Large-Margin Supervised Hashing

Abstract

Learning to hash embeds objects (e.g. images/documents) into a binary space with the semantic similarities preserved from the original space, which definitely benefits large-scale tough tasks such as image retrieval. By leveraging semantic labels, supervised hashing methods usually achieve better performance than unsupervised ones in real-world scenarios. However, most existing supervised methods do not sufficiently encourage inter-class separability and intra-class compactness which is quite crucial in discriminative hashcodes. In this paper, we propose a novel hashing method called Large-Margin Supervised Hashing (LMSH) based on a non-linear classification framework. Specifically, LMSH introduces the angular decision margin which could adjust inter-class separability and intra-class compactness through a hyper-parameter for more discriminative codes. Extensive experiments on three public datasets are conducted to demonstrate the LMSH’s superior performance to some state-of-the-arts in image retrieval tasks.