Affiliations: Centro de Informática, Universidade Federal de Pernambuco, Av. Jornalista Anibal Fernandes, s/n - Cidade Universitária, Recife, Brazil
Correspondence:
[*]
Corresponding author: R.G. Mesquita, entro de Informática, Universidade Federal de Pernambuco, Av. Jornalista Anibal Fernandes, s/n - Cidade Universitária, Recife, Brazil. E-mail:[email protected]
Abstract: The search for an object in an image commonly occurs by
detecting interest points in the whole scene, followed by matching all
descriptors extracted from the selected points with the descriptors of the
target. This paper proposes an iterative analysis of the scene by using
saliency maps to guide the search to more promising regions, instead of
processing the whole image at once. To do so, a patch-based version of SURF
(Speeded-Up Robust Features detector) is introduced just as a saliency
detection method, named Background Laplacian Saliency (BLS), which is based
on a pixel's distance from the estimated background of the scene combined
with a Laplacian-based operator. Using BLS and several state-of-the art
saliency algorithms, the application of saliency maps to guide the
recognition of objects instances using the patch-based SURF is investigated.
Experiments showed that when the search object is present in the scene, the
search's speed is significantly improved, but, on the other hand, when the
target is absent, the required processing time increases. Based on this
finding, this work also proposes the application of saliency maps to define
the level of detail in which each scene location should be processed. Using
a novel and practical evaluation, which considers both speed and accuracy of
saliency detection algorithms, it is demonstrated that the processing time
of visual search is reduced to about 52% of the time required by SURF
when the target's presence is guaranteed, and from 63% to 79% of that
time when no guarantee exists, without any significant loss of accuracy, for
both cases.
