Relevance Feedback

1
Relevance Feedback

• A relevance feedback mechanism for content-based
  image retrieval
• G. Ciocca, R. Schettini 1999

2
Weight

• Each query is different
• Some interest in color
• Some interest in shape
• Some interest in a specific region
• Associate each feature with a weight

3
Image segmentation

• User may interest only in a particular region
• Divide image into different region
• Advantage we can capture region of interest
• Disadvantage the feature vector is longer

4
Query Modification

• The query image is not the target image user want
• We modify the query base on the feedback

5
Weight updating thro feedback

• How to estimate users interest from feedback
• Base on the relevant images
• If color of relevant images varies, the user
  might not interest in color
• If feature vector of upper right corner are
  close, user might interest in that area

6
Overall Process
A heuristic used in the 1st iteration. If user
makes a query that is similar to previous query,
the weight vector will be used. Reduce time need
to adapt the similarity measure