Diapositive 1

1
A Conceptual Approach to Web Image Retrieval
Adrian Popescu Gregory Grefenstette CEA LIST,
France
OntoImage 2008 Marrakech, Morocco
Contacts adrian.popescu_at_telecom-bretagne.eu
gregory.grefenstette_at_cea.fr
2
Some facts

• 60 of Internet image queries include only one
  concept
  
• People are stupid or lazy?
• What are they looking for?
• Common nouns
• Places
• Other (famous) people
• Products
• Sexual content

3
Our Image Browsing Strategy

• Introduce conceptual hierarchy into image search
  applications
  
• Expand common words by more specific terms
• Present results reflecting hierarchy
• Dual browsing image or text query
• Exploit conceptual hierarchy and image spidering
  to structure image base

4
Architecture of OLIVE
Olive picture retrieval framework employing a l
arge scale conceptual hierarchy
to provide a dual access to Web images
5
Adapting WordNet

• Concept ranking
• Based on Web counts
• Relevance (C1) webCount(C1, parent)
• Web counts and WordNet structure
• Relevance (C1, C2) webCount(C1, parent)
  dist(C1,C2)/senses(C1)
  
• Example for dog
• (Web counts) pooch, pug, Newfoundland, basset
• (WebWordNet) collie, basset, german shepherd,
  doberman
  
• Query reformulation
• Query concept representative leaf nodes

6
Knowledge for dog
7
First response page for dog - Google Image
8
First response page for dog - Olive
9
First response page for Doberman - Olive
10
Visual similarity search for image of Doberman
11
Evaluation Precision on first two pages
12
Precision over ten pages
13
Qualitative evaluation

• Experiment setting
• 10 users compared Olive and Google Image
• Questionnaire about results structuring
  enriched interaction automatic reformulation
  
• Positive reactions from the users to the new
  features introduced in Olive
  
• The users also provided useful comments under the
  form of free text.

14
Conclusions

• A semantic retrieval architecture can cover a
  significant part of Web picture queries.
  
• Dual access to picture content, textual and
  visual
  
• Simple but efficient solution to introduce CBIR
  in large scale architectures
  
• With semantic structures, existing image indexes
  can be better used than in current applications
  
• OLIVE outperforms Google Image, qualitatively and
  quantitatively
  
• Future work
• The extension of the hierarchy (e.g. using
  Wikipedia)?

15
A glimpse of the future
16
Advertising

• Website
• http//moromete.net
• I am looking for work
• Starting November 2008