Object Recognition EEE 6209 – Digital Image ProcessingObject Recognition Choice of Features Degree...
Transcript of Object Recognition EEE 6209 – Digital Image ProcessingObject Recognition Choice of Features Degree...
EEE 6209 – Digital Image Processing
© Dr. S. M. Mahbubur Rahman
Object Recognition
Outline
Patterns and Classes
Decision-Theoretic Methods
Minimum Distance Classifier
Matching by Correlations
Bayes Classifiers
Assignments
EEE 6209 – Digital Image Processing
© Dr. S. M. Mahbubur Rahman
Object Recognition
Patterns & Classes Patterns are usually vectors comprising qualitative or structural descriptors. Say there are n number of patterns (often referred to as features)
[ ]Tnxxxx ,,,, 321 =x
Class is a family of patterns that share common properties. Say there are W number of classes
Wωωωω ,,,, 321
Classical works is known as Fisher discriminant analysis (1963)
EEE 6209 – Digital Image Processing
© Dr. S. M. Mahbubur Rahman
Object Recognition
widthpetal
length petal
2
1
==
x
x
setosa Iris
r versicoloIris
virginicaIris
3
2
1
===
ωωω
2-Patterns
3-Classes
Example
EEE 6209 – Digital Image Processing
© Dr. S. M. Mahbubur Rahman
Object Recognition
Choice of Features Degree of class separability depends strongly on the choice of descriptors selected for an application
Patterns can be noisy. See noisy signature in the figure. Other quantitative features can be statistical moments, Fourier descriptors, etc. ( )
( )
( )nn rx
rx
rx
θ
θθ
=
==
22
11
Features with structural relations include the size and positions of minutiae that describes the ridge properties of fingerprints called primitive components such as abrupt endings, branching, merging and disconnected segments.
EEE 6209 – Digital Image Processing
© Dr. S. M. Mahbubur Rahman
Object Recognition
Choice of Features String of primitive components (say a and b) associated with a boundary shape often adequately describes structural patterns from head-to-tail
patterns⇒abababab
Hierarchical ordering of patterns often produces tree structure of the problem of classification.
EEE 6209 – Digital Image Processing
© Dr. S. M. Mahbubur Rahman
Object Recognition
Choice of Features
Satellite image
Tree description of image
EEE 6209 – Digital Image Processing
© Dr. S. M. Mahbubur Rahman
Object Recognition
Feature Selection Noisy features are refined using the criterion of maximum relevancy and minimum redundancy of the features among the classes.
An example of selection of features use Fisher criterion that uses the maximum of the ratio between inter-class and intra-class distances of the features to select only the discriminative set of features.
Refinement of features not only improves the classification accuracy but also reduces the computational complexity.
EEE 6209 – Digital Image Processing
© Dr. S. M. Mahbubur Rahman
Object Recognition
Decision-Theoretic Recognition Let the decision functions be ( ) ( ) ( )xxx Wddd ,,, 21
A pattern x belongs to a class if iω
( ) ( ) jiWjdd ji ≠=> ;,,2,1 xx
Ties are resolved arbitrarily.
