Multiple Orthogonal K-means Hashing

Abstract

Hashing methods are efficient in dealing with large scale image retrieval problems. Current hashing methods, such as the orthogonal k-means, using coordinate descent algorithm to minimize quantization error usually yield unstable performance. It is because the coordinate descent algorithm only provides a local optimum solution. The orthogonal k-means develops a new model with a compositional parameterization of cluster centers to efficiently represent multiple centers. The objective of the orthogonal k-means is to minimize the quantization error by using the coordinate descent algorithm to find the optimal rotation, scaling and translation on descriptor vectors of images. The performance of the orthogonal k-means is dependent on the initialization of the rotation matrix. In this work, we propose the multiple ok-means hashing method to reduce the instability of performance of the orthogonal k-means hashing. For large scale retrieval problems, standard multiple hash tables methods using M tables require M times storage in comparison to single hash table schemes. We propose a binary code selection scheme to reduce the storage of the multiple orthogonal k-means to use the same size of storage as for single table’s. Experimental results show that the proposed method outperforms ok-mean using the same size of storage.