Efficient Indexing Structures for Multimedia Data

Abstract

Information retrieval for multimedia data is very subjective. Depending on the application specific features need to be accumulating. We study local and global feature extraction method to characterize feature space into metric, high dimensional and multi-dimensional vector space. We work on parallelization of geometric hashing based space partitioning index structure on GPU to index features extracted from SURF. We also implement and parallelize geometric hashing technique with transformation for mean invariant and alignment along principal directions on GPU. For GPU based classical geometric hashing, achieved speed-up is in the range 14x-61x for off-line indexing and 1.08x-10.06x for online searching compared to sequential version with different partitioning sizes. The variation with transformation function obtains speed-up in the range 12.12x-63.13x for off-line indexing and 1.02x-5.82x for online searching. Proposed index structure is extending with multi-query support. Speed-up in the range of 1.69x-460.19x is achieved in online search for 1 to 10 queries to be executed simultaneously. We propose newer index structure, geometric hashing with alignment along SVD based basis vector. Performance get boost by 10% on average in online searching as compared to classical approach. For GPU based geometric hashing with SVD based alignment speed-up is in range of 14.3x-82.28x for off-line indexing and 3.54x-4255.88x for online searching.