Decision boundary of classes and is iω jω( ) ( ) ( ) 0=−⇒ xxx jiij ddd
such that ( )( ) jij
iij
•d
•d
classpattern 0
classpattern 0
⇒<
⇒>
x
x
EEE 6209 – Digital Image Processing
© Dr. S. M. Mahbubur Rahman
Object Recognition
Minimum Distance Classifier Use Euclidean distance of feature vectors to determine a class
Let is the number of pattern vectors of class . Then mean of pattern vector is
jN
WjN
j
jj
j ,,2,1 1
== ∑∈ωx
xm
The distance of a given pattern vector from the mean vector is
x
jω
( )
( ) 21
,,2,1
jTj
Tjj
TT
jj WjD
mmxmmxxx
mxx
+−−=
=−=
EEE 6209 – Digital Image Processing
© Dr. S. M. Mahbubur Rahman
Object Recognition
Minimum Distance Classifier Since the term is independent of class label j the decision function is
( ) Wj d jTjj
Tj ,,2,1
2
1=−= mmmxx
The decision boundary is ( ) ( ) ( )
( ) ( ) ( ) 02
1 =+−−−=
−=
jiT
jijiT
jiij ddd
mmmmmmx
xxx
xxT
The decision surface is perpendicular to the bisector of the line segment joining and im jm
bisector hyperplane 3
bisector plane 3
bisector line2
⇒>⇒=⇒=
n
n
n
EEE 6209 – Digital Image Processing
© Dr. S. M. Mahbubur Rahman
Object Recognition
Minimum Distance Classifier Example
( ) ( ) ( )09.88.2 21
2112
=−+=−=
xx
ddd xxx
The decision functions are
( )
1.103.13.4 2
1
21
1111
−+=
−=
xx
d TT mmmxx
The decision boundary is
setosa Iris Class
r versicoloIris Class
2
1
⇒⇒
ωω
=
=
3.0
5.1 and
3.1
3.421 mm
( )
17.13.05.1 2
1
21
2222
−+=
−=
xx
d TT mmmxx
EEE 6209 – Digital Image Processing
© Dr. S. M. Mahbubur Rahman
Object Recognition
Minimum Distance Classifier Decision boundary of two classes is a line
( )( ) 212
112
classpattern 0
classpattern 0
•d
•d
⇒<⇒>
x
x
EEE 6209 – Digital Image Processing
© Dr. S. M. Mahbubur Rahman
Object Recognition
Minimum Distance Classifier Minimum distance classifier is computationally very fast
The classifier shows optimum performance if the distribution of patterns for each class about its mean is in the form of a spherical hyper-cloud in n-dimensional space
Example of large mean separation and small class spread happens in designing E-13B font character set used by the American Banker’s Association.
The 14 characters were designed on 9x7 grid to facilitate reading. The characters are printed using magnetic materials. Scanning produces 1D electrical waveform which are distinct.
EEE 6209 – Digital Image Processing
© Dr. S. M. Mahbubur Rahman
Object Recognition
E-13B font character used by American Banker’s Association.
The mean vectors are 14,,2,1 =j jm
High classification speed is obtained using the minimum distance classifier
Area of waveform remains approximately constant and zero derivative near the middle of the character
Minimum Distance Classifier
EEE 6209 – Digital Image Processing
© Dr. S. M. Mahbubur Rahman
Object Recognition
Let the template for recognition is
The correlation coefficient can be measured using Fourier transform efficiently
( )yxw ,
The correlation coefficients
Matching by Correlation
( )( )[ ] ( ) ( )[ ]
( )[ ] ( ) ( )[ ] 21
22,,,
,,,,
++−++−
++−++−=
∑∑∑∑
∑∑∑∑
s ts t
s ts t
tysxftysxfwtsw
tysxftysxfwtswyxγ
( ) ( ) ( )( ) ( )vuWvuF
yxwyxfyx
,,
,,,∗=
⊗=γ
The maximum correlation gives the positional information of the template
( )11 ≤≤− γ
EEE 6209 – Digital Image Processing
© Dr. S. M. Mahbubur Rahman
Object Recognition
Matching by Correlation
Template matching
