An Adaptive Hashing retrieval Method of Images Based on Multi-Bit Quantization

Multi-bit quantization is a popular quantization approach in hashing method to retrieve images, but it separately quantizes each dimension of real values, thus may destroy the original neighborhood structure. In this paper, an adaptive multi-bit quantization method is proposed. The method decomposes the original data space into several subspaces and then extends them to a product space. Since there exists a positive correlation between the variance of each subspace and the amount of information in the subspace, the proposed method adaptively allocates the numbers of bits according to the variance of the subspaces and gives more bits to the subspace with larger variance. The proposed adaptive multi-bit quantization scheme makes the hashing method effectively decrease the distortion compared to those which allocating same bits to different subspaces, and greatly increases coding efficiency. Experiments on two large public image datasets, LabelMe and Flickr, and comparisons with some state-of-the-art hashing methods show that the proposed method reduces the quantization error by 30%, and improves the average accuracy of the retrieval results by up to 9.8%, indicating that the proposed method can largely improve the retrieval efficiency by reducing the quantization error.