Relevance Feedback: New Trends - PowerPoint PPT Presentation

Loading...

PPT – Relevance Feedback: New Trends PowerPoint presentation | free to download - id: 75ca8e-MWIwN



Loading


The Adobe Flash plugin is needed to view this content

Get the plugin now

View by Category
About This Presentation
Title:

Relevance Feedback: New Trends

Description:

Relevance Feedback: New Trends Derive global optimization methods: More computationally robust Consider the correlation between different attributes – PowerPoint PPT presentation

Number of Views:111
Avg rating:3.0/5.0
Slides: 11
Provided by: uad3
Learn more at: http://www.cs.fiu.edu
Category:

less

Write a Comment
User Comments (0)
Transcript and Presenter's Notes

Title: Relevance Feedback: New Trends


1
Relevance Feedback New Trends
  • Derive global optimization methods
  • More computationally robust
  • Consider the correlation between different
    attributes
  • Incorporate semantic information
  • Text, key words
  • Other techniques
  • Query expansion
  • Store feedback information

2
New Trends Global Optimization
  • The key point is to derive a general distance
    function.
  • Traditional Way using traditional ellipses
    distance function aligned with the coordinate
    axis.
  • Global Optimization using general ellipses
    distance function that is not necessarily aligned
    with the coordinate axis.
  • Therefore, it allows for correlations between
    attributes in addition to different weight on
    each component.

3
New Trends Global Optimization--Traditional
Way MARS1997
  • Incompleteness of MARS1997
  • Directly applied Rocchios formula and used some
    heuristics (such as ß and ? in Rocchios
    formula).
  • The similarity values are in fact the weighted
    Euclidean distance function which has ellipses
    whose major axis must be aligned with the
    coordinate axis.
  • The weight adjustment based on standard deviation
    method cannot be proved optimal.

4
New Trends Global Optimization--More Robust
Method MindReader
  • Robustness of MindReader
  • Directly go for an optimal solution for minimum
    problem in hidden distance without using some
    heuristics.
  • Allow not only for different weights of each
    attribute, but also for correlations between
    attributes.
  • In the general ellipses distance function, the
    major axis are not necessarily aligned with the
    coordinate axis.

5
New Trends Global Optimization--More Robust
Method MindReader
  • Different distance functions

6
New Trends Global Optimization--More Robust
Method MARS1999
  • Improvement in MARS1999 combine the two
    best-known techniques of RF (MindReader and MARS)
    to overcome the shortcomings that each technique
    has.
  • Takes into account the multi-level image model.
  • Derives the optimal solutions for both the query
    vectors and the weights associated with the
    multi-level image model.

7
New Trends Combine Semantic Info.
  • Problems in previous systems
  • Only perform feedback at the low-level feature
    vector level, and fail to take into account the
    actual semantics information for the images.
  • Solution
  • Embed semantic information into the low-level
    feature based image retrieval system.

8
New Trends Combine Semantic Info. --a
framework Ifind
  • The basic idea construct a semantic network and
    integrated it with low-level feature vector based
    relevance feedback by using a modified form of
    the Rocchios formula

9
Other New Trends
  • Query Expansion
  • In each iteration of feedback, the relevant
    objects are added to the query and non-relevant
    ones are removed.
  • Store feedback information
  • Store the outcome of a feedback process when the
    process is terminated. The stored parameters can
    be used to predict the parameter settings for
    similar queries by interpolation.

10
Conclusion
  • Relevance Feedback is an excellent technique for
    improving the retrieval effectiveness.
  • Problems with RF
  • Low-level features cannot sufficiently represent
    the content of image.
  • To embed semantic information into the systems
    still has trouble dealing with the issues of
    weight normalization, thresholds selection and
    scalability.
  • No state-of-art systems can provide the object
    level queries. It is also a big research issue in
    content-based retrieval society.
About PowerShow.com