types of pattern recognition model lecture notes

Types of pattern

1. Lecture 2: Types of PR

2. Examples of applications
• Natural Scenes
• Biometrics
• Diagnostic systems
• Military applications
• Outdoor Scenes, Real World
• Autonomous Land Vehicle
• Aerial Imaging (Airport Scenes)
• Face recognition, verification, retrieval.
• Finger prints recognition.
• Medical diagnosis: X-Ray.
• Automated Target Recognition (ATR).
• Image segmentation and analysis (recognition
from aerial or satelite photographs).

3. Images
Input image
Features,attributes
and relationships
Objects
Labeled Objects
Low level
Intermediate Level
High level
Hierarchy of Image Understanding System (IUS)
I
U
S

4. DESCRIPTION
SEMANTIC
INTERPRETATION
SCENE
MODELS
SYMBOLIC
REPRESENTATION
REGION / EDGE
FEATURE
EXTRACTION
IMAGE
BLOCK DIAGRAM OF BOTTOM-UP APPROACH

5. SCENE
DESCRIPTION
INTERPRETATION
MODEL
SYMBOLIC
REPRESENTATION
REGION / EDGE
FEATURE
EXTRACTION
IMAGE
TOP-DOWN APPROACH

6. SCENE
DESCRIPTION
INTERPRETATION
MODEL
SYMBOLIC
REPRESENTATION
REGION / EDGE
FEATURE
EXTRACTION
IMAGE
REPRESENTATIVE BLOCK DIAGRAM OF
TOP DOWN BOTTOM UP APPROACH

7. BLACKBOARD MODEL APPROACH
BLACKBOARD
METHODS
KNOWLEDGE
SOURCES
SCHEDULER

8. Approaches
Statistical Model: based on underlying statistical
model of patterns and pattern classes.
Syntactic Model: pattern classes represented by
means of formal structures as grammars, automata,
strings, graphs, constraint matrix etc.
Structural Model: knowledge base,

9. Statistical PR

10. An Example
“Sorting incoming Fish on a conveyor
according to species using optical
sensing”
Sea bass
Species
Salmon

11. Problem Analysis
Set up a camera and take some sample images to
extract features
Length
Lightness
Width
Number and shape of fins
Position of the mouth, etc…
This is the set of all suggested features to explore for use in
our classifier!

12.  Preprocessing
Use a segmentation operation to isolate fishes
from one another and from the background
Information from a single fish is sent to a
feature extractor whose purpose is to
reduce the data by measuring certain
features
The features are passed to a classifier

14. Classification
Select the length of the fish as a possible
feature for discrimination

16. The length is a poor feature alone!
Select the lightness as a possible
feature.

18. Threshold decision boundary and cost
relationship
Move our decision boundary toward smaller
values of lightness in order to minimize the cost
(reduce the number of sea bass that are classified
salmon!)
Task of decision theory

19. Adopt the lightness and add the width of
the fish
Fish x
T
= [x
1, x
2]
Lightness Width

21. We might add other features that are not
correlated with the ones we already have. A
precaution should be taken not to reduce
the performance by adding such “noisy
features”
Ideally, the best decision boundary should
be the one which provides an optimal
performance such as in the following figure:

25. Decision given the posterior probabilities
X is an observation for which:
if P(
1 | x) > P(
2 | x) True state of nature = 
1
if P(
1 | x) < P(
2 | x) True state of nature = 
2
Therefore:
whenever we observe a particular x, the probability of
error is :
P(error | x) = P(
1 | x) if we decide 
2
P(error | x) = P(
2 | x) if we decide 
1

26. Minimizing the probability of error
Decide 
1 if P(
1 | x) > P(
2 | x);
otherwise decide 
2
Therefore:
P(error | x) = min [P(
1 | x), P(
2 | x)]
(Bayes decision)

27. Optimal decision property
“If the likelihood ratio exceeds a
threshold value independent of the
input pattern x, we can take optimal
actions”

28. Syntactic PR

29. Syntactic Pattern Recognition Technique
Syntactic Pattern Recognition consists of three major
steps:
Preprocessing which improves the quality of an
image, e.g. filtering, enhancement, etc.
Pattern representation which segments the picture
and assigns the segments to the parts in the model
Syntax analysis which recognizes the picture
according to the syntactic model: once a grammar has
been defined, some type of recognition device is
required, the application of such a recognizer is called
Parsing.
Syntactic methods are best suited to problems with
clear primitives and stable intermediate structures ,
with well defined and known alternative. A very
important issue in syntactic pattern recognition is that
of grammatical interface.

