Feature Selection for Image Retrieval

1
Feature Selection for Image Retrieval

• By Karina Zapién Arreola
• January 21th, 2005

2
Introduction

• Variable and feature selection have become the
  focus of much research in areas of applications
  for datasets with many variables are available
• Text processing
• Gene expression
• Combinatorial chemistry

3
Motivation

• The objective of feature selection is three-fold
• Improving the prediction performance of the
  predictors
• Providing a faster and more cost-effective
  predictors
• Providing a better understanding of the
  underlying process that generated the data

4
Why use feature selection in CBIR

• Different users may need different features for
  image retrieval
• From each selected sample, a specific feature set
  can be chosen

5
Boosting

• Method for improving the accuracy of any learning
  algorithm
• Use of weak algorithms for single rules
• Weighting of the weak algorithms
• Combination of weak rules into a strong learning
  algorithm

6
Adaboost Algorithm

• Is a iterative boosting algorithm
• Notation
• Samples (x1,y1),,(xm,ym), where, yi -1,1
• There are m positive samples, and l negative
  samples
• Weak classifiers hi
• For iteration t, the error is defined as
• et min (½)Si ?i hi(xi) yi
• where ?i is a weight for xi.

7
Adaboost Algorithm

• Given samples (x1,y1),,(xm,ym), where yi -1,1
• Initialize ?1,i1/(2m), 1/(2l), for yi 1,-1
• For t1,,T
• Normalize ?t,i ?t,i /(Sj ?t,j)
• Train base learner ht,i using distribution ?i,j
• Choose ht that minimize et with error ei
• Update ?t1,i ?t,i ßt1-ei
• Set ßt (et)/(1- et) and at log(1/ ßt)
• Output the final classifier H(x) sign( St at
  ht(x) )

8
Adaboost Application

• Searching similar groups
• A particular image class is chosen
• A positive sample of this group is given randomly
• A negative sample of the rest of the images is
  given randomly

9
Check list Feature Selection

• Domain knowledge
• Commensurate features
• Interdependence of features
• Prune of input variables
• Asses features individually
• Dirty data
• Predictor linear predictor
• Comparison
• Stable solution

10
Domain knowledge

• Features used
• colordb_sumRGB_entropy_d1
• col_gpd_hsv
• col_gpd_lab
• col_gpd_rgb
• col_hu_hsv2
• col_hu_lab2
• col_hu_lab
• col_hu_rgb2
• col_hu_rgb
• col_hu_seg2_hsv
• col_hu_seg2_lab
• col_hu_seg2_rgb

• Features used
• col_hu_seg_hsv
• col_hu_seg_lab
• col_hu_seg_rgb
• col_hu_yiq
• col_ngcm_rgb
• col_sm_hsv
• col_sm_lab
• col_sm_rgb
• col_sm_yiq
• text_gabor
• text_tamura
• edgeDB
• waveletDB

• Features used
• hist_phc_hsv
• hist_phc_rgb
• Hist_Grad_RGB
• haar_RGB
• haar_HSV
• haar_rgb
• haar_hmmd

11
Check list Feature Selection

• Domain knowledge
• Commensurate features
• Normalize features between an appropriated range
• Adaboost takes each feature independent so it is
  not necessary to normalize them

12
Check list Feature Selection

• Domain knowledge
• Commensurate features
• Interdependence of features
• Prune of input variables
• Asses features individually
• Dirty data
• Predictor linear predictor
• Comparison
• Stable solution

13
Feature construction and space dimensionality
reduction

• Clustering
• Correlation coefficient
• Supervised feature selection
• Filters

14
Check list Feature Selection

• Domain knowledge
• Commensurate features
• Interdependence of features
• Prune of input variables
• Features with the same value for all samples
  (variance0) were eliminated

• From
• Linear Features
• 3583 were selected

15
Check list Feature Selection

• Domain knowledge
• Commensurate features
• Interdependence of features
• Prune of input variables
• Asses features individually
• When there is no asses method, use Variable
  Ranking method. In Adaboost this is not necessary

16
Variable Ranking

• Preprocessing step
• Independent of the choice of the predictor
• Correlation criteria
• It can only detect linear dependencies
• Single variable classifiers

17
Variable Ranking

• Noise reduction and better classification may be
  obtained by adding variables that are presumable
  redundant
• Perfectly correlated variables are truly
  redundant in the sense that no additional
  information is gained by adding them. It doesnt
  mean absence of variable complementarily
• Two variables that are useless by themselves can
  be useful together

