Professional Documents
Culture Documents
IP Full
IP Full
IP Full
Department of CSE
DC
1
Syllabus
Program: B. Tech in CSE (BCS) Year, Semester: 3rd yr., 7th Sem.
Course Title: Image Processing Subject code: TIU-UCS-E40#
Contact Hours/week: 3-0-0 (L-I-P) Credit: Theory – 3
Course Content
Module-1: Digital Image Fundaments, Morphological Image
Processing & Image Transforms
Module-2: Image Enhancement
Module-3: Image Segmentation
Module-4: Representation and description
Module-5: Clustering and Classification
Module-6: Real life data applications
DC
2
Requirements
• Prerequisites
– Data Structures and Algorithms
– Linear Algebra and Statistics
– Experience with Programming Language
• Text
– Computer Vision by Shapiro & Stockman, Prentice Hall
– R. C. Gonzalez, W. E. Woods, “Digital Image Processing”,
Pearson-Prentice Hall, Eds. – 3, 2008
– Pattern Classification and Scene Analysis by Duda and Hart,
Wiley-Interscience Publication, New York
– A. K. Jain, “Fundamentals of Digital Image Processing”, Prentice Hall,
Addison-Wesley, 1989
– W. K. Pratt, “Digital Image Processing”, PIKS inside, Wiley, Newyork,
3rd Eds, 2001
– M. Petrou, P. Bosdogianni, “Image Processing- The Fundamentals”,
Wiley, 1999
– Research papers in journals
DC
3
What is Digital Image?
• A digital image is a representation of a two-
dimensional image as a finite set of digital values,
called picture elements or pixels
DC
4
What is Digital Image?
• Pixel values typically represent gray levels,
colors, heights, opacities etc.
• Remember digitization implies that a digital
image is an approximation of a real scene
1 pixel
DC
5
What is Digital Image?
• Common image formats include:
– 1 sample per point (B&W or Grayscale)
– 3 samples per point (Red, Green, and Blue)
– 4 samples per point (Red, Green, Blue, and “Alpha”,
a.k.a. Opacity)
•Human perception
•Employ methods able to enhance pictorial
information for human interpretation and
analysis such as:
•Noise filtering
•Content enhancement
• - Contrast enhancement
• - Deblurring
•Remote sensing
DC
8
Image Sharpening
DC
10
Deblurring
DC
11
History of Digital Image Processing
DC
12
History of Digital Image Processing
DC
15
History of Digital Image Processing
DC
17
Examples: The Hubble Telescope
•Launched in 1990 the Hubble
telescope can take images of
very distant objects
•However, an incorrect mirror
made many of Hubble’s
images useless
•Image processing
techniques were
used to fix this
DC
18
Examples: Artistic effects
•Artistic effects are
used to make
images more visually
appealing, to add
special effects and to
make composite
images
DC
19
Examples: Medical Image Processing
•Take slice from MRI scan of canine heart,
and find boundaries between types of tissue
– Image with gray levels representing tissue
density
– Use a suitable filter to highlight edges
DC
21
Examples: GIS
•Night-Time Lights of
the World data set
– Global inventory of
human settlement
– Not hard to imagine
the kind of analysis
that might be done
using this data
DC
22
Examples: Industrial Inspection
•Image processing
techniques are used
extensively by law
enforcers
– Number plate
recognition for speed
cameras/automated
toll systems
– Fingerprint recognition
– Enhancement of CCTV
images
DC
24
What is Computer Vision?