30. Block Diagram of a Syntactic Pattern Recognition
System
Preprocessing
Representation
Construction
Syntax
Analysis
(Parsing
)
ClassificationInput
Pattern
Automata
Construction
Grammatical
Interface
Primitive
(and relation)
selection
Segmentation
Primitive
(and relation)
recognition
Training
Patterns
Recognition
Training

31. Grammar
Grammars can be classified to four categories according to their
productions:
Unrestricted grammar
Context sensitive grammar
Context free grammar
Regular grammar
Depending on the options available at each stage of the rewriting
process, a grammar can be classified as:
Deterministic
Non-deterministic
Stochastic

32. I
K
L
H
J
A
B
FG
Scene
Objects
Object A Object B
 F G
Face HFace IFace J
Background
(Subpattern)
Floor LWall K
(Subpatterns)
(Subpatterns)
SEMANTIC PRIMITIVES
Hierarchical Structures

33. Pattern Primitives
unit
a
b
c
d
aaa
b
b
ccc
d
d
3 units
2 units
String form:aaabbcccdd
More explicitly: ‘+’ head-to-tail
concatenation
a+a+a+b+b+c+c+c+d+d
Tree Structure
Rectangle
a+a+a+b+bc+c+c+d+d+
An Example

34. Pattern and its Structural Description
Pattern
Subpattern
Pattern
primitivesc + c + d + a + a + b b + b + c + c
An Example

35. For this approach to be advantageous the simplest subpattern selected
called pattern primitives, should be much easier to recognize the
patterns themselves.
Composition
Of primitives
patterns
Accomplish by Grammar
Once the primitives of the pattern are identified than recognition is
accomplished by performing a syntax analysis or parsing.

36. a
a
a
a
b
b
b
b
b
b
b
b
d
d
c
c
a
a
b
b
b
b
e
c
Submedian Chromosome telocentric Chromosome
Context-free grammar
G = (V
N, V
T, P {<Submedian Chromosome>, <telocentric Chromosome>})
Where
V
N ={<Submedian Chromosome>, <telocentric Chromosome>,
<arm pair>, <left part>, <right part>, <arm>, <side>, <bottom>}
V
T =
a b ec d
An Exhaustive Example

37. Productions {P}:
<Submedian Chrom.> <arm pair> <arm pair>
<telocentric Chrom.> <bottom> <arm pair>
<arm pair> <side> <arm pair>
<arm pair> <arm pair> <side>
<arm pair> <arm> <right part>
<arm pair> <left part> <arm>
<left part> <arm> c
<right part> c <arm>
<bottom> b <bottom>
<bottom> <bottom> b
<bottom> e
<side> b <side>
<side> <side> b
<side> b
<side> d
<arm> b <arm>
<arm> <arm> b
<arm> a

38. babcbabdbabcbabd
BOTTOM-UP PARSING OR DATA DRIVEN
Submedian Chromosome
arm pairarm pair
arm pair arm pair
left part left part
arm arm arm arm
arm arm arm arm
arm
ba
arm armarm side
bcbabdbabcbabd
Now Given String
side

39. Structural PR

40. Example: Outdoor Scenes
What is Image Understanding?
Image Understanding is a process of
understanding the image by identifying
the objects in a scene and establishing
relationships among the objects.
Image Understanding is the task-
oriented reconstruction and
interpretation of a scene by means of
images.

41. Different Image understanding
Systems
VISION [C.C. Parma] (1980)
ARGOS [Steven M. Rubin] (1978)
ACRONYM [R. A. Brooks] (1981)
MOSAIC [T. Herman and T. Kanade] (1984)
SPAM [D. M. McKeown] (1985)
SCERPO [Lowe] (1987)
SIGMA [Takashi and Vincent Hwang] (1990)
Knowledge-Based Medical Image Interpretation
[Darryl N. Davis] (1991)
and many more …

42. An example: Labeling of a very small picture (satellite) of
agrarian nation
Knowledge Network
Assume satellite photograph: 4  2 pixels
ie
1234
5678
> 6000 possible labellings
Some possible labelling of 4  2 pixels image given the constraints
of the knowledge network
AABC
AABC
AABC
ABBC
AABC
BBBC
ABBC
ABBC
ABBC
BBBC
BBBC
BBBC
ABCC
ABCC
ABCC
BBCC
BBCC
BBCC
BCCC
BCCC
NTOP
BOTTOM
LEFT RIGHT
ALFALFA
CORN
BARLEY

