Adetokunbo Bamidele

1
A Content-Based Image Retrieval System
Using Models of Human Vision.

• Adetokunbo Bamidele
• EngD. Research Engineer
• http//www.ee.ucl.ac.uk/abamidel
• a.bamidele_at_ee or s.bamidele_at_cs
• Department of Electrical and Electronics
  Engineering
• Supervisors
• Prof. Fred Stentiford
• Dr. Lionel Sacks
• Monday, 31st March 2003

2
ABSTRACT

• The world is in the midst of a content
  explosion.  Widespread digital imaging technology
  and the tumbling cost of processing and storage
  devices has seen huge growth in the acquisition
  and storage of content. 

From amateur photography to crime detection and
medical imaging, image based content, whether
static or dynamic is accounting for millions of
terabytes across the worlds databases. 

Current image based databases use predominantly
traditional and manual means to classify,
archive, search and retrieve data.
3
AIMS
Models of human visual attention and models of
the human recognition of visual similarities with
the aim of characterizing content in ways that
promote fast and effective search of large image
databases. 
Study database configurations that accommodate
new metadata attributes and their associated
functionality
1
4
2
At the services and applications layer and
possibly in other unexplored environments.
3
As a means to providing value added services for
present and future telecommunications generations
The research will investigate
4
RESEARCH SCOPE

• The research will survey

• The state of the art in Content-Based Image
  Retrieval.

• Conduct experiments to test the validity of
  existing models of human vision within the
  context of visual content retrieval.

5
Activities could include

• Investigating how models of human vision can
  improve retrieval systems

• Deriving new metadata models and integrating with
  an object-based database.

• Producing demonstrators for searching and
  retrieving content.

• Proposing new metadata standards to international
  bodies (e.g. MPEG-7/21)

• Extending existing models of human vision to
  handle wider ranges of content.

• Investigating video applications.

6
A Visual Attention Enhanced CBIR System
Illustrated

• Some of these models may provide solutions to
  problems of the machine interpretation of images
  and the intuitive access and retrieval of visual
  material.

E.g. performance may be enhanced by identifying
what is important in an image prior to similarity
calculations and retrieval
Query By Example (QBE) for a red car
Query restricted to regions deserving attention.
7
Continued
QBE red cars and trees
most interesting regions identified on red cars
Region in vicinity of sun is most relevant to
retrieval.
QBE for sunrise
Results based on an existing VA model colour
spaceRGB
8
APPLICATION REQUIREMENTS

• Image/video retrieval must be immediate and
  relevant especially in a mobile environment.

• High perceived visual quality.

• Manual involvement must be minimal and need no
  special skills

• Intuitive interface.