Information Technology Reference
In-Depth Information
image contains a car, it might be automatically associated with the class
car
(labelled
with the label
car
).
In this paper we study the problem of image content recognition by using SIFT [14]
and SURF [5] local features, to represent image visual content, and kNN based classi-
fiers to decide about the presence of conceptual content. In more details we will define
20 different functions that measure similarity between images. These functions are de-
fined using various options and combinations of local feature matching and similarities.
Some of them also take into consideration geometric properties of the matching local
features. These functions are used in combination of a standard Single-label Distance-
Weighted kNN algorithm. In addition we also propose a new classification algorithm
that extend the traditional kNN classifiers by making direct use of similarity between
local features, rather than similarity between entire images. We will see that the similar-
ity functions that also make use of geometric considerations offer a better performance
than the others. However, the new kNN based classifier that exploit directly the simi-
larity between local features has an higher performance even without using geometric
information.
The paper is organized as follows. In Section 3 we briefly introduce local features.
In Section 4 we present various iamge similarity features relying on local features to
be used with a kNN classification algorithm. Section 5 propose a novel classification
approach. Finally, Sections 6 and 7 presents the experimental results. An earlier version
of this paper has been presented at the Third International Conference on Agents and
Artificial Intelligence [1].
2
Related Work
The first approach to recognizing location from mobile devices using image-based web
search was presented in [17]. Two image matching metrics were used: energy spec-
trum and wavelet decompositions. Local features were not tested. In the last few years
the problem of recognizing landmarks have received growing attention by the research
community. In [16] methods for placing photos uploaded to Flickr on the World map
was presented. In the proposed approach the images were represented by vectors of
features of the tags, and visual keywords derived from a vector quantization of the
SIFT descriptors. In [13] a combination of context- and content-based tools were used
to generate representative sets of images for location-driven features and landmarks.
Visual information is combined with the textual metadata while we are only consider-
ing content-based classification. In [19], Google presented its approach to building a
web-scale landmark recognition engine. Most of the work reported was used to imple-
ment the Google Goggles service [11]. The approach makes use of the SIFT feature.
The recognition is based on best matching image searching, while our novel approach
is based on local features classification. In [7] a survey on mobile landmark recog-
nition for information retrieval is given. Classification methods reported as previously
presented in the literature include SVM, Adaboost, Bayesian model, HMM, GMM. The
kNN based approach which is the main focus of this paper is not reported in that survey.
In [9], various MPEG-7 descriptors have been used to build kNN classifier committees.
However local features were not considered.