43. Example 2:
A1
A2
A3
H1
H2
cars (c)
(arbitrary)
streets (S)houses (H)
(at least one)(arbitrary)
other areas (A)
(at least two)
on
(seldom)
on
(often) on
(always)
inters.,
Junc.,
adj.
SCENE MODEL
SCENE
INSTANCE OF THE MODEL OBJECT MODEL (CAR)
on
on
on
adj adj
adj
adj
adj
adj
Junc.
H1
H2A1C
A2 S1 S2 A3
above
same
size
above
above
before
beforebody
rear wheel
front wheel
hood
pess. room
luggage boat
S1
S2

44. Given:
Object
Model
Image
How to match ?
Two Approaches:-
(i)Isomorphism
(ii)Largest Clique

45. SCENE CLASSES
HOUSE
SCENE
ROAD
SCENE
CAR
ROAD
GARAGE
TREE
HOUSE
DRIVEWAY
GUARDRAILS TELEPHONE
POLES
N
O
. O
F
N
O
. O
F
O
N
S
I
D
E
LIKE - A

46. TELEPHONE
POLES
NETWORK OF GENERAL ROAD_SCENE
HOUSE (H)
[ARBITRARY]
ROADS (R)
[AT LEAST ONE]
ROAD RAIL
[ARBITRARY LENGTH]
TREES
[ARBITRARY]
CARS (C)
[ARBITRARY]
OTHER AREAS
[AT LEAST TWO]
ON
ALWAYS
O
N
A
LW
A
Y
S
O
N
ALW
AY
S
OFTEN
ADJACENT TO
A
D
J
A
L
W
A
Y
S
BEHIND
ALWAYS
C
O
N
N
E
C
T
I
O
N
B
E
H
I
N
D
O
F
T
E
N
O
N
R
A
R
E
L
Y

47. Segmentation
Given input image is segmented into different
regions by using different segmentation methods.
1.Cluster based segmentation
•K-means clustering algorithm
•Porter and Canagarajah Method
•Validity Measure Method
2.Ohlander type segmentation
3.Comaniciu and Meer proposed Algorithm

48. By observing the results of above segmentation
algorithms, segmented image obtained by Comaniciu and
Meer proposed segmentation Algorithm is only used for
further processing.
Results obtained by Comaniciu and Meer
proposed segmentation algorithm are..
Original Image Segmented Image

49. Feature Extraction
Two types of features are extracted from the
segmented regions.
1.Primary features:
These features calculation directly deals with image arrays. It
includes Area, Mean intensities of R,G&B, contour length,
mass centers,Minimum bounding rectangle(MBR) and etc.
2. Secondary features :
These features are calculated from a set of primary features.
It includes compactness of regions, linearity, normalized
colors, Hue, saturation and intensity values and etc.

50. Formulae :
r = R/(R+G+B),
g = G/(R+G+B),
b = B/(R+G+B),
Intensity = (R+G+B)/3,
Hue = arc tan 2(3
1/2
(G-B),(2R-G-B)),
Saturation = 1-3 min(r, g, b),
Compactness = 4 (area)/(contour length)
2
,
and many more…

51. MINIMUM BOUNDING RECTANGLE (MBR)
AN IRREGULAR FIGURE
X
Y
X MAX
X MIN
Y

M
I
N
Y

M
A
X
AREA (A)
PERIMETER (P)

52. (a)
1
2 0
3
X
Y
L
E
N
G
T
H
(b)
KEY FEATURES OF BOUNDARY SEGMENTS
KEY FEATURES OF LINE SEGMENTS
CONTRAST
BOUNDARY CHAIN CODES
X
Y
LENGTH
END POINT



53. Image Synthesis
An image is constructed from the primary and secondary
features those are extracted in the earlier phase. This synthesized
image is compared with the original image.
Greater the similarities ,better the extracted features and
segmentation algorithm.
Synthesized Image Original image

54. This scene interpretation Module includes various
sub-modules. those are…
1.Outdoor Scene Knowledgebase
•Knowledge Extraction
•Knowledge Representation
2.Plan Generation
3.Relaxation Labeling
4.Image Interpretation
Scene Interpretation Module

