Deep Factorized Multi-view Hashing for Image Retrieval

Multi-view hashing has been paid much attention to due to its computational efficiency and lower memory overhead in similarity measurement between instances. However, a common drawback of these multi-view hashing methods is the lack of ability to fully explore the underlying correlations between different views, which hinders them from producing more discriminative hash codes. In our work, we propose the principled Deep Factorized Multi-view Hashing (DFMH) framework, including interpretable robust representation learning, multi-view fusion learning, and flexible semantic feature learning, to deal with the challenging multi-view hashing problem. Specifically, instead of directly projecting the features to a common representation space, we construct an adaptively weighted deep factorized structure to preserve the heterogeneity between different views. Furthermore, the visual space and semantic space are interactively learned to form a reliable hamming space. Particularly, the flexible semantic representation is obtained by learning regressively from semantic labels. Importantly, a well-designed learning strategy is developed to optimize the objective function efficiently. DFMH as well as compared methods is tested on benchmark datasets to validate the efficiency and effectiveness of our proposed method. The source codes of this paper are released at:https://github.com/chenyangzhu1/DFMH.