Sparsification of Graph Laplacian for Image Indexing using Multidimensional Spectral Hashing

Abstract

Many multimedia applications need an algorithm to search similar images from a large scale database. Hashing based techniques are used for searching similar images in practice. The existing sparsification of graph laplacian with spectral hashing for similarity search is suitable only
for smaller neighborhoods. But in the most of the cases, multimedia applications requires an algorithm for larger neighborhoods. This creates a research potential to develop a novel approach for generating optimal binary code to retrieve larger neighborhoods. This paper proposes multidimensional spectral hashing that uses sparsification of graph laplacian. Multidimensional spectral hashing uses outer product Eigen functions to improve the codes. The exponential growth of outer product functions are handled using kernel-trick. This makes our proposed algorithm to achieve storage-efficient multi- dimensional spectral hashing. The performance analysis of our proposed algorithm shows better result in terms of binary code generation, true
positives and retrieving neighbor’s images.