss curves comparison of different optimizers

下载: 全尺寸图片幻灯片

图 6 正则化参数$\alpha $对MAP的影响

Fig. 6 Result of regularization parameters $\alpha $ on MAP

下载: 全尺寸图片幻灯片

算法 1 HTR算法

输入: 训练集$X$, 网络层数$M$, 学习率$\eta $, 参数$\lambda $和$\alpha $, 迭代次数$R$, 查询$q$
输出: Output:查询结果

1:    (1)离线计算
2:    Step 1初始化:
3:    获得第一层网络输出矩阵
4:    Step 2反向传播优化:
5:    for $r=1, 2, \cdots, R$ do
6:        取集合$X$中的样本
7:        for $m=1, 2, \cdots, M$ do
8:            使用图 3的深度神经网络
9:            计算哈希编码
10:        end
11:        for $m=MI-1, \cdots, 1$do
12:            反向传播更新参数$W^m$和$c^m$
13:        end
14:        使用公式(8)计算损失
15:    End
16:    (2)在线计算
17:    获得海明编码
18:    使用公式(6)计算查询q的距离
19:    对结果根据海明距离从小到大排序

下载: 导出CSV

表 1 HTR默认参数设置

Tab. 1 Algorithm for HTR

参数	值
BATCH_SIZE	200
$\eta $	0.001
EPOCH	800
Bit长度	32

下载: 导出CSV

表 2 HTR正则化方法

Tab. 2 Default parameters for HTR

正则化方法	MAP	训练时间/s
L1	0.729	7 132
L2	0.874	11 897

下载: 导出CSV

表 3 查询质量对比

Tab. 3 Regularization method for HTR

Bit长度	DSH	HTR
12	0.608	0.741
32	0.825	0.874
64	0.733	0.800
128	0.727	0.818

下载: 导出CSV

表 4 在线查询性能

Tab. 4 Comparision of query results

Bit长度	时间/ms
12	0.213
32	0.203
64	0.209
128	0.210

下载: 导出CSV

参考文献(30)

