Deep Discriminative Supervised Hashing via Siamese Network

Yang LI; Zhuang MIAO; Jiabao WANG; Yafei ZHANG; Hang LI

doi:10.1587/transinf.2017EDL8126

Deep Discriminative Supervised Hashing via Siamese Network

Yang LI, Zhuang MIAO, Jiabao WANG, Yafei ZHANG, Hang LI

Full Text Views

0

Share
Cite this

Summary :

The latest deep hashing methods perform hash codes learning and image feature learning simultaneously by using pairwise or triplet labels. However, generating all possible pairwise or triplet labels from the training dataset can quickly become intractable, where the majority of those samples may produce small costs, resulting in slow convergence. In this letter, we propose a novel deep discriminative supervised hashing method, called DDSH, which directly learns hash codes based on a new combined loss function. Compared to previous methods, our method can take full advantages of the annotated data in terms of pairwise similarity and image identities. Extensive experiments on standard benchmarks demonstrate that our method preserves the instance-level similarity and outperforms state-of-the-art deep hashing methods in the image retrieval application. Remarkably, our 16-bits binary representation can surpass the performance of existing 48-bits binary representation, which demonstrates that our method can effectively improve the speed and precision of large scale image retrieval systems.

Publication: IEICE TRANSACTIONS on Information Vol.E100-D No.12 pp.3036-3040

Publication Date: 2017/12/01

Publicized: 2017/09/12

Online ISSN: 1745-1361

DOI: 10.1587/transinf.2017EDL8126

Type of Manuscript: LETTER

Category: Artificial Intelligence, Data Mining

Authors

Yang LI
  PLA University of Science and Technology (PLAUST)
Zhuang MIAO
  PLA University of Science and Technology (PLAUST)
Jiabao WANG
  PLA University of Science and Technology (PLAUST)
Yafei ZHANG
  PLA University of Science and Technology (PLAUST)
Hang LI
  PLA University of Science and Technology (PLAUST)

Keyword

hashing, convolutional neural network, siamese network, image retrieval

Cite this

Copy

Yang LI, Zhuang MIAO, Jiabao WANG, Yafei ZHANG, Hang LI, "Deep Discriminative Supervised Hashing via Siamese Network" in IEICE TRANSACTIONS on Information, vol. E100-D, no. 12, pp. 3036-3040, December 2017, doi: 10.1587/transinf.2017EDL8126.
Abstract: The latest deep hashing methods perform hash codes learning and image feature learning simultaneously by using pairwise or triplet labels. However, generating all possible pairwise or triplet labels from the training dataset can quickly become intractable, where the majority of those samples may produce small costs, resulting in slow convergence. In this letter, we propose a novel deep discriminative supervised hashing method, called DDSH, which directly learns hash codes based on a new combined loss function. Compared to previous methods, our method can take full advantages of the annotated data in terms of pairwise similarity and image identities. Extensive experiments on standard benchmarks demonstrate that our method preserves the instance-level similarity and outperforms state-of-the-art deep hashing methods in the image retrieval application. Remarkably, our 16-bits binary representation can surpass the performance of existing 48-bits binary representation, which demonstrates that our method can effectively improve the speed and precision of large scale image retrieval systems.
URL: https://globals.ieice.org/en_transactions/information/10.1587/transinf.2017EDL8126/_p

Copy

@ARTICLE{e100-d_12_3036,
author={Yang LI, Zhuang MIAO, Jiabao WANG, Yafei ZHANG, Hang LI, },
journal={IEICE TRANSACTIONS on Information},
title={Deep Discriminative Supervised Hashing via Siamese Network},
year={2017},
volume={E100-D},
number={12},
pages={3036-3040},
abstract={The latest deep hashing methods perform hash codes learning and image feature learning simultaneously by using pairwise or triplet labels. However, generating all possible pairwise or triplet labels from the training dataset can quickly become intractable, where the majority of those samples may produce small costs, resulting in slow convergence. In this letter, we propose a novel deep discriminative supervised hashing method, called DDSH, which directly learns hash codes based on a new combined loss function. Compared to previous methods, our method can take full advantages of the annotated data in terms of pairwise similarity and image identities. Extensive experiments on standard benchmarks demonstrate that our method preserves the instance-level similarity and outperforms state-of-the-art deep hashing methods in the image retrieval application. Remarkably, our 16-bits binary representation can surpass the performance of existing 48-bits binary representation, which demonstrates that our method can effectively improve the speed and precision of large scale image retrieval systems.},
keywords={},
doi={10.1587/transinf.2017EDL8126},
ISSN={1745-1361},
month={December},}

Copy

TY - JOUR
TI - Deep Discriminative Supervised Hashing via Siamese Network
T2 - IEICE TRANSACTIONS on Information
SP - 3036
EP - 3040
AU - Yang LI
AU - Zhuang MIAO
AU - Jiabao WANG
AU - Yafei ZHANG
AU - Hang LI
PY - 2017
DO - 10.1587/transinf.2017EDL8126
JO - IEICE TRANSACTIONS on Information
SN - 1745-1361
VL - E100-D
IS - 12
JA - IEICE TRANSACTIONS on Information
Y1 - December 2017
AB - The latest deep hashing methods perform hash codes learning and image feature learning simultaneously by using pairwise or triplet labels. However, generating all possible pairwise or triplet labels from the training dataset can quickly become intractable, where the majority of those samples may produce small costs, resulting in slow convergence. In this letter, we propose a novel deep discriminative supervised hashing method, called DDSH, which directly learns hash codes based on a new combined loss function. Compared to previous methods, our method can take full advantages of the annotated data in terms of pairwise similarity and image identities. Extensive experiments on standard benchmarks demonstrate that our method preserves the instance-level similarity and outperforms state-of-the-art deep hashing methods in the image retrieval application. Remarkably, our 16-bits binary representation can surpass the performance of existing 48-bits binary representation, which demonstrates that our method can effectively improve the speed and precision of large scale image retrieval systems.
ER -