Dual local consistency hashing with discriminative projections selection
Highlights
► We propose a novel hashing approach based on dual local consistency. ► We propose a discriminative projection selecting scheme to learn hash functions. ► Our method outperforms the state-of-the-art methods for fast image search.
Introduction
With the rapid evolution and development of the Internet, the visual contents (e.g. image, video) on web are explosively growing. In order to efficiently exploit such enormous web contents, fast search in large scale has become one of the most critical techniques. It is of great importance to many applications, such as image or video annotation [27], [28], social image retrieval [29]. The traditional content-based image retrieval techniques usually adopt exhaustive comparing the query image with pooled database, which is infeasible because the linear complexity is not scalable in practical situations. For example, the photo sharing website Flickr has over 4 billion images. Another visual content sharing website YouTube receives more than 20 h of uploaded videos per minute. Besides, most large-scale content-based image retrieval applications suffer from the curse of dimensionality since visual descriptors usually have hundreds or even thousands of dimensions. Therefore, beyond the infeasibility of exhaustive search, storage of the original data also becomes a challenging problem.
Over the past decades, several Approximate Nearest Neighbor (ANN) search techniques have been developed for large scale applications. Although there exist many tree-based methods [2], [6], [12], [16], [22] that can perform similarity search effectively in a low-dimensional feature space, for high-dimensional cases and applications with memory constrains, hashing-based ANN techniques have attracted more attention. Hashing-based methods are promising in accelerating similarity search for their capability of generating compact binary codes for a large number of images in the dataset so that similar images will have close binary codes. Retrieving similar neighbors is then accomplished simply by finding the images that have codes with a small Hamming distance from the query. It is extremely fast to perform similarity search over such binary codes [23], because (1) the encoded data are highly compressed and thus can be loaded into the main memory; (2) the hamming distance between two binary codes can be computed efficiently by using bit XOR operation and computing the number of set bits (an ordinary PC today would be able to do millions of Hamming distance computation in just a few milliseconds).
Many hashing algorithms have been proposed to address the fast retrieval issue in recent years. These hashing-based methods for fast image retrieval can be considered as embedding high dimensional feature vectors to a low dimensional Hamming space, while retaining as much as possible the semantic similarity structure of data. In terms of if labeled data is needed, hashing methods can be roughly divided into two categories: unsupervised methods [1], [31], [33] and supervised methods [9], [21], [25], [26]. Unsupervised methods, such as Locality Sensitive Hashing (LSH) [1] and spectral hashing (SH) [31], only use the unlabeled data to generate binary codes for given samples, while supervised methods which incorporate the label information, such as Restricted Boltzmann Machines (RBMs) [9] and Sequential Projection Learning for Hashing (SPLH) [26], are able to preserve the semantic similarity and thus facilitate semantic retrieval and classification.
Although the existing hashing methods have shown success in large-scale image search, there are some shortcomings of them. For example, the traditional spectral hashing (SH) [31] and its extension self-taught hashing (STH) [33] learn the binary codes by making the similar images close to each other in the Hamming space. However, the codes generated in this way may not have the full discriminative power for retrieval and classification task, because a good embedding should not only make the similar images close but also make the dissimilar images far away. In addition, a common step in many binary coding methods [25], [31] is performing principal component analysis (PCA) to get the projection vectors. However, the projections learned in this way do not have the complementary capability, i.e., the errors generated by the previous projection should be corrected as many as possible by the following projection. Therefore, the learned binary codes also lack discriminative power to some extent.
In order to address the above-mentioned problems, we propose a novel discriminative hashing approach based on dual local consistency, which does not only make the similar images have the same codes but also make the dissimilar images have different codes. Moreover, we adopt a more discriminative projection selecting scheme, in which we select the PCA projection for each bit of the codes sequentially so that the current projection can correct the errors generated by the previous projection as many as possible.
The rest of the paper is organized as follows. Section 2 reviews related work. Section 3 describes our dual local consistency hashing approach with discriminative projections selection. In Section 4, we introduce the experiments on two publicly available datasets. Finally, conclusions are given in Section 5.
