Speeding up the combination of multiple descriptors for different boundary conditions

Abstract

Content-based complex data retrieval is becoming increasingly common in many types of applications. The content of these data is represented by intrinsic characteristics, extracted from them which together with a distance function allows similarity queries. Aimed at reducing the “semantic gap”, characterized by the disagreement between the computational representation of the extracted low-level features and how these data are interpreted by the human perception, the use of multiple descriptors has been the subject of several studies. This paper proposes a new method to carry out the combination of multiple descriptors for different boundary conditions in which the balancing is carried out in pairs, starting by the best candidate descriptor. In the experiments, the proposed method achieved computational cost up to 3650 times smaller than the exhaustive search for the best linear combination of descriptors, keeping almost the same average precision, with variations lower than 0.9%.