55. Outdoor Scene Knowledgebase
Outdoor scene Knowledgebase contains all the knowledge
about the scenes like raw picture data(RGB values,edges
etc), features information, constraints on those features and
etc. Knowledgebase for outdoor scenes is divided into 3 parts.
1. Knowledge Extraction
•Region Features
•Object Features
•Binary object relations
2. Knowledge Representation
•Image domain knowledge
•Scene domain knowledge
3.Constraint Matrix Generation

56. Knowledge Extraction
• Region Features: This relation is domain-independent, which
includes region attributes such as region number, the three-color
features, average intensity, centroid, area, minimum bounded
rectangle, hue, saturation and linearity.
• Object Features: This relation contains the object name as a
domain, followed by a set of domain triplets, each of which
contains the feature name, whether or not this feature has a veto,
and acceptable variation for each of the feature.
Features extracted for objects ‘Sky’, ‘Building’, ‘Tree’ and
‘Grass’ are
 Sky: R = [49,154], G = [96,194], B = [162,239], Hue =
[202,218], Compactness = [0.0853, 0.8289], Linearity = [43.038,
82.1362].

57. Grass: Hue = [71,105], Compactness = [0.0813, 0.168],
Linearity = [39.4317, 75.77], R = [76, 141], G = [92, 166], B =
[36, 101].
 

Tree: Hue = [51, 93], Compactness = [0.002, 0.3801], Linearity
= [25.965, 35.963], R = [49, 154], G = [96, 194], B = [162, 239].
Building: Compactness = [0.0631, 1.1719], Linearity =
[39.8625, 84.8485].
• Binary Object relation: This relation includes binary spatial
relationship between each pair of regions: left, right, above and
below. This relation describes the constraints imposed on object
pairs.

58. Knowledge Representation
•Image domain Knowledge: This type of knowledge is used to
extract features from an image and to group them to construct the
structural description of the image. Knowledge extracted from the
segmented regions of the image is represented as follows
Rno Min:c,r Max:c,r Hole R G B area Mass_Ctr Hue Satrtion Linearity
1 0,0 41,21 0 125 172 223 151 17,3 211 111 63.8095
2 5,0 11,2 1 113 163 225 13 8,0 213 126 54.5455
3 35,0 299,52 0 95 151 222 5505 197,12 213 145 77.4242
4 15,2 87,22 0 230 234 236 957 50,12 170 8 77.551
5 0,3 13,35 0 110 166 232 334 6,17 212 133 48.8095

59. •Scene domain Knowledge: This type of knowledge includes
intrinsic properties of and mutual relations among objects in the
world such as names of objects and their constituent parts, the
geometric coordinate system to specify location and spatial
relations between those objects.
OBJ-1 OBJ-2 pos diff val Rpos Rdif Rval Gpos Gdif Gval Bpos Bdiff Bval Hpos Hdiff Hval Lpos Ldiff Lval
Relation
SKY SKY diff LE 70 LE LE 30 LE LE 35 LE LE 20 LE LE 20 LE LE 35 NULL
SKY BUILD above GT 30 GT GT 30 GT GT 20 GT GT 50 GT GT 20 LT LT 20 NULL
SKY TREE above GT 20 GT GT 40 GT GT 50 GT GT 10 GT GT 50 GT GT 20 NULL
SKY GRASS above GT 150 GT GT 50 GT GT 70 GT GT 10 GT GT 120 GT GT 5 NULL
SKY ROAD NULLNULL 0 NULLNULL0 NULL NULL0 NULL NULL0 NULL NULL 0 NULL NULL 0 NULL
BUILD BUILD diff LE 60 LT LE 70 LE LE 65 LE LE 75 LE LE 30 GE GE 5 CONTAINS
BUILD SKY below GT 30 LT GT 30 LT GT 20 LT GT 50 LT GT 30 GT LT 20 NULL
BUILD TREE left GT 50 GT GE 60 GT GE 50 GT GT 40 NULL NULL 0 GT GT 20 NULL
BUILD TREE right GT 30 GT GE 60 GT GE 50 GT GT 40 NULL NULL 0 GT GT 20 NULL

