深度非对称压缩型哈希算法

doi:10.19665/j.issn1001-2400.2021.05.024

摘要/Abstract

摘要：

针对图像检索中很多深度监督哈希算法不能有效地利用大型数据集监督信息和困难样本的问题,提出了一种端到端的非对称压缩型哈希算法。该算法将网络的输出空间分为查询集与数据库集,构造数据监督矩阵,并采用非对称的方式使全局监督信息得到有效的利用。同时,在损失函数中对同类哈希码聚拢程度与不同类哈希码的分离程度进行显式的约束,提高训练过程中模型对困难样本的判别能力。首先,改进骨干特征提取网络SKNet-50,通过添加哈希层和阈值化层,输出查询集矩阵;然后,使用交叉方向乘子方法优化损失函数得到数据库集矩阵;最后,利用交替优化的方法完成深度模型的训练。使用48比特哈希码检索图像时,在CIFAR-10和NUS-WIDE数据集上的平均精度均值达到0.946和0.923,在MS-COCO数据集上检索的平均精度均值可以达到0.881。实验结果表明,提出算法可以学习到更加判别紧凑的哈希码,在检索精度上要优于目前主流的算法。

关键词: 非对称哈希, 图像检索, 深度学习

Abstract:

Most existing deep supervised hashing algorithms in image retrieval fail to effectively utilize difficult samples and the supervised information.In order to solve the this problem,an end-to-end asymmetric compression hashing algorithm is proposed which divides the output space of the network into the query set and database set,constructs the supervised data matrix,and effectively uses the global supervised information in an asymmetric way.Meanwhile,the gathering degree of within-class hash codes and the separation degree of inter-class hash codes are explicitly constrained in the loss function,which improves the discriminative ability of the model on difficult samples under training.First,the hashing layer and thresholding layer are added into the improved backbone feature extraction network,SKNet-50,which outputs the query set matrix.Then,the matrix of the database set is obtained by optimizing the loss function with the alternating direction method of multipliers(ADMM).Finally,the deep model is trained with the alternative optimization method.The proposed method can achieve 0.946,0.923 and 0.811 MAP on the CIFAR-10,NUS-WIDE and MS-COCO datasets,respectively,when the 48-bit hash code is used to retrieve images.Experimental results show that the proposed method can learn more discriminative and compact hash codes,and that the retrieval accuracy is superior to the current mainstream algorithm.

Key words: asymmetric hashing, image retrieval, deep learning

中图分类号:

TP391

闫佳,曹玉东,任佳兴,陈冬昊,李晓会. 深度非对称压缩型哈希算法[J]. 西安电子科技大学学报, 2021, 48(5): 212-221.

YAN Jia,CAO Yudong,REN Jiaxing,CHEN Donghao,LI Xiaohui. Deep asymmetric compression Hashing algorithm[J]. Journal of Xidian University, 2021, 48(5): 212-221.

图/表 8

图1

图2

表1

改进的SKNet-50的网络结构"

输出尺寸	SKNetH-50的层结构
112×112	7×7,64,stride 2
56×56	3×3 max pool,stride 2
56×56	$1 × 1, 128 SK [M = 2, G = 32, r = 16], 128 1 × 1, 256$ ×3
28×28	$1 × 1, 256 SK [M = 2, G = 32, r = 16], 256 1 × 1, 512$ ×4
14×14	$1 × 1, 512 SK [M = 2, G = 32, r = 16], 512 1 × 1, 1024$ ×6
7×7	$1 × 1, 1024 SK [M = 2, G = 32, r = 16], 1024 1 × 1, 2048$ ×3
1×1	7×7 global average pool,1 000-d fc,hash-k

