Two-phase mapping hashing

Abstract

Hashing methods have become important retrieval methods for large image databases due to its fast searching time. However, current hashing methods mapping high dimensional real-valued vectors of images to low dimensional hash codes directly may not yield good performance. These methods perform dimensionality reduction together with the Hamming quantization which lead to a difficult nonlinear problem. In this work, we propose a Two-phase Mapping Hashing (TMH) method which first maps images from the high dimensional real-valued space to a high dimensional Hamming space (hash code) using the SKLSH to preserve pairwise similarity. Then, the mapping from the high dimensional Hamming space to a low dimensional Hamming space is found via a minimization of reconstruction error between the two Hamming spaces. The shorter hash code created by the TMH preserves pairwise similarity of images in the input space and yields better precision and recall rates in comparison to SKLSH. Experimental results show that the TMH outperforms the LSH, the Compressed Hashing, the ITQ and the SKLSH.