60. • Knowledge about the mapping between Image and Scene:
This type of knowledge is used to transform image features into
scene features and vice versa.
OBJNO Object RMIN RMAX GMIN GMAX BMIN BMAX HMIN HMAX LMIN LMAX COND UNQ$
1 SKY 102 250 96 250 162 255 180 235 44 83 B GT G NULL
1 SKY 102 250 96 250 162 255 180 235 44 83 B GT R NULL
1 SKY 102 250 96 250 162 255 180 235 44 83 M.Y LT 70 NULL
1 SKY 102 250 96 250 162 255 180 235 44 83 B GT 20 NULL
1 SKY 102 250 96 250 162 255 180 235 44 83 G GT 130 NULL
1 SKY 102 250 96 250 162 255 180 235 44 83 R GT 100 HUE
1 SKY 0 100 0 130 0 255 180 300 5 98 B GT G NULL
1 SKY 0 100 0 130 0 255 180 300 5 98 B GT R NULL
1 SKY 0 100 0 130 0 255 180 300 5 98 M.Y LT 70 NULL
1 SKY 0 100 0 130 0 255 180 300 5 98 B GT 10 HUE
2 TREE 57 195 63 230 27 190 51 183 5 40 G GT B NULL
2 TREE 57 195 63 230 27 190 51 183 5 40 G GT R NULL
2 TREE 57 195 63 230 27 190 51 183 5 40 LIN LT 44 LIN

61. Plan Generation
A plan is generated by assigning all possible
object labels to the segmented regions with some
confidence value. This object labels assignment is
based on the set of primary/secondary features
extracted in the previous work.
For each region, there exists a set of object
interpretations and their associated confidence
values(probabilities). For Example,
Region1= [(sky=0.38), (building=0.18), (tree=0.27),
(grass=0.04), (road=0.13)]

62. Confidence value of i
th
region being object ‘x’ is


k
m
C
i
x
= Q
j
.S
j
.W
x
j
such that Σ C
i
p
(t)=1 where,

j=1 p=1
m be the number of regions,
W
x
j
be the weight associated with object ‘x’ having j
th
feature,
Sky: Hue=[202,218], Compactness=[0.0854,0.8289], Linearity=[43.038,82,1362],
R=[49,154], G=[96,194], B=[162,239]
Q
j
=1 for |f
x
j-
f
i
j
| <= ε
x
i
(acceptable variation in j
th
feature),
=0 Otherwise, where, f
x
j
be the j
th
feature of object ‘x’
S
j = 2f
i
j
if
(f
x
j-
f
i
j
) >= 0

2-

f
x
j+
f
i
j
= 2f
x
j
if
(f
x
j-
f
i
j
) < 0

2-

f
x
j+
f
i
j and

63. Reg.no SKY TREE GRASS BUILD ROAD
1 0.651613 0.0574839 0.0574839 0.174194 0.0574839
2 0.119529 0.119529 0.119529 0.521886 0.119529
3 0.127622 0.127622 0.127622 0.48951 0.127622
4 0.125114 0.125114 0.125114 0.499543 0.125114
5 0.2 0.2 0.2 0.2 0.2
6 0.2 0.2 0.2 0.2 0.2
7 0.612903 0.063871 0.063871 0.193548 0.063871
8 0.122329 0.122329 0.122329 0.510685 0.122329
9 0.12013 0.12013 0.12013 0.519481 0.12013
10 0.2 0.2 0.2 0.2 0.2
Confidences obtained in plan generation for the above image are

64. Relaxation Labeling
Def:- Labeling the regions based on adjacent
regions.
Relaxation labeling technique iteratively updates
the confidence values of object labels assigned to each
of the large regions, based on the confidence values of
object labels of the neighboring regions.
This adjustment process requires the constraint
matrix which can be obtained from the Outdoor scene
knowledgebase.

65. For example,
Before relaxing
reg6=[(sky=0.0385514),(building=0.116822),(tree=0.766355),
(grass=0. 0385514), (road=0. 0385514)]
After relaxing
reg6=[(sky=0.00380988), (building=2.24154e-05),
(tree=0.948548),(grass=0.00127976),(road=0.00109291)]
Relaxation contd..

66. Mathematically, this relaxation can be done by using
following formulae …
•C
i
k
(I
x
)

be the confidence that region R
i
has an interpretation
equals to

I
x
at iteration ‘k’ ,
• I
x


I ,where ‘I’ be the set of all pertinent object interpretation
in the knowledgebase,
0 C
i
k
(I
x
)  1.0 for all I
x
II and for all iM.
C
i
k
(I
x
)=1.0 for all iM.

Ix I

•‘A’ be the set of all adjacent regions to R
i
Relaxation contd..