DC
25
Relationships to other fields
DC
29
Origin
x f ( x, y )
DC
30
Fundamental steps
Representation
And Description
Segmentation
Preprocessing
Knowledge
Recognition
Base
And
Image Interpretation
Acquisition
Problem Result
Domain
DC
31
Key Stages in Digital Image
Processing
Image Morphological
Restoration Processing
Image
Segmentation
Enhancement
Image Object
Acquisition Recognition
Representation
Problem Domain
& Description
Colour Image Image
Processing Compression
DC
32
Key Stages in Digital Image
Processing: Image Acquisition
Image Morphological
Restoration Processing
Image
Segmentation
Enhancement
Image Object
Acquisition Recognition
Representation
Problem Domain
& Description
Colour Image Image
Processing Compression DC
33
Key Stages in Digital Image
Processing: Image Enhancement
Image Morphological
Restoration Processing
Image
Segmentation
Enhancement
Image Object
Acquisition Recognition
Representation
Problem Domain
& Description
Colour Image Image
Processing Compression DC
34
Key Stages in Digital Image
Processing: Image Restoration
Image Morphological
Restoration Processing
Image
Segmentation
Enhancement
Image Object
Acquisition Recognition
Representation
Problem Domain & Description
Colour Image Image
Processing Compression DC
35
Key Stages in Digital Image
Processing: Morphological Processing
Image Morphological
Restoration Processing
Image
Segmentation
Enhancement
Image Object
Acquisition Recognition
Representation
Problem Domain
& Description
Colour Image Image
Processing Compression DC
36
Key Stages in Digital Image
Processing: Segmentation
Image Morphological
Restoration Processing
Image
Segmentation
Enhancement
Image Object
Acquisition Recognition
Representation
Problem Domain
& Description
Colour Image Image
Processing Compression DC
37
Key Stages in Digital Image
Processing: Object Recognition
Image Morphological
Restoration Processing
Image
Segmentation
Enhancement
Image Object
Acquisition Recognition
Representation
Problem Domain
& Description
Colour Image Image
Processing Compression DC
38
Key Stages in Digital Image Processing:
Representation & Description
Image Morphological
Restoration Processing
Image
Segmentation
Enhancement
Image Object
Acquisition Recognition
Representation
Problem Domain
& Description
Colour Image Image
Processing Compression DC
39
Key Stages in Digital Image
Processing: Image Compression
Image Morphological
Restoration Processing
Image
Segmentation
Enhancement
Image Object
Acquisition Recognition
Representation
Problem Domain
& Description
Colour Image Image
Processing Compression DC
40
Key Stages in Digital Image
Processing: Colour Image Processing
Image Morphological
Restoration Processing
Image
Segmentation
Enhancement
Image Object
Acquisition Recognition
Representation
Problem Domain
& Description
Colour Image Image
Processing Compression DC
41
Element of digital image processing
systems
The basic operations performed in a digital image processing
systems include (1) acquisition, (2) storage, (3) processing, (4)
communication and (5) display.
Storage
Optical disks
Tape
Videotape
Mag disks
Image acquisition
equipment Processing Unit Display Unit
Video TV Monitors
Computer
Scanner Printers
Workstation
Camera Slide projectors
Communication Channel
Basic fundamental elements of image processing DC
42
Imaging Basics
DC
43
Pretest: Matrices
a11 a12 . . . a1n
a a22 . . . a2 n
21
.
A
. . . . .
A [aij ]mn
. . . . . .
. . . . . .
am1 am 2 . . . amn mn
Row Matrix, Column Matrix, Square Matrix, Diagonal Matrix
A 1 4 3 7 9 1
4 A aij mn ; m n
m n ; m n
A aij
aii 0, aij a ji 0
A 3
7
9 DC
44
Pretest: Matrices
mn ; m n
A aij
Identity Matrix:
aii 1, aij a ji 0
m n ; m n
A aij
Symmetric Matrix:
aij a ji i, j
Skew-symmetric Matrix:
m n ; m n
A aij
aii 0 and aij a ji i, j
0 4 0 0
Sparse Matrix: Most elements are 0
0 0 3 0
A
0 0 0 0
0 2 0 0
DC
45
Pretest: Matrices
A [ aij ]mn and B [bij ]mn A B [cij ]mn
Addition:
where cij aij bij
1 3 0 0 2 1 1 5 1
A 4 0 3 , B 2 3 1 A B 6 3 4
6 1 2 1 2 0 7 3 2
A [aij ]mn and B [bij ]mn A B [cij ]mn
Subtraction:
where cij aij bij
1 3 0 0 2 1 1 1 1
A 4 0 3, B 2 3 1 A B 2 3 2
6 1 2 1 2 0 5 1 2
A [aij ]mn 1 3 0 2 6 0
Scalar
and kA [ kaij ]mn A 4 0 3, k 2, kA 8 0 6
Multiplier: DC
6 1 2 12 2 4 46
Pretest: Matrices
A [ aij ]mn and B [bij ]n p AB [cij ]m p
Multiplication: n
where cij aik bkj , i 1,2,..., m; j 1,2,..., p
k 1
DC
48
Image Formation
• Physics of light
– Light sources
– Surface properties
DC
49
Image Formation
DC
50
Image Geometry and perspective
projection
y
x Object Point
f (x,y,z)
r
z
x r
(x,y) z
f r'
y
z r
Also
x' y' r'
x y r
f f
Therefore, x' x, y' y DC
z z 51
Imaging Geometry (Basic Transformations)
X * X X 0 , Y * Y Y0 , Z * Z Z 0
Translation:
( X , Y , Z ) : Old point, ( X *,Y *, Z *) : New point
and ( X 0 , Y0 , Z 0 ) : Displacement
X * 1 0 0 X 0 X
Y * 0 1 0 Y0 Y
Z * 0 0 1
Z0 Z V * TV
1 0 0 0 1 1
X * 1 0 0 X0 X
Y * 0 1 0 Y0 Y
V* T V
Z * 0 0 1 Z0 Z
1 0 0 0 1 1
DC
52
Imaging Geometry (Basic Transformations)
S x 0 0 0
Scaling: 0 Sy 0 0
S
The rotation angle
0 0 Sz 0
is measured
0 0 0 1 clockwise when
looking at the
Z cos sin 0 0 origin from a point
Rotation: sin cos 0 0 on the +Z axis.
