期刊名称：International Journal of Computer and Information Technology
印刷版ISSN：2279-0764
出版年度：2016
卷号：5
期号：1
页码：33
出版社：International Journal of Computer and Information Technology
摘要：Hashing is becoming a popular and effective method for nearest neighbor search in large-scale databases, due to its computational and memory efficient. In this paper, we present an efficient unsupervised nonlinear hashing method to transform high-dimensional data to low-dimensional binary data for fast retrieval. Firstly, we use the Nystr.m method to transform the feature space into nonlinear kernel feature space, to capture the similarity property of the data. Secondly, all training data are formulated to maximize the entropy over each hash bit, which can be relaxed to maximize the variance on each hash bit. Then, we solve the objective function by an improved sequential projection learning method. In each projection learning iteration, we reduce the HAE (Hamming Accumulated Errors) through some pseudolabel pairs generated from all previous learning projections. During the process, we only need to store the sum of covariance matrix instead of the similarity matrix to save memory storage. We also use a more efficient method to update the covariance matrix after each iteration. We carry out extensive experiments on two benchmarks, and demonstrate that the proposed method achieves better performance than some state-of-the-art hashing approaches