(a) (b)
(c) (d)
(a) Satellite image of hurricane (b) Template of eye of storm
(c) Correlation coefficients (d) Position of eye estimated
EEE 6209 – Digital Image Processing
© Dr. S. M. Mahbubur Rahman
Object Recognition
Let the probability that a pattern in test comes from the class is
Bayes Classifier
Let denote the loss when the classifier decides came from the class , when its actual class is
xiω
ijL xjω iω
Since the pattern may belong to anyone of the classes, the average loss incurred in assigning to the class is
x Wx jω
( )x|ip ω
( ) ( )∑=
=W
kkkjj pLr
1
| xx ω
Using Bayes formula the conditional average risk or loss can be written as
( ) ( ) ( ) ( )k
W
kkkjj PpL
pr ωω∑
=
=1
|1
xx
x
EEE 6209 – Digital Image Processing
© Dr. S. M. Mahbubur Rahman
Object Recognition
Here
Bayes Classifier kω( )kp ω|x PDF of patterns in class
( )kP ω Probability of occurrence of class kω
Since the term is positive and common to all , when it can be dropped for estimating smallest or largest value of the loss. Hence, the average loss function reduces to
( )xp ( )xjrWj ,,2,1 =
( ) ( ) ( )k
W
kkkjj PpLr ωω∑
=
=1
|xx
The classifier that minimizes the total average risk is called the Bayes classifier
EEE 6209 – Digital Image Processing
© Dr. S. M. Mahbubur Rahman
Object Recognition
The classifier assigns an unknown pattern in the class only if , i.e.,
Bayes Classifier iωx
( ) ( ) jiWjrr ji ≠=< ;,,2,1for xx
( ) ( ) ( ) ( ) jijPpLPpL q
W
qqqjk
W
kkki ≠∀<∑∑
==
||11
ωωωω xx
The loss function is zero and unity for correct and incorrect decisions, respectively. Hence,
ijijL δ−=1
The average loss function becomes
( ) ( ) ( ) ( )
( ) ( ) ( )jj
k
W
kkkjj
Ppp
Ppr
ωω
ωωδ
|
|11
xx
xx
−=
−=∑=
EEE 6209 – Digital Image Processing
© Dr. S. M. Mahbubur Rahman
Object Recognition
The Bayes classifier assigns the unknown pattern in the class only if
Bayes Classifier
iωx
( ) ( ) ( ) ( ) ( ) ( ) || jjii PppPpp ωωωω xxxx −<−
In other words, the classifier considers
In conclusion, the decision function of Bayes classifier under the 0-1 loss function is a conditional density given by
jiWj ≠= ;,,2,1for
( ) ( ) ( ) ( ) || jjii PpPp ωωωω xx > jiWj ≠= ;,,2,1for
( ) ( ) ( )jjj Ppd ωω|xx = Wj ,,2,1for =
EEE 6209 – Digital Image Processing
© Dr. S. M. Mahbubur Rahman
Object Recognition
The Bayes classifier needs PDF of patterns in each of the classes and the probability of occurrence of each class
Bayes Classifier
( ) jP ω
The PMFs are usually inferred from previous knowledge or training set of the problem. Fro example, if all classes are equally likely then
The PDFs need multivariate analysis since the dimension of is n. In practice, analytical solutions are obtained assuming parametric PDFs for patterns in a class.
( )W
P j
1=ω
( )jp ω|x
( )jp ω|xx
EEE 6209 – Digital Image Processing
© Dr. S. M. Mahbubur Rahman
Object Recognition
Naïve Bayes Classifier The Bayes classifier assuming Gaussian PDF of patterns in each of the classes is known as Naïve Bayes classifier. Example for two equi-probable patterns has a decision boundary at 0x The decision function is
( ) ( ) ( )( )
( )j
mx
j
jjj PePpd j
j
ωπσ
ωω σ 2
2
2
22
1|
−−
== xx
On the decision boundary ( ) ( )0201 xdxd =
If classes are equi-probable
2,1for =j
( ) ( )21 ωω PP =
Then, is s.t. 0x
( ) ( )2010 || ωω xpxp =
EEE 6209 – Digital Image Processing
