Introduction - Multimedia Database Retrieval: Technology and Applications - page 8

Fig. 1.2 Mobile visual search via vocabulary coding through a wireless upstream query transmis-
sion pipeline. The scale invariance feature transform (SIFT) is applied to the captured landmark
for feature extraction and the resulting features are encoded to obtain a compact signature for the
recognition of the landmark image at the server end

visual recognition algorithm. Recognition requires discriminative BoW, to give a
more distinctive representation for local descriptors through discriminative learning
of image patches or saliency mapping.
Chapter 4 presents a soft-BoW method for mobile visual search based on
the discriminative learning of image patches. Specifically, a multi-touch screen
and the user's interaction on a mobile device are utilized by a user to select
regions-of-interest (ROIs) as prioritized information, and the surrounding context
is used as secondary information. Along with the BoW model, a context-embedded
vocabulary tree (CVT) for soft weighting is adopted by using both the ROI and its
surrounding image context to allow the mining of mobile visual intents. A system
is built upon an initial visual query input to obtain the recognition results. Once
the context metadata is associated with the intent, the system takes advantage of
more reliable contextual text and Global Positioning System (GPS) features in
searching and re-ranking. Ultimately, interesting and relevant social activities are
recommended to the users. The discriminative BoW presented in this work not
only enables mobile access to large multimedia repositories, but also provides more
effective user interaction.
Chapter 5 focuses on mobile visual search systems, which can identify land-
marks in a user's surroundings from the images captured by the camera on the
user's devices, and retrieve interesting information related to those landmarks. In
particular, saliency information is used with a re-ranking approach and incorporated
at various stages of recognition: saliency-aware local descriptor, saliency-aware
SVT, saliency-aware BoW, and discriminative learning via re-ranking. An important
novelty of this work is that, instead of relying on a plain structure of a compact
database of image signatures, the saliency information and re-ranking method are
used for increasing discriminating power and improving recognition accuracy.

1.3.4 Multimedia Retrieval in a Cloud Datacenter

In today's multimedia network systems, multimedia files are distributed over the
nodes in an overlay network as a cloud datacenter. Apparently, the searching of