A Unified Binary Embedding Framework for Image Retrieval

Hashing, as an efficient retrieval strategy, has been extensively studied in the image retrieval community. It aims to map the original high-dimensional features into compact binary codes, which dramatically reduces the cost of computing similarity with economic memory consumption. In general, most existing algorithms contain two stages: projecting high-dimensional descriptors into low-dimensional features, and encoding the low-dimensional features as binary strings. Although these two-stage approaches are intuitive and effective, certain beneficial information to binary encoding is inevitably discarded in the dimensionality reduction, which tends to lossy encoding. In this paper, we propose a novel hash algorithm where integrate subspace reconstruction and binary quantization into a unified framework and compact hash codes can be generated directly from high-dimensional features. More specifically, the proposed method discovers a latent subspace from the original feature space which is suitable for hash learning, thus alleviating the heavy information loss caused by traditional dimensionality reduction strategy. Meanwhile, thanks to the excellent property of rotational invariance in Euclidean space, the proposed method avoid the challenge of directly optimizing the binary matrix. Specially, our approach essentially is built on l 2 estimator obeying the unsupervised paradigm, without involving any labeled data. Furthermore, we design an effective iterative alternating strategy to optimize the proposed model, so as to generate compact binary codes for the subsequent image matching. Extensive experiments on public benchmarks demonstrate that our method outperforms the compared method by a significant margin.