Deep hash: semantic similarity preserved hash scheme

Abstract

A novel hashing scheme based on a deep network architecture is proposed to tackle semantic similarity problems. The proposed methodology utilises the ability of deep networks to learn nonlinear representations of the input features. The equivalence of the neuron layer and the sigmoid smoothed hash functions is introduced, and by incorporating the saturation and orthogonality regulariser, the final compact binary embeddings can be achieved. The experiments illustrate that the proposed scheme exhibits superior improvement compared with conventional hashing methods.Introduction: Hash learning has attracted much attention from researchers in recent years owing to its wide application in large-scale computer vision problems, including information retrieval, object recognition, feature matching and so on. They aim to generate binary embedding from the original feature inputs and generally have a sublinear query time. Although various hash learning methods have been proposed in the literature, there still exists a well-known problem of the 'semantic gap' between the high-level semantics and the extracted lowlevel features in the approximate nearest neighbours (ANNs) search. To address this issue, supervised methods were proposed by incorporating semantic supervision in the training phase. Liu et al.[1] incorporated the supervised information and presented a sequential kernelised hashing method to boost the hashing performance. Wang et al.[2] minimised the empirical error on the labelled data by minimising the entropy of the generated hash bits over the unlabelled dataset to avoid overfitting. Zhang et al.[3] learnt hash functions by training a support vector machine classifier for each bit using the pre-learnt binary codes as class labels. Salakhutdinov and Hinton [4] unsupervisedly trained a deep autoencoder of the word-count vectors to produce compact binary codes for document retrieval.In this Letter, we propose a novel hash scheme, 'deep hash', (DH) that takes advantage of the abstraction ability of the deep architecture and the natural intrinsics of the conventional hash learning to bridge the 'semantic gap'. Stacked restricted Boltzmann machines (RBMs) [5] are employed to build up a deep network architecture for producing highlevel abstraction of the initial input features, whereas for modelling realvalued inputs Gaussian RBM (GRBM) [5] is used instead. We introduce coupling the traditional hash learning with the deep architecture and show the natural relation between the smoothed relaxed hash function and the neural networks. By further employing saturation and an orthogonality regulariser, the final compact binary codes are produced. Extensive experiments on commonly used datasets demonstrate the superior performance of our proposed hash scheme.

Comprehensive analysis and evaluation to unsupervised binary hashing method in image similarity measurement

“…The ground truth is acquired using the Euclidean distance. Datasets used in this experiment are CIFAR‐10 dataset [28, 29], SIFT‐1M dataset [4, 30, 31], and GIST‐1M dataset [4, 31]. Each of detailed datasets is described in Table 1.…”

“…Also, the second test bed is to evaluate the performance of the NN search with CIFAR‐10 dataset [28, 29]. However, the ground truth is class labels.…”

Unsupervised binary hashing method using locality preservation and quantisation error minimisation

“…Then, vectors of the ground truth are acquired as the 50th NN of the testing vectors using the Euclidean distance. The dataset used in this experiment is the CIFAR-10 dataset [7]. This dataset provides 60 000 colour images whose resolution is 32 × 32.…”