Multi-bit quantization based on neighboring structure preservation

Hashing based approximate nearest neighbor search has become a research hotspot in computer vision. Most existing hashing methods concentrate on projection learning, and few efforts are dedicated to quantization coding. In this paper, we present a multi-bit quantization strategy to improve the quantization quality of projection values by adaptively learning quantization thresholds and quantizing each projection dimension with multiple bits. Our method exploits both the similarity and the local structure of samples in the original feature space and the pair-wise samples coding consistency. Extensive experiments on two canonical image datasets have shown that our method consistently outperforms the state-of-the-art quantization methods in terms of query performance.