[1]	黄元元, 刘宁钟.利用特征点平均矩特征的商标图像检索[J].中国图象图形学报. 2010, 15(4):637-644. http://d.old.wanfangdata.com.cn/Periodical/zgtxtxxb-a201004014
[2]	宋瑞霞, 孙红磊, 王小春, 等.边界特征和区域特征相结合的商标检索算法[J].软件学报. 2012, 23(2):85-93. http://d.old.wanfangdata.com.cn/Conference/7754349
[3]	张玲, 邹北骥, 孙家广, 等.一种基于极坐标下分块的商标图像检索新方法[J].小型微型计算机系统. 2007, 28(1):66-69. doi: 10.3969/j.issn.1000-1220.2007.01.014
[4]	TURSUN O, KALKAN S. METU dataset: A big dataset for benchmarking trademark retrieval[C]//14th IAPR International Conference on Machine Vision Applications. New York: IEEE, 2015: 514-517.
[5]	YAN Y, REN J, LI Y, et al. Adaptive fusion of color and spatial features for noise-robust retrieval of colored logo and trademark images[J]. Multidimensional Systems and Signal Processing. 2016, 27(4):1-24. https://www.researchgate.net/publication/292677375_Adaptive_fusion_of_color_and_spatial_features_for_noise-robust_retrieval_of_colored_logo_and_trademark_images
[6]	ANUAR F M, SETCHI R, LAI Y K. Trademark image retrieval using an integrated shape descriptor[J]. Expert Systems with Applications. 2013, 40(1):105-121. doi: 10.1016/j.eswa.2012.07.031
[7]	孙兴华, 郭丽.基于子图像多特征组合的商标图像检索[J].模式识别与人工智能. 2002, 15(1):14-20. doi: 10.3969/j.issn.1003-6059.2002.01.003
[8]	WANG J, ZHANG T, SONG J, et al. A survey on learning to hash[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2018, 40(4):769-790. doi: 10.1109/TPAMI.2017.2699960
[9]	DATAR M, IMMORLICA N, INDYK P, et al. Locality-sensitive hashing scheme based on p-stable distributions[C]//Twentieth Symposium on Computational Geometry. New York: ACM, Symposium On Computational Geometry, 2004: 253-262.
[10]	STRECHA C, BRONSTEIN A M, BRONSTEIN M M, et al. LDAHash:Improved matching with smaller descriptors[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2012, 34(1):66-78. doi: 10.1109/TPAMI.2011.103
[11]	WEISS Y, TORRALBA A, FERGUS R, et al. Spectral hashing[C]//Proceedings of the 22nd Annual Conference on Neural Information Processing Systems (NIPS). Vancouver: ACM.2008: 1753-1760.
[12]	WAN J, WANG D, HOI S C, et al. Deep learning for content-based image retrieval: A comprehensive study[C]//Acm Multimedia. New York: ACM, 2014: 157-166.
[13]	CRUZROA A, OVALLE J E, MADABHUSHI A, et al. A deep learning architecture for image representation, visual interpretability and automated basal-cell carcinoma cancer detection[C]//Medical Image Computing and Computer Assisted Intervention. Berlin: Springer Berlin Heidelberg. 2013: 403-410.
[14]	ZHU H, LONG M, WANG J, et al. Deep hashing network for efficient similarity retrieval[C]//National Conference on Artificial Intelligence. CA: AAAI, 2016: 2415-2421.
[15]	彭天强, 粟芳.基于深度卷积神经网络和二进制哈希学习的图像检索方法[J].电子与信息学报. 2016, 38(8):2068-2075. http://d.old.wanfangdata.com.cn/Periodical/dzkxxk201608034
[16]	LIU H, WANG R, SHAN S, et al. Deep Supervised Hashing for Fast Image Retrieval[C]//IEEE Computer Society. IEEE Conference on Computer Vision and Pattern Recognition. New York: IEEE, 2016: 2064-2072.
[17]	XIA R, PAN Y, LAI H, et al. Supervised hashing for image retrieval via image representation learning[C]//National Conference On Artificial Intelligence, CA: AAAI, 2014: 2156-2162.
[18]	龚震霆, 陈光喜, 任夏荔, 等.基于卷积神经网络和哈希编码的图像检索方法[J].智能系统学报. 2016, 11(3):391-400. http://d.old.wanfangdata.com.cn/Periodical/xdkjyc201603015
[19]	ZHANG R, LIN L, ZHANG R, et al. Bit-scalable deep hashing with regularized similarity learning for image retrieval and person re-identification[J]. IEEE Transactions on Image Processing. 2015, 24(12):4766-4779. doi: 10.1109/TIP.2015.2467315
[20]	LIONG V E, LU J, WANG G, et al. Deep hashing for compact binary codes learning[C]//Computer Vision And Pattern Recognition. New York: IEEE, 2015: 2475-2483.
[21]	GUO J, ZHANG S, LI J, et al. Hash learning with convolutional neural networks for semantic based image retrieval[C]//Pacific-Asia Conference on Knowledge Discovery and Data Mining. Berlin: Springer, 2016: 227-238.
[22]	向雷, 肖诗斌, 林春雨, 等.基于轮廓与SIFT特征组合的商标图像检索[J].计算机工程与应用. 2013, 49(19):167-172. doi: 10.3778/j.issn.1002-8331.1112-0552
[23]	ANDONI A, INDYK P, NGUYEN H L, et al. Beyond locality-sensitive hashing[C]//Society for Industrial and Applied Mathematics. Proceedings of the twenty-fifth annual ACM-SIAM symposium on Discrete algorithms. New York: ACM, 2014: 1018-1028.
[24]	LI H, LIN Z, SHEN X, et al. A convolutional neural network cascade for face detection[C]//Computer Vision And Pattern Recognition. New York: IEEE, 2015: 5325-5334.
[25]	周飞燕, 金林鹏, 董军.卷积神经网络研究综述[J].计算机学报. 2017, 40(6):1229-1251. http://d.old.wanfangdata.com.cn/Periodical/jsjxb201706001
[26]	KRIZHEVSKY A, SUTSKEVER I, HINTON G E, et al. ImageNet Classification with Deep Convolutional Neural Networks[C]//Conference and Workshop on Neural Information Processing Systems. Cambridge: MIT Press, 2012: 1097-1105.
[27]	YANG H, LIN K, CHEN C, et al. Supervised learning of semantics-preserving hash via deep convolutional neural networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2018, 40(2):437-451. doi: 10.1109/TPAMI.2017.2666812
[28]	GOOGLE. Tensorflow.[EB/OL]. (2015-04-10)[2018-04-20]. https://tensorflow.google.cn/get-started/.
[29]	CHRISTIAN E. Flickrlogos.[EB/OL]. (2011-05-15)[2018-04-10]. https://www.multimedia-computing.de/flickrlogos/data/.
[30]	KINGMA D P, BA J. Adam: A method for stochastic optimization[J/OL] CoRR, 2014, abs/1412. 6980: 1-15.[2018-05-20]. https://arxiv.org/pdf/1412.6980v2.pdf.