Improving Similarity Search in Face-Images Data

Abstract

Similarity search involves finding all the face-images in a database, which are similar to a desired face-image, based on some distance measure. Comparing the desired face-image to all the face-images in a large dataset is prohibitively slow. If face-images can be placed in a metric space, search can be sped up by using a metric data structure. In this work, we evaluate the performance of range queries with metric data structures (LAESA, VPtree, DSAT, HDSAT1, HDSAT2, LC, RLC and GNAT) when the metric spaces are face-images data with the Euclidean distance. The experimental results show that all data structures reduce the ratio between the number of distances computed and the database size. Moreover, the LAESA has the best performance in the majority of the experimental cases, but the RLC competes with the other metric data structures, and has the best results when compared with the other dynamic metric data structures.