Collaborative multiview hashing. (March 2018)

• Record Type: Journal Article
• Authors: Chen, Zhixiang; Zhou, Jie
• Is Part Of: Pattern recognition. Volume 75(2018:Mar.)
• Page Start: 149
• Page End: 160
• Publication Date: 2018-03
• Subjects: Multiview hashing -- View collaboration -- Nonlinear hashing -- Binary code; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2017.02.026
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: We exploit the diverse information of multiview representations by utilizing the collaboration between view representations to learn the binary codes in each view such that they are predictive to each other. We exploit the correlation between the view representations and the semantic labels to preserve the semantic relationship between data samples. We employ nonlinear hashing functions as the projection in each view to preserve the local data structure in each view. Abstract: In this paper, we propose a collaborative multiview hashing (CMH) approach to incorporate multiview representations into the binary code learning for scalable visual retrieval. Unlike most existing multiview hashing methods which learn linear projections to preserve the fused similarity relationship across different views in unsupervised manner, we employ the nonlinear hashing functions as the projection in each view and exploit the diverse information of multiview representations by utilizing the collaboration between view representations and the correlation between the view representations and the semantic labels. Specifically, the binary codes in each view are constrained to be predictive to each other to exploit the collaboration between the descriptors in different views that describe the same sample. Furthermore, the binary codes in all views are enforced to preserve the semantic relationship between data samples. The hashing functions are implemented in the form of multi-layer neural network with nonlinear transformations at each layer and trained with both the view collaboration and semantic preserving constraints on the outputs. Experimental results on two datasets validate the superiority of the proposed approach in comparison with several state-of-the-art hashing methods.