The Bayesian Image Retrieval System,PicHunter

1
The Bayesian Image Retrieval System,PicHunter

• Theory, Implementation, and Psychophysical
  Experiments

2
Introduction

• Relevance feedback
• users give additional information
• Main idea
• With an explicit model of a users actions,
  assuming a desired goal, PicHunter uses Bayes
  rule to predict the goal image, given their
  actions

3
Nature of search

• Target-specific search (Target search)
• exact match
• Category search
• same category is ok
• Open-ended search (browsing)

4
Bayes formula

• Fj hypothesis (Target image is j)
• E experiment (users response behavior)
• Show us how the correctness of a hypothesis
  change after carrying out an experiment
• How to model P(EFj)?

5
Theoretical basis for PicHunter

• During each session
• a set Dt of ND images, Action At
• H t ----History of the session

6
User ModelAssessing Image similarity

• Key term
• P(AtTTi,Dt,U)
• Uspecific user
• Purposeupdate the probability of each Ti being
  target

7
Relevance feedback

• e.g. 2AFC (two-alternative forced-choice)
• Given two image, user need to choose which one is
  similar to target
• P(EFj) ? P(A1X1,X2,TTi)
• 1 if d(X1,Ti) lt d(X2,Ti)
• 0.5 if d(X1,Ti) d(X2,Ti)
• 0 d(X1,Ti) gt d(X2,Ti)
• Another one is relative distance

8
Relative distance measure

• using the pictorial features distance
• as the form of the probability
• When ND2, At1 or 2
• Psigmoid(A1X1,X2,T)

9
Pictorial features

• HSV-HIST
• Hue, Saturation, Value histogram
• HSV-CORR
• RGB-CCV
• Color histogram

10
Display Updating Model

• Most-Probable Display Updating Model
• Give the most similar one for user to choose
• Most-informative Display Updating Model
• CP(T)
• Give both similar and dissimilar images for use
  to choose

11
Results

• Cox formulated an experiment XYZ
• X - with memory or with out
• Use all the response or just response in one
  iteration
• Y - with using relative / absolute distance
  measure
• Z use pictorial or semantic measure
• Benchmark - how many images need to be displayed
  before target is found
• MRS is the best
• With memory, use relative distance and semantic
  measure