Section snippets
Fast similarity search
Given a set of data points, the objective in nearest neighbor search is to find a subset that is most similar to the query. Exhaustively comparing the query with each sample in the database is computationally infeasible because linear complexity is not acceptable for large-scale database. To avoid excessive computational and memory costs, an Approximate Nearest Neighbor (ANN) search is more appealing than exhaustive comparing with sublinear query complexity [20].
There has been extensive
The proposed approach
In this section, we will introduce dual local consistency hashing, which do not only make the similar images have similar binary codes but also dissimilar images with different binary codes. In addition, we introduce a more effective projection selecting scheme that can sequentially choose the most discriminative projection for each hash function. Therefore, the binary codes learned by our method perform much better than the other approaches.
Experiments
In this section, we evaluate the proposed approach on the USPS digit and CIFAR-10 datasets. We compare its performance with five state-of-the-art binary coding methods:
- 1.
Locality Sensitive Hashing (LSH) [1];
- 2.
Restricted Boltzmann Machines (RBMs) [9];
- 3.
Spectral Hashing (SH) [31];
- 4.
Semi-Supervised Hashing (SSH) [25];
- 5.
Sequential Projection Learning for Hashing (SPLH) [26].
Conclusion
In this paper, we propose a novel hashing approach for fast image retrieval. Unlike many traditional hashing methods which only preserve the similarity structure of images in a global manner, our method is based on dual local consistency. In our approach, not only the similar images are projected to the same hash codes, but also the dissimilar images are projected to different hash codes. Moreover, our approach adopts a more discriminative projection selecting scheme, which can choose more
Acknowledgement
This work was supported in part by the 973 Programm under Project 2010CB327905, by the National Natural Science Foundation of China under Grant 61170127, 60975010, 60833006, and 61070104.
References (33)
- et al.
Semantic hashing
International Journal of Approximate Reasoning
(2009) - A. Andoni, P. Indyk, Near-optimal hashing algorithms for approximate nearest neighbor in high dimensions, in:...
- et al.
An optimal algorithm for approximate nearest neighbor searching
Journal of ACM
(1998) Spectral Graph Theory
(1997)- et al.
Introduction to Algorithms
(2001) - M. Datar, N. Immorlica, P. Indyk, V. Mirrokni, Locality-sensitive hashing scheme based on p-stable distributions, in:...
- et al.
An algorithm for finding best matches in logarithmic expected time
ACM Transactions on Mathematical Software (TOMS)
(1977) - A. Gionis, P. Indyk, R. Motwani, Similarity search in high dimensions via hashing, in: Proceedings of the 25th...
- et al.
A fast learning algorithm for deep belief nets
Neural Computation
(2006) - et al.
Reducing the dimensionality of data with neural networks
Science
(2006)
Cited by (4)
R<sup>2</sup>PCAH: Hashing with two-fold randomness on principal projections
2017, NeurocomputingCitation Excerpt :Thus, in order to achieve satisfied results, the LSH based methods typically require long codes and multiple hash tables, which degrade the search efficiency in practice. Data-dependent hashing methods [1,17–22,28–30], on the contrary, attempt to learn good data-aware hash functions by utilizing machine learning techniques such as spectral graph analysis and semi-supervised learning, which are usually demonstrated to be more effective than data-independent LSH. In order to handle the above problems simultaneously in a unified framework, in this paper, we present a novel hashing approach named R2PCAH, where R2 represents the two-fold randomness in terms of both rotations and shifts on the top principal projections of PCA data.
Optimal transforms of random vectors: The case of successive optimizations
2017, Signal ProcessingCitation Excerpt :Techniques associated with data compression are used in a number of areas of signal processing such as, to name a few, robust estimation of principal and minor components of a stochastic signal [1–6], recognition of non-deterministic signals generated by synthetic non-chaotic and chaotic stochastic processes [7], color image watermarking [8], similarity search in a large number if signals [9,10], image resolution enhancement [11], the extrapolated impulse response filter design [12], multisensor networks [13–16].
Beyond semantic attributes: Auxiliary feature discovery for image classification
2014, NeurocomputingAttribute expansion with sequential learning for object classification
2013, Electronic Proceedings of the 2013 IEEE International Conference on Multimedia and Expo Workshops, ICMEW 2013