Content-Based Image Retrieval

Charlie Dagli and Thomas S. Huang

{dagli,huang}@ifp.uiuc.edu

1
Outline
 What is CBIR ?
 Image Features
 Feature Weighting and Relevance
Feedback
 User Interface and Visualization

2
What is Content-based Image
Retrieval (CBIR)?
 Image Search Systems that search
for images by image content
<-> Keyword-based Image/Video Retrieval
(ex. Google Image Search, YouTube)

3
Applications of CBIR
 Consumer Digital Photo Albums

Digital Cameras

Flickr
 Medical Images
 Digital Museum
 Trademarks Search
 MPEG-7 Content Descriptors

4
Basic Components of CBIR
 Feature Extractor

Create the metadata
 Query Engine

Calculate similarity
 User Interface

5
 How does CBIR work ?
 Extract Features from Images
 Let the user do Query
 Query by Sketch
 Query by Keywords
 Query by Example
 Refine the result by Relevance Feedback

Give feedback to the previous result

6
 Query by Example
 Pick example images, then ask the system
to retrieve “similar” images.

“Get similar images”

CBIR

Query Sample

Results

7
 Relevance Feedback
 User gives a feedback to the query results
 System recalculates feature weights

Query Feedback Feedback

Initial
sample 1st Result 2nd Result

8
Basic Components of CBIR
 Feature Extractor

Create the metadata
 Query Engine

Calculate similarity
 User Interface

9
Image Features (Metadata)
 Color
 Texture
 Structure
 etc

10
Color Features
 Which Color Space?

RGB, CMY, YCrCb, CIE, YIQ, HLS, …
 Our Favorite is HSV

Designed to be similar to human perception

11
 HSV Color Space
 H (Hue)
 Dominant color (spectral)
 S (Saturation)
 Amount of white
 V (Value)
 Brightness

How to Use This?

12
Straightforward way to use
HSV as color features
 Histogram for each H, S, and V
 Then compare in each bin
 Is this a good idea?

13
Are these two that different?
 Histogram comparison is very
sensitive

14
 Color Moments [Stricker ‘95]
 For each image, the color distribution in
each of H, S and V is calculated
 1st (mean), 2nd (var) and 3rd moment for HSV

N
1
Ei = ∑
N j=1
pij

N
i : color channel {i=h,s,v}
1
σi = ( ∑
N j=1
(pij − Ei ) 2 )1 2 N = # of pixels in image
1 N
Total 9 features
si = ( ∑ (pij − Ei ) 3)1 3
N j=1

15
 Shape Features
 Region-Based Shape

Outer Boundary
 Contour-Based Shape

Features of Contour
 Edge-Based Shape

Ex. Histogram of edge length and
orientation

16
Region-based vs. Contour-based
 Region-based

Suitable for Complex objects with
disjoint region
 Contour-based

Suitable when semantics are contained in
the contour

17
Region-based vs. Contour-based
Region Complexity

Contour
Complexity

18
Region-based vs. Contour-based

Good Examples for Region-based shape

Good Examples for Contour-based shape

19
 Angular Radial Transformation (ART)
[Kim’99]
 A Region-based shape
 Calculate the coefficients based on image
intensities in polar coordinates (n<3, m<12)
2π 1
Fnm = ∫ 0 ∫ o
Vnm (ρ , θ ) f (ρ ,θ ) ρdρdθ
f (ρ,θ )L image intensity in polar coordinates
Vnm (ρ,θ )L ART basis function
Vnm (ρ,θ ) = 1 / 2 π exp( jmθ )Rn (ρ)
€  1 n=0
Rn (ρ) = 
 2 cos(πnρ) n ≠ 0
Total 35 coefficients in 140 bits (4 bits/coeff)

€ 20
Curvature Scale-Space (CSS)
[Mokhtaarian ‘92]

 A contour-based shape
1) Apply lowpass filter repeatedly until
concave contours smoothed out
2) “How contours are filtered” becomes
the features
• Zero crossing in the curvature functions
after each application of the lowpass
filter
• CSS Image

21
 CSS Image
Tracks zero-crossing locations of each concavity in the contour
Contour Curvature

B
A
C
D
F
3 ITR
E

29 ITR
B
A

E
F

100
ITR

22
S
CSS Features
 # of peaks in CSS images
 Highest peak
 Circularity (perimeter2/ area)
 Eccentricity
 Etc.

23
Texture Features
 Wavelet-based Texture Features
[Smith’94]

24
Wavelet Filter Bank
Coarse Info (low
freq)

Wavelet
Filter
Original Image
Detail (high freq)
25
Texture Features from Wavelet

Original Wavelet Feature

Image Image Extraction
f6,v6 f3,v3

Wavelet f5,v5 f4,v4

f1,v1
Filter
Bank f2,v2 f0,vo

Take the mean

Decompose the
and variance of
images into
each subband
frequency subbands

26
Other approaches: Region-Based
 Global features often times fail to
capture local content in an image
GLOBAL DESCRIPTION
{Green, Grassy, Hillside}
color, texture, shape

No sheep? No fence? No houses?

