ICT role in 21st century education and it's challenges.
Image re ranking system
1. An Approach to Re-Rank Retrieved
Images - Image Search Results with
Visual Similarity
Presented by Guided by
Veningston .K Mr. M. Newlin Rajkumar
M.E student, Dept of CSE, Lecturer, Dept of CSE,
Anna University – Coimbatore, Anna University – Coimbatore,
India. India.
2. Objective
To address a ranking problem in web image
retrieval
System to re-rank images returned by image
search engine
Re-ranking images by incorporating,
visual aspects
visual similarity
04/17/12 Department of Computer Science and Engineering 2
3. Introduction
Image Retrieval System
Comparative study on text & image based search
Interest point extraction - visual content of images
Re-ranking the results of text based systems
using visual information
Finding the largest set of most similar images
Rearranging images based on the similarity
04/17/12 Department of Computer Science and Engineering 3
4. Why image Re-ranking?
To maximize relevancy of image results
To achieve diversity of image results
04/17/12 Department of Computer Science and Engineering 4
5. Proposed Scheme
Goal
Retrieve image results that are relevant
Finding common features among images
Overview
Interest points on the images are extracted
Similarity of each pair of images are computed
Generate graph model
Apply page ranking
04/17/12 Department of Computer Science and Engineering 5
6. Finding common features among
images
Similarity measurement to handle potential rotation,
scale and perspective transformations.
04/17/12 Department of Computer Science and Engineering 6
7. Scale Invariant Feature Transform
Image matching and features to similarity
04/17/12 Department of Computer Science and Engineering 7
8. Image similarity
Given two images u and v,
Corresponding descriptor vector, Du = (d1u,
d2u, ...dmu ) and Dv = (d1v, d2v, ...dnv ),
Define the similarity between two images
simply as the number of interest points
shared between two images divided by their
average number of interest points.
04/17/12 Department of Computer Science and Engineering 8
9. Graph model
Given the visual similarities of the images to
be ranked.
Treat images as web documents
Treat similarities as visual hyperlinks
Estimate the probability of images being visited by
a user // using page rank
Images with more estimated visits are ranked
higher
04/17/12 Department of Computer Science and Engineering 9
10. Similarity to Centrality
Centrality – importance of images
Given a graph with vertices and a set of
weighted edges, define and measure the
“importance” of each of the vertices
Vertices = images
Egde weights = similarity
A vertex closer to an important vertex should
rank higher than others
04/17/12 Department of Computer Science and Engineering 10
11. Similarity to Centrality
Ranking scores correspond to probability of
arriving in each vertex by traversing through
the graph
Matrix constructed from the
weights of the edges in the
Adjacency matrix for graph
unweighted graph Decision to take a
particular path defined by
weighted edges
04/17/12 Department of Computer Science and Engineering 11
12. Page Ranking (PR)
A B
Determines the order of E
search results
Method of measuring a C D
page’s importance
Results are based on Search result Priority
this priority order B 3
A 2
C 1
D 1
E 1
04/17/12 Department of Computer Science and Engineering 12
13. Overall scheme
Computationally infeasible to compute
similarities for all images indexed by search
engine
Pre-cluster web images based on metadata
Define the similarity of images
Given a query, extract top - N results
returned, create graph of visual similarity on
the N images
Compute image rank only on this subset
04/17/12 Department of Computer Science and Engineering 13
14. Performance metrics
Precision and recall
number of relevant images in the returned images
Recall = ----------------------------------------------------------------------
total number of relevant images in the database
number of relevant images in the returned images
Precision = ---------------------------------------------------------------------
total number of returned images
Result analysis
Screen shots
04/17/12 Department of Computer Science and Engineering 14
15. Merits
Minimizing irrelevant images
Selecting small set of images
Computational cost
04/17/12 Department of Computer Science and Engineering 15
16. Conclusion
Ability to reduce the number of irrelevant
images
A tiny set of important images can be
selected from a very large set of candidates
04/17/12 Department of Computer Science and Engineering 16
17. Future work
Extensions of this technique to a query driven
feature selection.
04/17/12 Department of Computer Science and Engineering 17
18. References
[1] Yushi Jing, Shumeet Baluja. VisualRank: Applying PageRank to
Large-Scale Image Search. IEEE Transaction on Pattern Analysis and
Machine Intelligence, Vol 30, No. 11:1887–1890, 2008.
[2] R. Datta, D. Joshi, J. Li, and J. Wang. Image Retrieval: Ideas,
Influences, and Trends of the New Age. ACM Computing Surveys,
vol. 40, no. 2, 2008.
[3] Yushi Jing, Shumeet Baluja. PageRank for Product Image Search.
WWW 2008, April 21–25, 2008, Beijing, China. ACM International conf.
2008.
[4] K. Mikolajczyk and C. Schmid. A performance evaluation of local
descriptors. IEEE Transaction on Pattern Analysis and Machine
Intelligence, 27(10):1615–1630, 2005.
[5] S. Belongie, J. Malik, and J. Puzicha. Shape matching and object
recognition using shape contexts. IEEE Transactions on Pattern
Analysis and Machine Intelligence (PAMI), 24(24):509–522, 2002.
04/17/12 Department of Computer Science and Engineering 18
19. Thank you for your
Attention
04/17/12 Department of Computer Science and Engineering 19