© Dr. S. M. Mahbubur Rahman
Object Recognition
Naïve Bayes Classifier The Naïve Bayes classifier assuming multivariate Gaussian PDF of patterns with dimension n can be obtained. The PDF for j-th pattern class is
( )( )
( ) ( )jjT
j
ep
j
nj
mxCmx
Cx
−−− −
=1
21
21
22
1|
πω
where
{ } ∑∈
≈=jj
jj NE
ωx
xxm1Mean vector
Covariance-Matrix ( )( ){ }( )∑
∈
−≈
−−=
j
Tjj
T
j
Tjjjj
N
E
ωx
mmxx
mxmxC
1
The off-diagonal elements cjk are zero when xj and xk are uncorrelated
EEE 6209 – Digital Image Processing
© Dr. S. M. Mahbubur Rahman
Object Recognition
Naïve Bayes Classifier Using ln as monotonic function the decision function of Bayes classifier becomes
Considering multivariate Gaussian as prior for pattern classes
( ) ( ) ( )( )( ) ( )jj
jjj
pP
Ppd
ωω
ωω
|lnln
|ln
x
xx
+=
= Wj ,,2,1for =
( ) ( ) ( )( ) ( ) ( ) ( )[ ]( ) ( ) ( )[ ]jj
Tjjj
jjT
jjj
jjj
P
nP
pPd
mxCmxC
mxCmxC
xx
−−−−=
−−−−−=
+=
−
−
1
1
2
1ln
2
1ln
2
1ln
2
12ln
2ln
|lnln
ω
πω
ωω Wj ,,2,1for =
Decision function is hyperquadratic
EEE 6209 – Digital Image Processing
© Dr. S. M. Mahbubur Rahman
Object Recognition
Naïve Bayes Classifier If the covariance matrix is equal for all classes, i.e.,
The linear decision function in terms of hyper-plane is obtained by eliminating all terms except with terms having j
Wj ,,2,1for =
( ) ( ) jT
jjT
jj Pd mCmmCxx 11
2
1ln −− −+= ω
Bayes classifer becomes minimum distance classifier (hyper-spehers) when (i) pattern classes are Gaussian, (ii) covariance matrix is identity matrix and (iii) all classes are equally likely
CC =j
Wj ,,2,1for =
If the covariance matrix and classes have IC = ( )W
P j
1 =ω
( ) jT
jjT
jd mmmxx2
1 −= Wj ,,2,1for =
EEE 6209 – Digital Image Processing
© Dr. S. M. Mahbubur Rahman
Object Recognition
Naïve Bayes Classifier Examples of 2 pattern classes each having 3 dimensions of patterns.
−−==311
131
113
16
121 CC
=
1
1
3
4
11m
=
3
3
1
4
11m
( ) ( )2
1 11 == ωω PP
Given that
The decision function
( ) jT
jjT
jd mCmmCxx 11
2
1 −− −= 2,1for =j
Here
−−
−−=−
844
484
4481C
Since ( )( ) 5.5884
5.14
3212
11
−++−=−=
xxxd
xd
x
x
The decision surface
( ) ( ) 4888 32121 +−−=− xxxdd xx
EEE 6209 – Digital Image Processing
© Dr. S. M. Mahbubur Rahman
Object Recognition
Naïve Bayes Classifier Examples of 3 pattern classes each having 4 dimensions of patterns in classifying objects of multispectral images.
The patterns are
====
4
3
2
1
x
x
x
x Visible blue (Band-1)
Visible green (Band-2)
Visible red (Band-3)
Infra red (Band-4)
The classes are
===
3#
2#
1#
class
class
class Water body
Urban development
Vegetation
EEE 6209 – Digital Image Processing
© Dr. S. M. Mahbubur Rahman
Object Recognition
Bayes Classifier (a) (b) (c)
(d) (e) (f)
(g) (h) (i)
(a)-(d) Four spectral images, (e) training regions for 3-type of objects (f) classification results of training sets (observe errors shown in black dots) (g) water body (h) urban area (i) vegetation for whole image
EEE 6209 – Digital Image Processing
© Dr. S. M. Mahbubur Rahman
Object Recognition
Bayes Classifier Confusion matrix for the results of classification
( )3
1 =jP ω
Condition is
3,2,1for =j
EEE 6209 – Digital Image Processing
© Dr. S. M. Mahbubur Rahman
Object Recognition
Assignment Problem #1
Problem #2
theory
EEE 6209 – Digital Image Processing
© Dr. S. M. Mahbubur Rahman
Object Recognition
Assignment Problem #3
#2