27
Other approaches: Region-Based
 Segmentation-Based

Images are segmented by color/texture similarities:
Blobworld [Carson ‘99], Netra [Ma and Manjunath ‘99]
 Grid-Based

Images are partitioned, features are calculated from
blocks: [Tian ‘00],[Moghaddam ‘99]

28
Other approaches: Region-Based
 Combine Grid and Segmentation: [Dagli and Huang, ‘04]

29
Basic Components of CBIR
 Feature Extractor

Create the metadata
 Query Engine

Calculate similarity
 User Interface

30
Now, We have many features
(too many?)

 How to express visual “similarity”

with these features?

31
Visual Similarity ?
 “Similarity” is Subjective and Context-
dependent.
 “Similarity” is High-level Concept.
 Cars, Flowers, …
 But, our features are Low-level features.

Semantic Gap!

32
Which features are most important?
 Not all features are always important.
 “Similarity” measure is always changing
 The system has to weight features on the fly.
How ?

33
 Online Feature Weighting
 Approach #1 - Manual
 Ask the user to specify number
“35% of color and 50% of texture…”
 Very difficult to determine the numbers
 Approach #2 - Automatic
 Learn feature weights from examples
 Relevance Feedback

34
Online Feature Weighting
 From Query Examples, the system
determines feature weighting matrix W

CBIR
Calculate W

Query
Result
r r r rT r r
distance( x, y ) = ( x − y ) W ( x − y )
35
 How to Calculate W ?
 No Negative Examples (1-class)
 Positive and Negative Examples (2-class)
 One Positive and Many Negative classes
(1+x)-class
 Many Positive and Many Negative classes
(x+y)-class

36
When there are only relevant
images available…
 We want to give more weights to
common features among example
images.
 Use the variance.

Features with low variance
-> Common features
-> Give higher weight

37
 One Class Relevance Feedback in
MARS [Rui ‘98]
 Calculates the Variance among relevant examples.
 The inverse of variance becomes the weight of each
feature.
 This means “common features” between positive
examples have larger weights.

1/σ12 0 
 2 
 1/ σ 2 
W = 2
1/σ 3 
  W is a k x k diagonal matrix
 O 
2
 0 1/σ k 

38
 Relevance Feedback as Two-Class
Problem (positive and negative)
Fisherʼs Discriminant Analysis (FDA)

 Find a W that …
 minimizes the scatter
of each class cluster
positive (within scatter)
 maximizes the scatter
negative between the clusters
(between scatter)

39
Two-Class problem
 Target function W T SBW
W = argmax T
 W is full matrix W W SW W
SB LBetween Scatter Matrix
SW LWithin Scatter Matrix
2
SW = ∑ ∑ j ∈group # i (x j − m i )(x j − m i ) T
i=1

€ SB = (m1 − m2 )(m1 − m2 )T
m1,m2 Lmean of each class

40
 Solution
 The problem is reduced to
generalized eigenvalue problem

SB w i = λi SW w i
1/ 2
W = ΦΛ
ΛLdiagonal matrix of eigenvalues
ΦLeigenvectors

€ 41
From Two-class to (1+x)-class
 Positive examples are usually from
one class such as flower
 Negative examples can be from any
classes such as “car”, “elephant”,
“orange”…
 It is not desirable to assume negative
images as one class.

42
 RF as (1+x)-Class Problem
• Biased Discriminant Analysis [Zhou et al. ‘01]
• Negative examples can be any images
• Each negative image has its own group
positive
negative

SW = ∑ (x − m)(x − m)T
x ∈ positive

SB = ∑ (x − m)(x − m)T
x ∈negative

mLmean of positive class

The solution is similar to FDA
43
RF as (x+y)-Class Problem
 Group BDA [Nakazato, Dagli ‘03]
 Multiple Positive classes
 Scattered Negative classes

positive

negative

44
RF as (x+y)-Class Problem
 Group BDA [Nakazato, Dagli ‘03]
 Multiple Positive classes
 Scattered Negative classes
positive
negative
SW = ∑ i ∑ x ∈i (x − mi )(x − mi )T
SB = ∑ ∑ (y − mi )(y − mi )T
i y ∈negative

mi Lmean of positive class i

45
Basic Components of CBIR
 Feature Extractor

Create the metadata
 Query Engine

Calculate similarity
 User Interface

46
User Interface and Visualization
 Basic GUI
 Direct Manipulation GUI

El Nino [UC San Diego]

