基于深度自学习的图像哈希检索方法

J4 ›› 2015, Vol. 37 ›› Issue (12): 2386-2392.

基于深度自学习的图像哈希检索方法

欧新宇1，2，伍嘉3，朱恒4，李佶5

(1.云南开放大学云南省干部在线学习学院,云南昆明 650223；
2.华中科技大学计算机科学与技术学院,湖北武汉 430074；3.云南开放大学经济与管理学院,云南昆明 650223；
4.云南大学信息学院,云南昆明 650091；5.昆明长水国际机场信息部,云南昆明 650000)

收稿日期:2014-10-28 修回日期:2015-03-13 出版日期:2015-12-25 发布日期:2015-12-25
基金资助:
云南省教育厅应用基础研究计划资助项目(2012Y503)；云南省科技厅应用基础研究计划项目青年资助项目(2012FD064)；云南开放大学科学研究基金资助项目(201405)；国家自然科学基金资助项目(61274092)

Image hashing retrieval method
based on deep selflearning

OU Xinyu1，2，WU Jia3，ZHU Heng4，LI Ji5

(1.Yunnan Province Cadres Online Learning College,Yunnan Open University,Kunming 650223;
2.School of Computer Science & Technology,Huazhong University of Science and Technology,Wuhan 430074;
3.School of Economics and Management,Yunnan Open University,Kunming 650223;
4.School of Information Science and Engineering,Yunnan University,Kunming 650223;
5.Department of Information,Kunming Changshui International Airport,Kunming 650000,China)

Received:2014-10-28 Revised:2015-03-13 Online:2015-12-25 Published:2015-12-25

摘要/Abstract

摘要：

基于监督学习的卷积神经网络被证明在图像识别的任务中具有强大的特征学习能力。然而，利用监督的深度学习方法进行图像检索，需要大量已标注的数据，否则很容易出现过拟合的问题。为了解决这个问题，提出了一种新颖的基于深度自学习的图像哈希检索方法。首先，通过无监督的自编码网络学习到一个具有判别性的特征表达函数，这种方法降低了学习的复杂性，让训练样本不需要依赖于有语义标注的图像，算法被迫在大量未标注的数据上学习更强健的特征。其次，为了加快检索速度，抛弃了传统利用欧氏距离计算相似性的方法，而使用感知哈希算法来进行相似性衡量。这两种技术的结合确保了在获得更好的特征表达的同时，获得了更快的检索速度。实验结果表明，提出的方法优于一些先进的图像检索方法。

关键词: 自学习, 感知哈希算法, 栈式自编码算法, 无监督学习, 图像检索

Abstract:

Convolutional neural networks are an established powerful selflearning ability in image recognition tasks. However, supervised deep learning methods are prone to overfitting when the labeled data are small or noisy. To solve these problems, we propose a novel deep selflearning image hashing retrieval method, an unsupervised learning. First, we can obtain a function with discriminative features via unsupervised autoencoding networks, which reduces learning complexity, thus enabling training images not to rely on their semantic labels. The algorithm is, therefore, forced to learn more robust features from the massive unlabeled data. In order to speed up the query, a perceptual hash algorithm is employed. The combination of these two techniques guarantee a better feature description and a faster query speed without depending on labeled data. Experimental results demonstrate that the proposed approach is superior to some of stateofthe-art methods.

Key words: self-learning;perceptual hash algorithm;stacked auto-encoding algorithm;unsupervised learning;image retrieval

欧新宇1，2，伍嘉3，朱恒4，李佶5. 基于深度自学习的图像哈希检索方法[J]. J4, 2015, 37(12): 2386-2392.

OU Xinyu1，2，WU Jia3，ZHU Heng4，LI Ji5. Image hashing retrieval method
based on deep selflearning [J]. J4, 2015, 37(12): 2386-2392.

