Please use this identifier to cite or link to this item:
Title: Annotating web images using NOVA: NOn-conVex group spArsity
Authors: Wu, F.
Yuan, Y.
Rui, Y.
Yan, S. 
Zhuang, Y.
Keywords: consistent
group feature selection
image annotation
non-convex group sparsity
Issue Date: 2012
Citation: Wu, F.,Yuan, Y.,Rui, Y.,Yan, S.,Zhuang, Y. (2012). Annotating web images using NOVA: NOn-conVex group spArsity. MM 2012 - Proceedings of the 20th ACM International Conference on Multimedia : 509-518. ScholarBank@NUS Repository.
Abstract: As image feature vector is large, selecting the right features plays a fundamental role in Web image annotation. Most existing approaches are either based on individual feature selection, which leads to local optima, or using a convex penalty, which leads to inconsistency. To address these difficulties, in this paper we propose a new sparsity-based approach NOVA (NOn-conVex group spArsity). To the best of our knowledge, NOVA is the first to introduce non-convex penalty for group selection in high-dimensional heterogeneous features space. Because it is a group-sparsity approach, it approximately reaches global optima. Because it uses non-convex penalty, it achieves the consistency. We demonstrate the superior performance of NOVA via three means. First, we present theoretical proof that NOVA is consistent, satisfying un-biasness, sparsity and continuity. Second, we show NOVA converges to the true underlying model by using a ground-truth-available generative-model simulation. Third, we report extensive experimental results on three diverse and widely-used data sets Kodak, MSRA-MM 2.0, and NUS-WIDE. We also compare NOVA against the state-of-the-art approaches, and report superior experimental results. © 2012 ACM.
Source Title: MM 2012 - Proceedings of the 20th ACM International Conference on Multimedia
ISBN: 9781450310895
DOI: 10.1145/2393347.2393419
Appears in Collections:Staff Publications

Show full item record
Files in This Item:
There are no files associated with this item.

Google ScholarTM



Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.