期刊文献+

一种有效深度哈希图像拷贝检测算法

AN EFFECTIVE DEEP HASH IMAGE COPY DETECTION ALGORITHM
下载PDF
导出
摘要 目前拷贝检测中的图像哈希方法由于手工设计特征和线性映射带来的限制,检测精度不高。为了解决这一难题,提出一种端到端的深度哈希拷贝检测算法——DHCD。构建多尺度孪生卷积神经网络,以空间金字塔分层池化的方式来获得图像对的显著性特征;在新设计的哈希损失函数作用下,既保持了特征在语义结构上的相关性,又使得特征输出接近于目标哈希码;通过挖掘难分样本,[JP2]对难分样本再训练,提升了模型的识别效果。在拷贝数据集上的实验结果表明,该算法与当前主流的图像哈希算法相比,准确率提升了10%左右,且效率没有降低。 At present,image hashing method in copy detection has low detection accuracy due to limitations imposed by manual design features and linear mapping.In order to solve this problem,we propose an end-to-end deep hash copy detection algorithm—DHCD.It constructed a multi-scale twin-convolution neural network to obtain the salient features of image pair in the way of spatial pyramid hierarchical pooling.Then,under the effect of the newly designed hash loss function,it not only kept the correlation of features in the semantic structure,but also made the output of features close to the target hash code.In addition,the recognition effect of the model was improved by digging difficult samples and retraining the difficult samples.Experimental results on copy dataset show that the DHCD algorithm improves accuracy by about 10%compared with current mainstream image hashing algorithm,and the efficiency is not decreased.
作者 刘琴 袁家政 刘宏哲 李兵 王佳颖 叶子 Liu Qin;Yuan Jiazheng;Liu Hongzhe;Li Bing;Wang Jiaying;Ye Zi(Beijing Key Laboratory of Information Service Engineering,Beijing Union University,Beijing 100101,China;Department of Academic Research,Beijing Open University,Beijing 100081,China;State Key Laboratory of Pattern Recognition,Institute of Automation,Chinese Academy of Science,Beijing 100190,China;Ministry of Science,Technology and Information,State Grid General Aviation Company Limited,Beijing 102209,China)
出处 《计算机应用与软件》 北大核心 2020年第3期213-219,303,共8页 Computer Applications and Software
基金 国家自然科学基金项目(61571045,61871039) 国家科技支撑项目(2015BAH55F03) 北京成像技术高精尖创新中心高精尖项目(BAICIT-2016002)。
关键词 拷贝检测 深度哈希 多尺度 哈希损失 挖掘难分样本 Copy detection Deep hash Multi-scale Hash loss Digging difficult samples
  • 相关文献

参考文献4

二级参考文献47

  • 1He, Xiaofei, Partha Niyogi. Locality preserving pro- jections[C]//NIPS, 2003,16 : 234-241.
  • 2Cvetkovic, D. M. , et al. Recent results in the theory of graph spectr[J]. North Holland, 1988, 36: 193- 211.
  • 3Smadja, F. Retrieving collocations from text: Xtract[J]. Computational linguistics, 1993,19(1) .. 143-177.
  • 4Church, K. W. , P. Hanks. Word association norms, mutual information, and lexicographyEJ:. Computa- tional linguistics, 1990,16 (1) .. 22-29.
  • 5Berger, A. L. , Pietra, V. J. I3. , Pietra, S. A. D. A maximum entropy approach to natural language pro- cessing[J]. Computational linguistics, 1996,22 (1) : 39- 71.
  • 6Belkin, M. , P. Niyogi. Laplacian eigenmaps and spec- tral techniques for embedding and clustering[J]. Ad- vances in neural information processing systems, 2001, 14:585-591.
  • 7Powers, 13. M. Evaluation: From precision, recall and f-factor to roc, informedness, markedness : correla- tion[J]. School of Informatics and Engineering, Fliw ders University, Adelaide, Australia, 2007, Tech. Rep. SIE-07-001.
  • 8S. Baluja, M. Covell. Learning to hash: forgivinghash functions and applications [C]//Data Mining and Knowledge Diseovery, 2008,17 (3) ; 402-480.
  • 9C. D. Manning, P. Raghavan, H. Sch"utze. Introduc- tion to information retrieval[M], volume 1. London: Cambridge University Press, 2008.
  • 10Zhang, L. , Qiao, L. , Chen, S. Graph-optimized lo- cality preserving projections [C]//Pattern Recogni- tion,2010,43(6) :1993-2002.

共引文献6

相关作者

内容加载中请稍等...

相关机构

内容加载中请稍等...

相关主题

内容加载中请稍等...

浏览历史

内容加载中请稍等...
;
使用帮助 返回顶部