Multi-Level Supervised Hashing with Deep Features for Efficient Image Retrieval

Wing Ng, Jiayong Li, Xing Tian, Hui Wang, Sam Kwong, Jonathan Wallace

Research output: Contribution to journalArticle

Abstract

Image hashing based on deep convolutional neural networks (CNN), deep hashing, has achieved breakthrough performance for image retrieval. Although CNN features from different layers have different levels of information, most of the existing deep hashing methods utilize only the output of the penultimate fully-connected layer as the feature vector, focusing primarily on semantic information whilst ignoring detailed texture information. This calls for research on multi-level hashing, utilizing multi-level features with different levels of generality. Existing multi-level hashing methods typically concatenate multiple feature vectors into a single one. As a result, different levels of features are not utilized to the full as the complementarity relationship between different levels is not exploited, Furthermore, different levels of similarity will weaken each other. To fill this gap, a novel image hashing method, Multi-Level Supervised Hashing with deep feature (MLSH), is proposed in this paper to further exploit multiple levels of deep image features. It uses a multiple-hash-table mechanism to integrate multi-level features from a single deep convolutional neural network. It takes advantage of the complementarity among multi-level features extracted from different layers of the deep network. High-level features reveal the semantic content of the image, while low-level features provide the structural information that is missing in high-level features. Instead of simple concatenation, several hash tables are trained individually using different levels of features from different layers, which are then integrated in a way for efficient image retrieval. The method has been systematically evaluated through experiments on three real-world image databases, and has thus been demonstrated to set a new state of the art in image hashing, outperforming all of the state-of-the-art hashing methods. Furthermore, the recall and precision can be balanced and improved simultaneously.
Original languageEnglish
JournalNeurocomputing
Publication statusAccepted/In press - 12 Feb 2020

    Fingerprint

Cite this