R
0 0 1
0
Affects only the
values of X and Y
0 0 0 1 coordinates
Y
X
1 0 0 0 cos 0 sin 0
0 cos sin 0 0 1 0 0
R R
0 sin cos 0 sin 0 cos 0
DC
0 0 0 1 0 0 0 1 53
Coordinate Systems
• Model coordinates
– Own coordinate
• Absolute coordinates
– World coordinates
• Camera coordinates
– Measured w. r. t. the camera
DC
54
Simple Image Model
Image refers to a 2 - D light intensity function f ( x, y ) and 0 f ( x, y )
f ( x, y ) i ( x, y ) r ( x, y ) where 0 i ( x, y ) and 0 r ( x, y ) 1
We call the intensity of a monochrome image f at (x, y) the gray level (L ) of
the image at that point. Then Lmin L Lmax
f ( 0 ,0 ) f ( 0,1) f ( 0, M 1)
f (1,0 ) f (1,1) f (1, M 1)
f ( x, y )
f ( N 1,0 ) f ( N 1,1) f ( N 1, M 1) DC
55
Sampling & Quantization
The digitization process requires decisions about values for N, M, and the
number of discrete gray levels allowed for each pixel. Common practice in
digital image processing is to let these quantities be integer powers of two; that
is
N 2n , M 2k G 2m
The number, b, of bits required storing a digitized image:
b N M m If M = N b N2 m
For example, a 128128 image with 64 gray levels requires 98,304 bits of storage.
DC
56
Quantization
6 bpp 8 bpp
DC
57
Sampling
128 X 128 64 X 64 32 X 32
DC
58
Image Properties
• An array of pixels
• Pixel
– 1 bit per pixel (bpp): binary
– 8 bpp : grayscale
– 24 bpp : red, green and blue
– Satellite images have many more bpp
DC
59
Image Coordinates
• Many systems
– (a) Origin at top left
– (b) Origin at bottom left
– (c) Origin at the center
• Useful for different situations
– (a) most common
– (c) windows processing
DC
60
Image Coordinate Systems
DC
61
Levels of processing
• Point level
– thresholding
• Local
– averaging
• Global
– histogramming
• Object
– segmentation
DC
62
Binary Image Analysis
DC
1
Motivation
DC
2
Binary Images
DC
4
Binary Image
• Pixels are 0 or 1
– 1 foreground
– 0 background
• Assume mxn image
– Indexed from 0 .. m-1 for rows
– and 0 …. n-1 for columns
– I[i,j] : refers to individual pixels
DC
5
Neighborhood
How many neighbors does a pixel have?
DC
6
Neighborhood
• 4 neighborhood j
i
N i-1, j
W P E i, j-1 i, j i, j+1
S i+1, j
DC
7
Neighborhood
j
• Diagonal neighbor
i
DC
8
Neighborhood
j
• 8 neighborhood
i
DC
9
Basic Relationships Between Pixels
Connectivity:
• Connectivity between pixels is important
0 1 1 0 1 1 0 1 1
0 1 0 0 1 0 0 1 0
0 0 1 0 0 1 0 0 1
DC
11
Basic Relationships Between Pixels
DC
12
Relations and Equivalence
Binary relation : A binary relation R on a set A is a set of pairs of
elements from A. If the pair (a,b) is in R, denoted aRb as which
mean " a is related to b".
