| Summary: | The recent advances in technology have led to an exponential growth in the
number of digital images being stored on the Internet as well as in personal
computers. Effective methods to organize and index photos based on semantic
content have become essential to provide users with the convenience of searching
their albums for specific content without prior manual annotation. However,
querying for the image content is still a challenging task which has attracted much
research effort.
In this paper, we present a photo query framework based on prior annotation.
When the user provides a text query (e.g. “water”), the framework performs a
search within the annotation database and finds relevant photos. To accomplish
this goal, we built a set of classifiers to annotate user photos in advance, and used
these annotations for query. We leveraged the NUS-WIDE dataset, which contains
publicly available web images and their associated labels, to train the classifiers.
These classifiers are used to detect the presence of concepts in each photo in a
photo folder, and annotate the photos with suitable labels. To increase the accuracy
of the annotation process, we conducted experiments on two simple but effective
classification methods, k Nearest Neighbor (kNN) and Support Vector Machine
(SVM), and determine the best method by considering their accuracy and speed.
|
|---|