Um Arcabouço escalável, eficiente e adaptativo para busca aproximada em conteúdo multimídia

Abstract

Content-based multimedia retrieval (CBMR) applications are becoming increasingly popular in several online services which handle large volumes of data and are submitted to high query and objects insertions (multimedia descriptors) rate. Although the complexity of these applications is in different aspects, find the objects closest to a query one, is undoubtedly one of the most costly operations. To solve this problem, several recent work proposes the parallelization in distributed memory of Approximate nearest neighbors (ANN) search. It was observed that the highlighted solutions present in the literature employ parallelization strategies that lead to a dataset partitioning creating processing bottlenecks and also show upsuboptimal to the use of computational resources, present in each processing node,to handle the high variables queries and insertions streans. Thus, in this thesis work, strategies were proposed that overcome the limitations pointed out and extend the ability of ANNS distributed memory parallelization approaches to efficiently handle intense and variable query demands in large and growing volumes of multimedia objects.