Biased Support Vector Machine for Relevance Feedback in Image Retrieval

1
Biased Support Vector Machine for Relevance
Feedback in Image Retrieval
Presentation in IJCNN 2004

• Hoi, Chu-Hong Steven
• Department of Computer Science Engineering
• The Chinese University of Hong Kong
• Shatin, Hong Kong
• Budapest, 25-29 July, 2004

2
Outline

• Background Motivation
• Biased Support Vector Machine (Biased SVM)
• Relevance Feedback by Biased SVM
• Experimental Results
• Conclusions

3
Background

• Challenges in Content-based Image Retrieval
  (CBIR)
• Semantic gap, low-level features, high-level
  concepts
• Subjectivity of human being,
• Relevance Feedback (RF)
• Refine retrieval results by incorporating users
  interactions
• A technique to narrow down the semantic gap,
  subjectivity
• Methods heuristic weighting Rui98, MARS99,
  optimization MindReader98, Rui00,
  classification MacArthur99, other learning
  techniques Huang01,
• Popular method proposed recently Support Vector
  Machines (SVM) Hong00, Chen01, Tong01, Zhang01

4
Motivation
Optimal separating hyperplane
margin
A case of regular SVM

• Imbalance dataset problem?
• negative gtgt positive
• Positive overwhelmed by negative

5
Motivation

• Limitation of regular SVMs for RF
• Regular binary SVM
• Simply treat as a strict binary classification
  problem
• without imbalance consideration
• Regular 1-SVM
• Exclude negative information
• Our solution Biased SVM
• A modified 1-SVM incorporating negative
  information with bias control

6
Biased SVM

• Problem formulation
• Training data
• The objective function

c
R
7
Biased SVM (cont.)

• Optimization by Lagrange multipliers
• Take the partial derivatives of L with respect to
  R,?,c, and?

8
Biased SVM (cont.)

• Dual problem (Quadratic Programming (QP) )
• The decision function

f (x)lt0
f (x)gt0
9
Relevance Feedback by Biased SVM

• One of differences with regular SVM
• Visual comparison

Biased SVM
Regular SVM
10
Relevance Feedback by Biased SVM

• Obtained decision function
• Simplified evaluation function

11
Experimental Results

• Datasets
• One synthetic dataset 40-Cat, each contains 100
  data points randomly generated by 7 Gaussian in a
  40-dimensional space.
• Two real-world image datasets selected from COREL
  image CDs
• 20-Cat 2,000 images
• 50-Cat 5,000 images

12
Experimental Results (cont.)

• Image Representation
• Color Moment
• 9-dimension
• Edge Direction Histogram
• 18-dimension
• Canny detector
• 18 bins, each of 20 degrees
• Wavelet-based texture
• 9-dimension
• Daubechies-4 wavelet, 3-level DWT
• 9 subimages to generate the feature

13
Experimental Results (cont.)

• Compared Schemes
• Relevance Feedback by regular nu-SVM
• Relevance Feedback with 1-SVM
• Relevance Feedback with Biased SVM
• Experimental Setup
• Metric Average precision relevant / returned
• Pick 10 instances, label pos. or neg.
• First iteration, 2 pos. and 8 neg.
• Same kernel and settings for compared schemes
• 200 relevance feedback simulation rounds are
  executed for each compare scheme.

14
Experimental Results (cont.)
Synthetic dataset
20-Cat COREL Images
15
Experimental Results (cont.)
50-Cat COREL Images
16
Experimental Results (cont.)
17
Conclusions

• Address the imbalance problem of relevance
  feedback in CBIR.
• Propose a modified SVM technique, i.e. Biased
  SVM, to attack the imbalance problem of relevance
  feedback problem in CBIR.
• Demonstrate effectiveness of the proposed scheme
  from experiments.

18
Budapest, Hungary, July, 2004
Thank you!
19
References

• Rui98 Y. Rui, T. S. Huang, M. Ortega, and S.
  Mehrotra, Relevance Feedback A Power Tool in
  Interactive Content-Based Image Retrieval, IEEE
  Tran Circuits and Systems for Video Technology,
  Vol 8 No 5, 1998, 644-655
• MARS99 K. Porkaew, S. Mehrotra, and M. Ortega,
  Query Reformulation for Content Based Multimedia
  Retrieval in MARS, IEEE Intl Conf. Multimedia
  Computing and Systems (ICMCS99), June, 1999
• MindReader98 Y. Ishikawa, R. Subramanya, and C.
  Faloutsos, MindReader Query databases through
  multiple examples, 24th VLDB Conf. (New York),
  1998
• Zhang01 L. Zhang, F. Lin, and B. Zhang,
  SUPPORT VECTOR MACHINE LEARNING FOR IMAGE
  RETRIEVAL, ICIP2001, 2001
• Rui00 Y. Rui, T. S. Huang, Optimizing learning
  in image retrieval, CVPR00, Hilton Head Island,
  SC, June 2000
• MacArthur99 S. MacArthur, C. Brodley, and C.
  Shyu, Relevance Feedback Decision Trees in
  Content-Based Image Retrieval, workshop CBAIVL,
  CVPR00, June 12, 2000.
• Tong01 S. Tong, and E. Chang, Support vector
  machine active learning for image retrieval, ACM
  MM2001, 2001
• Chen01 Y. Chen, X. S. Zhou, T. S. Huang,
  One-class SVM for Learning in Image Retrieval,
  ICIP'2001, Thessaloniki, Greece, October 7-10,
  2001
• Hong00 P. Hong, Q. Tian, T. S. Huang,
  "Incorporate Support Vector Machines to
  Content-Based Image Retrieval with Relevance
  Feedback", ICIP'2000, Vancouver, Sep 10-13, 2000.