SERVE: Soft and Equalized Residual VEctors for image retrieval

doi:10.1016/j.neucom.2016.04.047

CORC > 北京大学 > 信息科学技术学院

	SERVE: Soft and Equalized Residual VEctors for image retrieval
	Li, Jun ; Xu, Chang ; Gong, Mingming ; Xing, Junliang ; Yang, Wankou ; Sun, Changyin
刊名	NEUROCOMPUTING
	2016
关键词	SERVE Manifolds Multi-graph embedding Graph ensemble Image retrieval OBJECT RETRIEVAL RELEVANCE FEEDBACK RE-RANKING REPRESENTATION FEATURES DATABASE VLAD
DOI	10.1016/j.neucom.2016.04.047
英文摘要	In the last decade, a wide variety of image signatures, e.g., Bag-of-Visual-Words (BOVW), Fisher Vector (FV), and Vector of Locally Aggregated Descriptor (VLAD), have been developed for effective image retrieval. These image signatures, however, are either computationally expensive or simplified for the purpose of trading accuracy for efficiency. To simultaneously guarantee efficiency and effectiveness, we propose a novel image signature termed Soft and Equalized Residual VEctors (SERVE) which is more discriminatively formulated and maintains higher accuracy. It improves VLAD by encoding the variability in within-cluster feature points into the summation of Residual Vectors.(RV) while manifesting superiority in computational efficiency over FV. To find the latent low-dimensional manifolds underlying in the SERVE feature space, we propose to partition the original feature space into separate subspaces by random projections and employ multi-graph embedding to obtain additional performance gain. In particular, we make use of two fusion strategies for graph ensemble to generate a holistic representation. Extensive empirical studies carried out on the three retrieval-specific public benchmarks reveal that our method outperforms existing state-of-the-art methods and provides a promising paradigm for the image retrieval task. (C) 2016 Published by Elsevier B.V.; National Natural Science Foundation of China [61473086, 61375001]; Natural Science Foundation of Jiangsu Province [BK20140566, BK20150470]; SCI(E); ARTICLE; lijunautomation@gmail.com; changxu1989@gmail.com; gongmingnju@gmail.com; jlxing@nlpr.ia.ac.cn; wankou.yang@yahoo.com; cysun@seu.edu.cn; 202-212; 207
语种	英语
内容类型	期刊论文
源URL	[http://ir.pku.edu.cn/handle/20.500.11897/447931]
专题	信息科学技术学院
推荐引用方式 GB/T 7714	Li, Jun,Xu, Chang,Gong, Mingming,et al. SERVE: Soft and Equalized Residual VEctors for image retrieval[J]. NEUROCOMPUTING,2016.
APA	Li, Jun,Xu, Chang,Gong, Mingming,Xing, Junliang,Yang, Wankou,&Sun, Changyin.(2016).SERVE: Soft and Equalized Residual VEctors for image retrieval.NEUROCOMPUTING.
MLA	Li, Jun,et al."SERVE: Soft and Equalized Residual VEctors for image retrieval".NEUROCOMPUTING (2016).