Unsupervised deep hashing with node representation for image retrieval. (April 2021)

• Record Type: Journal Article
• Authors: Wang, Yangtao; Song, Jingkuan; Zhou, Ke; Liu, Yu
• Is Part Of: Pattern recognition. Volume 112(2021)
• Publication Date: 2021-04
• Subjects: Deep hashing -- GCN -- Node representation -- Image retrieval; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2020.107785
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: An unsupervised deep hashing framework that consists of node representation learning stage and hash function learning stage is proposed. In the first stage, we utilize graph convolution network to integrate the relationships between samples into node representations. In the second stage, we use above node representations to fast achieve an end-to-end hash model to generate semantic hash codes. Extensive experiments show NRDH can achieve higher performance and outperform the state-of-the-art unsupervised deep hashing methods. Abstract: Supervised graph convolution network (GCN) based hashing algorithms have achieved good results by recognizing images according to the relationships between objects, but they are hard to be applied to label-free scenarios. Besides, most existing unsupervised deep hashing algorithms neglect the relationships between different samples and thus fail to achieve high precision. To address this problem, we propose NRDH, an unsupervised D eep H ashing method with N ode R epresentation for image retrieval, which adopts unsupervised GCN to integrate the relationships between samples into image visual features. NRDH consists of node representation learning stage and hash function learning stage. In the first stage, we treat each image as a node of a graph and design GCN-based AutoEncoder, which can integrate the relationships between samples into node representation. In the second stage, we use above node representations to guide the network and help learn the hash function to fast achieve an end-to-end hash model to generate semantic hash codes. Extensive experiments on CIFAR-10, MS-COCO and FLICKR25K show NRDH can achieve higher performance and outperform the state-of-the-art unsupervised deep hashing methods.