ewimt 2005

1
AN ATTENTION BASED SIMILARITY MEASURE USED TO
IDENTIFY IMAGE CLUSTERS
Abstract This work outlines a new attention
based similarity measure. This measure
describes an application to the problem of
identifying image clusters in a four class
problem. A diverse set of images was obtained
using camera phones in four separate
locations. The classification performance was
tested against the true location of the images.
The approach promises to have application to
the unsupervised extraction of unknown numbers
of clusters in larger datasets.
Introduction The annotation of visual data for
subsequent retrieval is almost entirely carried
out through manual effort. This is slow, costly
and error prone and presents a barrier to the
stimulation of new multimedia services.
Similarity measures are central to most pattern
recognition problems not least in computer vision
and the problem of categorising and retrieving
large number of digital images. These problems
have motivated considerable research into content
based image retrieval . There are many approaches
to similarity and pattern matching and much of
this is covered in several survey papers .
Experimental Objectives
Automatic identification of the location of
mobile images. Implement a Cognitive visual
attention based classification method by visual
sub-clustering
Training and Classification
Cognitive Visual Attention Model
Experimental Results
Neighbourhood at location x matching at y
Conclusions This work has described an approach
to the identification of visual clusters within
an extremely diverse set of images in the context
of the identification of photo locations.
Matrices of inter-image similarities were
generated without the use of a priori selection
of features. It has been shown that the
selection of exemplars for classification is
sufficiently convergent to yield corresponding
results in a test set with corresponding clusters
being visually similar to those in the training
set.
Visual Sub-Cluster Extraction - 2
Visual Sub-Cluster Extraction - 1
Future Work A feasibility study will be carried
out to measure the effectiveness of unsupervised
clustering with application to much larger bodies
of data where the number of clusters is unknown
and the data is unlabelled. Some aspects of
the work will also include the management of home
photo collections and other image databases where
non-visual metadata is available. It is
envisioned that the technique may be applied in
video retrieval in application to Key frame
retrieval.
Some visual examples of extracted exemplars and
their visual sub-clusters from the 202 South
Hall location class (entire set).

• Acknowledgements
• The authors wish to acknowledge the support of
  Research and Venturing within British Telecom and
  colleagues at the University of California at
  Berkeley.
• The work also falls within the scope of the
  MUSCLE Network of Excellence in the European 6th
  Framework.

ADETOKUNBO BAMIDELE Engineering Doctorate
Programme in Communications (UCL) Department of
Electrical and Electronics Engineering Content
Understanding Group, Academic Supervisors
Professor Fred Stentiford, Dr. Lionel
Sacks, Industrial Supervisor Dr. Jason Morphett
(BT CTO Research and Venturing) http\\www.ee.ucl.
ac.uk\abamidel