Anchor-Based Self-Ensembling for Semi-Supervised Deep Pairwise Hashing

Abstract

Deep hashing has attracted considerable attention to tackle large-scale retrieval tasks, because of automatic and powerful feature extraction of convolutional neural networks and the gain of hashing in computation and storage costs. Most current supervised deep hashing methods only utilize the semantic information of labeled data without exploiting unlabeled data. However, data annotation is expensive and thus only scarce labeled data are available, which are difficult to represent the true distribution of all data. In this paper, we propose a novel semi-supervised deep pairwise hashing method to leverage both labeled and unlabeled data to learn hash functions. Our method utilizes the transduction of anchors to preserve the pairwise similarity relationship among both labeled and unlabeled samples. Additionally, to explore the semantic similarity information hidden in unlabeled data, it adopts self-ensembling to create strong ensemble targets for latent binary vectors of training samples and form a consensus predicting similarity relationship to multiple anchors. Unlike previous pairwise based hashing methods without maintaining the relevance among similar neighbors, we further explain and exhibit the capability of our method on preserving their relevance through calculating their similarities to anchors. Finally, extensive experiments on benchmark databases demonstrate the superior performance of the proposed method over recent state-of-the-art hashing methods on multiple retrieval tasks. The source codes of the proposed method are available on: https://github.com/xsshi2015/Semi-supervised-Deep-Pairwise-Hashing .