Flexible Discrete Multi-view Hashing with Collective Latent Feature Learning

摘要

Multi-view hashing has gained considerable research attention in efficient multimedia studies due to its promising performance on heterogeneous data from various sources. However, its application in discriminative hash codes learning remains challenging as it fails to efficiently capture preferable components from multiple representations. In this work, we propose a novel discriminative multi-view hashing framework, dubbed flexible discrete multi-view hashing, in conjunction with collective latent feature learning by combining multiple views of data and consistent hash codes learning by fusing visual features and flexible semantics. Specifically, an adaptive multi-view analysis dictionary learning model is developed to skillfully combine diverse representations into an established common latent feature space where the complementary properties of different views are well explored based on an automatic multi-view weighting strategy. Moreover, we introduce a collaborative learning scheme to jointly encode the visual and semantic embeddings into an aligned consistent Hamming space, which can effectively mitigate the visual-semantic gap. Particularly, we employ the correntropy induced regularization to improve the robustness of the formulated flexible semantics. An efficient learning algorithm is proposed to solve the optimization problem. Extensive experiments show the state-of-art performance of the proposed method on several benchmark datasets.