18
Check list Feature Selection

• Domain knowledge
• Commensurate features
• Interdependence of features
• Prune of input variables
• Asses features individually
• Dirty data
• Predictor linear predictor
• Comparison
• Stable solution

19
Check list Feature Selection

• Domain knowledge
• Commensurate features
• Interdependence of features
• Prune of input variables
• Asses features individually
• Dirty data
• Predictor linear predictor
• Comparison
• Stable solution

20
Adaboost Algorithm

• Given samples (x1,y1),,(xm,ym), where xi, yi
  -1,1
• Initialize ?1,i1/(2m), 1/(2l), for yi -1,1
• For t1,,T
• Normalize ?t,i ?t,i /(Sj ?t,j)
• Train base learner ht,i using distribution ?i,j
• Choose ht that minimize et with error ei
• Update ?t1,i ?t,i ßt1-ei
• Set ßt (et)/(1- et) and at log(1/ ßt)
• Output the final classifier H(x) sign( St at
  ht(x) )

21
Weak classifier

• Each weak classifier hi is defined as follows
• hi.pos_mean mean value for positive samples
• hi.neg_mean mean value for negative sample
• A sample is classified as
• 1 if it is closer to hi.pos_mean
• -1 if it is closer to hi.neg_mean

22
Weak classifier

• hi.pos_mean mean value for positive samples
• hi.neg_mean mean value for negative sample
• A Linear Classifier was used

23
Check list Feature Selection

• Domain knowledge
• Commensurate features
• Interdependence of features
• Prune of input variables
• Asses features individually
• Dirty data
• Predictor linear predictor
• Comparison
• Stable solution

24
Adaboost experiments and results
10 positives
4 positives
25
Few positive samples
Use of 4 positive samples
26
More positive samples
Use of 10 positive samples
False Positive
27
Training data
Use of 10 positive samples
28
Changing number of Training Iterations
The number of iterations Used was from 5 to 50
Iterations 30 was set
29
Changing Sample Size
30
Few negative samples
Use of 15 negative samples
31
More negative samples
Use of 75 negative samples
32
Check list Feature Selection

• Domain knowledge
• Commensurate features
• Interdependence of features
• Prune of input variables
• Asses features individually
• Dirty data
• Predictor linear predictor
• Comparison (ideas, time, comp. resources,
  examples)
• Stable solution

33
Stable solution

• For Adaboost is important to have a
  representative sample
• Chosen parameters
• Positives samples 15
• Negative samples 100
• Iteration number 30

34
Stable solution with more samples and iterations
Dinosaurs
Roses
Buses
Horses
Elephants
Buildings
Food
Humans
Mountains
Beaches
35
Stable solution for Dinosaurs

• Use of
• 15 Positive samples
• 100 Negative samples
• 30 Iterations

36
Stable solution for Roses

• Use of
• 15 Positive samples
• 100 Negative samples
• 30 Iterations

37
Stable solution for Buses

• Use of
• 15 Positive samples
• 100 Negative samples
• 30 Iterations

38
Stable solution for Beaches

• Use of
• 15 Positive samples
• 100 Negative samples
• 30 Iterations

39
Stable solution for Food

• Use of
• 15 Positive samples
• 100 Negative samples
• 30 Iterations

40
Unstable Solution
41
Unstable solution for Roses

• Use of
• 5 Positive samples
• 10 Negative samples
• 30 Iterations

42
Best features for classification

• Humans
• Beaches
• Buildings
• Buses
• Dinosaurs
• Elephants
• Roses
• Horses
• Mountains
• Food

43

• And the winner is

44
Feature frequency
45
Extensions

• Searching similar images
• Pairs of images are built
• The difference for each feature is calculated
• Each difference is classified as
• 1 if both images belong to the same class
• -1 if both images belong to different classes
• Multiclass adaboost

46
Extensions

• Use of another weak classifier
• Design weak classifier using multiple features
• ? classifier fusion
• Use different weak classifier such as SVM, NN,
  threshold function, etc.
• Different feature selection method SVM

47
Discussion

• Is important to add feature Selection for Image
  retrieval
• A good methodology for selecting features should
  be used
• Adaboost is a learning algorithm
• ? data dependent
• It is important to have representative samples
• Adaboost can help to improve the classification
  potential of simple algorithms

48

• Thank you !