Image Grouper [Nakazato and Huang]
 3D Virtual Reality Display

47
Traditional GUI for Relevance
Feedback
 User selects
relevant images
 If good images
are found, add
them
 When no more
images to add, the
search converges
Slider or Checkbox

48
ImageGrouper [Nakazato and Huang]
 Query by Groups

Make a query by creating groups of images

Easier to try different combinations of
query sets (trial-and-Error Query)

49
ImageGrouper

Result View Pallete Panel

50
 Note
 Trial-and-Error Query is very important
because
 Image similarity is subjective and context-
dependent.
 In addition, we are using low-level image features.
(semantic gap)
 Thus, it is VERY difficult to express the user’s
concept by these features.

51
Image Retrieval in 3D
 Image retrieval and browsing in 3D
Virtual Reality
 The user can see more images without
occlusion
 Query results can be displayed in
various criteria

Results by Color features, by texture, by
combination of color and texture

52
3D MARS

Texture

Structure

color
Initial Display
Result

53
 3D MARS in CAVE™

 Shuttered glasses for

immersive 3D
experience
 Click and Drag images
by WAND
 Fly-through by
Joystick

54
 Demos
 Traditional GUI

IBM QBIC
• http://wwwqbic.almaden.ibm.com/

UIUC MARS
• http://chopin.ifp.uiuc.edu:8080
 ImageGrouper

http://www.ifp.uiuc.edu/~nakazato/grouper

55
References (Image Features)
 Bober, M., “MPEG-7 Visual Descriptors,” In IEEE Transactions on Circuits and
Systems for Video Technology, Vol. 11, No. 6, June 2001.
 Stricker, M. and Orengo, M., “Similarity of Color Images,” In Proceedings of SPIE,
Vol. 2420 (Storage and Retrieval of Image and Video Databases III), SPIE Press,
Feb. 1995.
 Zhou, X. S. and Huang, T. S., “Edge-based structural feature for content-base
image retrieval,” Pattern Recognition Letters, Special issue on Image and Video
Indexing, 2000.
 Smith, J. R. and Chang S-F. Transform features for texture classification and
discrimination in large image databases. In Proceedings of IEEE Intl. Conf. on
Image Processing, 1994.
 Smith J. R. and Chang S-F. “Quad-Tree Segmentation for Texture-based Image
Query.” In Proceedings of ACM 2nd International Conference on Multimedia, 1994.
 Dagli, C. K. and Huang, T.S., “A Framework for Color Grid-Based Image Retrieval,”
In Proceedings of International Conference on Pattern Recognition, 2004.
 Tian, Q. et. al. “Combine user defined region-of-interest and spatial layout in image
retrieval,” in IEEE Intl. Conf. on Image Processing, 2000.
 Moghaddam B. et. al. “Defining image content with multiple regions-of-interest,” in
IEEE Wrkshp on Content-Based Access of Image and Video Libraries, 1999.

56
 References (Relevance Feedback)
 Rui, Y., et al., “Relevance Feedback: A Power Tool for
Interactive Content-Based Image Retrieval,” In IEEE
Trans. on Circuits and Video Technology, Vol.8, No.5,
Sept. 1998
 Zhou, X. S., Petrovic, N. and Huang, T. S. “Comparing
Discriminating Transformations and SVM for Learning
during Multimedia Retrieval.” In Proceedings of ACM
Multimedia ‘01, 2001.
 Ishikawa, Y., Subrammanya, R. and Faloutsos, C.,
“MindReader: Query database through multiple
examples,” In Proceedings of the 24th VLDB
Conference, 1998.

57
 References (User Interfaces and
Visualizations)
 Nakazato, M. and Huang, T. S. “3D MARS: Immersive Virtual
Reality for Content-Based Image Retrieval.“ In Proceedings of
2001 IEEE International Conference on Multimedia and Expo
(ICME2001), Tokyo, August 22-25, 2001
 Nakazato, M., Manola, L. and Huang, T.S., “ImageGrouper: Search,
Annotate and Organize Images by Groups,” In Proc. of 5th Intl.
Conf. On Visual Information Systems (VIS’02), 2002.
 Nakazato, M., Dagli C.K., and Huang T.S., “Evaluating Group-Based
Relevance Feedback for Content-Based Image Retrieval,” In
Proceedings of International Conference on Image Processing,
2003.
 Santini, S. and Jain, R., “Integrated Browsing and Querying for
Image Database,” IEEE Multimedia, Vol. 7, No. 3, 2000, page
26-39.

58
Similar Region Shape, Different
Contour

59