表1

图3

表2

图4

图5

图6

参考文献 21

[1]	尹奇跃, 黄岩, 张俊格, 等. 基于深度学习的跨模态检索综述[J]. 中国图象图形学报, 2021, 26(6):1368-1388.
	YIN Qiyue, HUANG Yan, ZHANG Junge, et al. Survey on Deep Learning Based Cross-Modal Retrieval[J]. Journal of Image and Graphics, 2021, 26(6):1368-1388.
[2]	石娟, 谢德, 蒋庆. 深度共性保持哈希[J]. 西安电子科技大学学报, 2021, 48(3):71-77.
	SHI Juan, XIE De, JIANG Qing. DeepConsistency-Preserving Hashing[J]. Journal of Xidian University, 2021, 48(3):71-77.
[3]	刘颖, 程美, 王富平, 等. 深度哈希图像检索方法综述[J]. 中国图象图形学报, 2020, 25(7):1296-1317.
	LIU Ying, CHENG Mei, WANG Fuping, et al. Deep Hashing Image Retrieval Methods[J]. Journal of Image and Graphics, 2020, 25(7):1296-1317.
[4]	XIA R, PAN Y, LAI H, et al. Supervised Hashing for Image Retrieval via Image Representation Learning[C]// Proceedings of the AAAI Conference on Artificial Intelligence.Palo Alto:AAAI, 2014:2156-2162.
[5]	LIU H, WANG R, SHAN S, et al. Deep Supervised Hashing for Fast Image Rretrieval[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition.Piscataway:IEEE, 2016:2064-2072.
[6]	LI Y, PEI W, ZHA Y, et al. Push for Quantization:Deep Fisher Hashing (2019)[J/OL]. [2019-08-31]. https://arxiv.org/pdf/1909.00206.pdf.
[7]	SHEN F, GAO X, LIU L, et al. Deep Asymmetric Pairwise Hashing[C]// Proceedings of the 25th ACM International Conference on Multimedia.New York:ACM, 2017:1522-1530.
[8]	MA L, LI H, MENG F, et al. Discriminative Deep Metric Learning for Asymmetric Discrete Hashing[J]. Neurocomputing, 2020, 380:115-124. doi: 10.1016/j.neucom.2019.11.009
[9]	JIANG Q Y, LI W J. Asymmetric Deep Supervised Hashing[C]// Proceedings of the AAAI Conference on Artificial Intelligence.Palo Alto:AAAI, 2018:1-8.
[10]	WU B, GHANEM B. φp-Box ADMM:A Versatile Framework for Integer Programming[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019, 41(7):1695-1708. doi: 10.1109/TPAMI.34
[11]	KRIZHEVSKY A, SUTSKEVER I, HINTON G E. ImageNet Classification with Deep Convolutional Neural Networks[J]. Advances in Neural Information Processing Systems, 2012, 25:1097-1105.
[12]	LI X, WANG W, HU X, et al. Selective Kernel Networks[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition.Piscataway:IEEE, 2019:510-519.
[13]	QIAN Q, CHEN L, LI H, et al. DR Loss:Improving Object Detection by Distributional Ranking[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition.Piscataway:IEEE, 2020:12164-12172.
[14]	SHEN F, XU Y, LIU L, et al. Unsupervised Deep Hashing with Similarity-Adaptive and Discrete Optimization[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2018, 40(12):3034-3044. doi: 10.1109/TPAMI.34
[15]	KRIZHEVSKY A. Learning Multiple Layers of Features From Tiny Images:Technical Report TR-2009[R]. Toronto:University of Toronto, 2009.
[16]	CHUA T S, TANG J, HONG R, et al. Nus-Wide:A Real-World Web Image Database from National University of Singapore[C]// Proceedings of the ACM International Conference on Image and Video Retrieval.New York:ACM, 2009:1-9.
[17]	LIN T Y, MAIRE M, BELONGIE S, et al. Microsoft COCO:Common Objects in Context[C]// European Conference on Computer Vision.Heidelberg:Springer, 2014:740-755.
[18]	LI W J, WANG S, KANG W C. Feature Learning Based Deep Supervised Hashing with Pairwise Labels (2015)[J/OL]. [2015-11-12]. https://arxiv.org/pdf/1511.03855.pdf.
[19]	CAO Z, LONG M, WANG J, et al. HashNet:Deep Learning to Hash by Continuation[C]// Proceedings of the IEEE International Conference on Computer Vision.Piscataway:IEEE, 2017:5608-5617.
[20]	ZHU H, LONG M, WANG J, et al. Deep Hashing Network for Efficient Similarity Retrieval[C]// Proceedings of the AAAI Conference on Artificial Intelligence.Palo Alto:AAAI, 2016:1-8.
[21]	WANG X, YI S, KITANI K M. Deep Supervised Hashing with Triplet Labels[C]// Asian Conference on Computer Vision.Heidelberg:Springer, 2016:70-84.

算法	CIFAR-10			NUS-WIDE			MS-COCO
算法	12 bit	24 bit	48 bit	12 bit	24 bit	48 bit	12 bit	24 bit	48 bit
DSH	0.655	0.742	0.800	0.641	0.742	0.798	0.619	0.641	0.655
DPSH	0.643	0.721	0.775	0.794	0.824	0.844	0.627	0.668	0.711
Net	0.700	0.764	0.782	0.782	0.812	0.830	0.641	0.684	0.724
DHN	0.702	0.766	0.781	0.786	0.818	0.841	0.644	0.662	0.712
DTSH	0.731	0.783	0.800	0.800	0.829	0.829	0.684	0.732	0.760
DFH	0.749	0.786	0.801	0.801	0.837	0.837	0.633	0.674	0.717
ADSH	0.864	0.911	0.921	0.829	0.878	0.895	0.832	0.855	0.860
ACDH	0.879	0.935	0.946	0.847	0.897	0.923	0.851	0.867	0.881