67. The incremental change for each interpretation of the
region R
i can be calculated by
C
i
k
(I
x
)=  D
ij
 
ij
(I
x
,I
y
) C
j
k
(I
y
)

RjA IxI
C
i
k+1
(I
x
)= C
i
k
(I
x
)[1+C
i
k
(I
x
)]

 C
i
k
(I
x
)[1+C
i
k
(I
x
)]

IxI
 
The denominator is chosen in order to ensure that
 
 
C
i
k+1
(I
x
)=1.
IxI
Relaxation contd..

68. Let R
j
A be the j
th
adjacent region to the current region R
i

D
ij
be the adjacency of the regions

R
i
&

R
j
,given by the ratio of the
common perimeter of the region R
j
to the total perimeter of the
region R
i
.
 D
ij =
1.0
RjA
and 
ij
(I
x
,I
y
) be the compatibility of regions R
i
&

R
j
when
interpreted as objects I
x
and I
y
, respectively.
 ij(Ix ,Iy)=
+1 if Ri and Rj are compatible as objects Ix and Iy
-1 if Ri and Rj are not compatible as objects Ix and Iy


ij
(I
x
,I
y
) values can be obtained from the constraint matrix which is
stored in the knowledgebase.
 
Relaxation contd..

69. Example for outdoor Scene
Knowledgebase

Outdoor scene Knowledgebase for the following house
image can be stored as follows
For the house image constraints should be..
1.Roof should be upside of the image,
2.Window is always embedded in the wall,
3.Roof should not be adjacent to the floor
4.Window cannot be in the floor/Roof
5.Wall should be adjacent to roof and
6. Wall contains door and window,
And etc..

70. All these binary constraints should be stored in the
knowledgebase in the form of constraints matrix.
Constraint matrix for House image is..

 
RoofWallWindow FloorDoor
Roof
1 1 -1 -1 -1
wall 1 1 1 1 1
Window -1 1 -1 -1 -1
Floor -1 1 -1 1 1
Door -1 1 -1 1 -1

71. Interpretation
Finally, regions are labeled as objects based on their
confidence values.
Regions are interpreted as objects, in such a way
that the corresponding region has high confidence value
for that object interpretation.
For example:
Region4={(sky=0.783),(tree=0.154),
(building=0.0124),(road=0.0432),(car=0.0264)}

72. Segmentation
Input image
Feature Extraction
Segmented regions
Synthesized Image
Features & relations
Original Image
Comparison
Plan Generation
Features
OK
Labeled Objects
Relaxation Labeling
Outdoor scenes
Knowledgebase
Features
Interpretation
Convergence
Interpreted Objects
Image analysis
Module
Scene Interpretation
Module

73. Results
Original Image
Synthesized Image
Segmented Image Gray Image

74. Interpreted Image after
Relaxation
Interpreted Image
before Relaxation

75. References
• M.D.Levine and S.I.Shaheen,“A Modular Computer Vision System for
Picture Segmentation and Interpretation”, IEEE Transactions on Pattern Analysis
and Machine Intelligence vol.PAMI-3, no.5
• Robert A.Hummel and Steven W.Zucker, “On the Foundations of
relaxation Labeling” , IEEE Transactions on Pattern Analysis and Machine Intelligence
vol.PAMI-5, no.3.
•Takashi Matsoyama and Vincent Shang-Shouq Hwang, “SIGMA- A
knowledge-based Aerial Image Understanding System”, Series Editor: Martin D.
Levine, New York and London: Plenum Press, 1990.
•Ahmed E. Ibrahim, “An Intelligent Framework for Image Understanding”.
• Rafael C.Gonzalez, Richard E.Woods,“Digital Image Processing”, second
edition.
• K. SaiCharan, “Plan Generation for Outdoor `Natural Scene in Image
Understanding”, M.Tech thesis, HCU, (2004).

76. • S. W. Zucker, “Relaxation labeling and the reduction of local
ambiguities”, in pattern recognition and Artificial Intelligence, C. H. Chen, Ed. New
York: Academic, 1977.
• A. Agarwal, “A Framework for Distributed Machine Perception in
Understanding Static Natural Scenes”, Ph. D Thesis, IIT Delhi, 1989.
• Su Linying, Bernadette Sharp, and Claude C. Chibelushi: “Knowledge-
Based Image Understanding: A Rule-Based Production System for X-Ray
Segmentation”, ICEIS 2002: 530-533. www.soc.staffs.ac.uk/lys1/iceis2002.pdf