参考文献

［1］Jegou H,Douze M,Schmid C.Product quantization for nearest neighbor search ［J］.IEEE Transactions on Pattern Analysis and Machine Intelligence,2011,33(1):117128.
［2］Wang Jianfeng,Xiao Guoqiang,Jiang Jianmin.An image retrieval algorithm based on the HSI color space accumulative histogram ［J］.Computer Engineering & Science,2007,29(4):5558.(in Chinese)
［3］Wu Yongying,Ma Xiaofei.Implementation of the CBIR system based on colour,space and texture features ［J］.Computer Engineering & Science,2005,27(6):4042.(in Chinese)
［4］Cai Lijun.Research on contentbased image retrieval with relevance feedback［D］.Kaifeng:Henan University,2013.(in Chinese)
［5］Liu Jinling.Chinese junk SMS retrieval based on query words expansion［J］.Computer Engineering,2011,37(8):5254.(in Chinese)
［6］Weiss Y,Torralba A,Fergus R.Spectral hashing［J］.Proc NIPS, 2008,282(3):17531760.
［7］Krizhevsky A,Sutskever I,Hinton G E.Image netclassification with deep convolutional neural networks［C］∥Advances in Neural Information Processing System,2012:1.
［8］Oquab M,Bottou L E O,Laptev I,et al.Learning and transferring midlevel image representations using convolutional neural networks［C］∥Proc of IEEE Conference on Computer Vision and Pattern Recognition,2014:1.
［9］Girshick R,Donahue J,Darrell T,et al.Rich feature hierarchies for accurate object detection and semantic segmentation［J］.arXiv preprint arXiv:13112524.2013.
［10］Lu Min, Huang Yalou, Xie Maoqiang, et al. Rank hash similarity for fast similarity search ［J］.Information Processing & Management,2013,49(1):158168.
［11］Ipeirotis P G. Analyzing the Amazon mechanical turk marketplace［J］.XRDS:Crossroads,The ACM Magazine for Students,2010,17(2):1621.
［12］Hinton G E,Osindero S,Teh Y.A fast learning algorithm for deep belief nets［J］.Neural Computation,2006,18(7):15271554.
［13］Hinton G E.Reducing the dimensionality of data with neural networks［J］.Science,2006,313(5786):504507.
［14］Bengio Y, Lamblin P, Popovici D, et al.Greedy layerwise training of deep networks［C］∥Advances in Neural Information Processing Systems,2007:153.
［15］Lee H,Battle A,Raina R,et al.Efficient sparse coding algorithms［C］∥Advances in Neural Information Processing Systems,2007:801.
［16］Coates A,Ng A Y,Lee H.An analysis of singlelayer networks in unsupervised feature learning［J］.Journal of Machine Learning Research,2011,15:215223.
［17］Wang J, Kumar S, Chang S.Semisupervised hashing for scalable image retrieval［C］∥Proc of CVPR’10,2010:1.
［18］Gionis A,Indyk P,Motwani R,et al.Similarity search in high dimensions via hashing［C］∥Proc of International Conference on Very Large Data Bases,2000:518529.
［19］Hinton G E,Salakhutdinov R R.Reducing the dimensionality of data with neural networks［J］.Science,2006,313(5786):504507.
［20］Torralba A,Fergus R,Weiss Y.Small codes and large image databases for recognition［C］∥Proc of IEEE Conference on Computer Vision & Pattern Recongnition,2008:18.
附中文参考文献：
［2］王剑峰，肖国强，江健民.基于HSI色彩空间累加直方图的图像检索算法［J］.计算机工程与科学,2007,29(4):5558.
［3］吴永英，马笑飞.基于颜色、空间和纹理特征的CBIR系统实现［J］.计算机工程与科学,2005,27(6):4042.
［4］蔡利君.基于内容的交互式图像检索技术的若干问题研究［D］.开封：河南大学,2013.
［5］刘金岭.基于查询词扩展的中文垃圾短信检索［J］.计算机工程,2011,37(8):5254.

[1]	印杰, 黄肖宇, 刘家银, 牛博威, 谢文伟, . 基于预训练语言模型的安卓恶意软件检测方法[J]. 计算机工程与科学, 2023, 45(08): 1433-1442.
[2]	强梓林, 刘建国, 刘云峰, 卫栋, 强彦. 基于时域-频域哈希编码的电网图像检索方法[J]. 计算机工程与科学, 2022, 44(10): 1877-1884.
[3]	李方, 吴国栋, 涂立静, 刘玉良, 查志康, 李景霞. 图自编码器推荐研究综述[J]. 计算机工程与科学, 2022, 44(02): 335-344.
[4]	姬子恒, 王斌. 基于深度学习的草图检索方法研究进展[J]. 计算机工程与科学, 2021, 43(12): 2190-2205.
[5]	方梦华, 姜添, . 基于无监督学习的无人机目标跟踪[J]. 计算机工程与科学, 2021, 43(06): 1024-1031.
[6]	陈金龙, 刘雄飞, 詹曙. 基于无监督生成对抗网络的人脸素描图像真实化[J]. 计算机工程与科学, 2021, 43(01): 125-133.
[7]	贺周雨1,冯旭鹏2,刘利军1,黄青松1,3. 基于SHN模型的商品图像检索方法[J]. 计算机工程与科学, 2019, 41(11): 1991-1999.
[8]	孙一飞，武继刚，张欣鹏. 面向眼底图像小目标检测的无监督学习方法[J]. 计算机工程与科学, 2019, 41(11): 2000-2006.
[9]	袁晖1，廖开阳1,3，郑元林1,2，曹从军1,3，汤梓伟1，邓轩1. 基于CNN特征加权和区域整合的图像检索[J]. 计算机工程与科学, 2019, 41(01): 113-121.
[10]	夏海蛟1,2，谭毅华1,2. 一种面向识别的无监督特征学习算法[J]. 计算机工程与科学, 2018, 40(06): 1103-1110.
[11]	张南1,韩晓军1,2. 融合颜色词袋特征的视觉词汇树图像检索[J]. 计算机工程与科学, 2018, 40(03): 487-493.
[12]	戴月明，张明明，王艳. 协同进化混合蛙跳算法[J]. 计算机工程与科学, 2018, 40(01): 139-147.
[13]	曹玉东，刘艳洋，孙福明，贾旭. 低空间复杂度的LSH算法及其在图像检索中的应用[J]. J4, 2015, 37(02): 379-383.
[14]	彭晏飞1，尚永刚1,王德建2. 一种新的基于SVM和主动学习的图像检索方法[J]. J4, 2014, 36(07): 1371-1376.
[15]	张振华,朱信忠,赵建民,徐慧英. 一种基于局部保形映射（LPP）的相关反馈方法[J]. J4, 2012, 34(6): 101-105.