Example : A { p1, p2 , p3 , p4 } arranged as
p1 p2
p3
p4
R is the set of pairs of points from that are 4 - connected; that is,
R {( p1, p2 ), ( p2 , p1 ), ( p1, p3 ), ( p3 , p1 )}
DC
13
Relations and Equivalence
DC
14
Arithmetic and Logical Operation
NOT AND
OR XOR
NOT
AND
A B [NOT(A)]
DC
AND (B) 15
Applying Masks
• Basic concept
• Based on convolution
• Variety of tasks can be framed in terms of
mask operations
• Mask = a set of pixels
– Have associated weights
– Generally square
– Has an origin (usually the center pixel)
– Generally 2D or 1D
DC
16
Example of Masks
1 1 1 -1 0 1 -1 1
1 1 1 -1 0 1
1 1 1 -1 0 1
1
1 2 1
1
2 4 2
1
1 2 1
1
1
DC
17
Mask Operation
DC
18
Boundary Conditions
DC
19
Example
-1 0 1
-1 0 1
9 9 9 9 1 1 1
-1 0 1
9 9 9 9 1 1 1
9 9 9 9 1 1 1
9 9 9 9 1 1 1
9 9 9 9 1 1 1
9 9 9 9 1 1 1
9 9 9 9 1 1 1
DC
20
Example -1 0 1
Mask -1 0 1
-1 0 1
9 9 9 9 1 1 1
9 9 9 9 1 1 1
9 9 9 9 1 1 1
9 9 9 9 1 1 1 0 0 -24 -24 0
9 9 9 9 1 1 1 0 0 -24 -24 0
9 9 9 9 1 1 1 0 0 -24 -24 0
9 9 9 9 1 1 1 0 0 -24 -24 0
Image 0 0 -24 -24 0
Result DC
21
Example
40 40 80 80 80 1 2 1
40 40 80 80 80 2 4 2
40 40 80 80 80 1 2 1
(b) 33 mask
40 40 80 80 80
40 40 80 80 80
(a) Original gray-tone image
DC
22
Example
640 800 1120 1280 1280
40 50 70 80 80
640 800 1120 1280 1280
40 50 70 80 80
640 800 1120 1280 1280 40 50 70 80 80
40 50 70 80 80
640 800 1120 1280 1280 40 50 70 80 80
(d) Normalized result after
division by the sum of the
640 800 1120 1280 1280 weights in the mask (b)
DC
23
Counting Objects
DC
24
Counting Objects
0 0 0 0 1 1 1 1
0 1 1 0 1 0 0 1
1 0 0 1 0 1 1 0
0 0 0 0 1 1 1 1
DC
25
Counting Objects
DC
26
Counting Objects
DC
27
Counting Objects
DC
28
Connected Components
• Given
– A binary image B
– A pair of pixels B[r,c] B[r ' ,c' ] v where v is either 0 or 1
(r0 ,c0 ), (r1, c1 ), , (rn ,cn ) (r',c') is a connected path from (r,c) to (r',c')
DC
29
Connected Components
DC
30
Connected Components
DC
31
Connected Components Labeling
DC
32
Connected Components Labeling
Classical Algorithm
Left Current
DC
33
Connected Components Labeling
Classical Algorithm
Top Top
0 0
Left Current Left Current
0 1 1 1
Top Top
1 1
Left Current Left Current
0 1 1 1
DC
34
Connected Components Labeling
DC
35
Label Collision
1 2 • Labelcollision is a very common
1 2 occurrence – example U-shaped object
DC
36
Classical Algorithm
First Pass
currentlabel=1
for (i=0; i<nrows; i++)
for(j=0; j<ncols; j++)
if (Image[i,j]==0) break;
if ((top==0) && (left==1) label (current) = label(left);
if ((top==1) && (left==0) label (current) = label(top);
if ((top==0) && (left==0) {
label (current) = currentlabel;
currentlabel++;
}
if ((top==1) && (left==1) {
label (current) = label(left);
equivelent(label(left), label(top));
}
DC
37
Classical Algorithm
Second Pass
for (i=0; i<nrows; i++)
for(j=0; j<ncols; j++)
if (Image[i,j]==0) break;
currentlabel=label(i,j);
label(i,j)=equivalentclass(currentlabel);
DC
38
Classical Algorithm
DC
39
Recursive Scanning Method
Begin by selecting a pixel with label 1 (water) and label it as “marked.”
2
2 1 2
2 1 0 1 2
2 1 2
2 DC
42
Distance Measures
2 2 2 2 2
2 1 1 1 2
2 1 0 1 2
2 1 1 1 2
2 2 2 2 2
DC
43
Gray Scale Images
DC
1
Histogram
DC
2
Histogram
Image Histogram
DC
3
Thresholding
DC
4
Thresholding
DC
5
Thresholding
Threshold = 180 DC
6
Image Enhancement
Image enhancement is basically the improving the
perception of information in images for human
viewers and providing better input for the further
image processing techniques.
Techniques:
Spatial domain
Frequency domain
DC
7
Image Enhancement
DC
8
Image Enhancement
DC
9
Image Enhancement
DC
10
Image Enhancement
DC
11
Point Processing Enhancement
DC
12
Point Processing Enhancement
DC
13
Point Processing Enhancement
DC
14
Examples
Show better detail
DC
15
Examples
Show boundary of an object
DC
16
Image Enhancement
DC
17
Image Enhancement
Image Enhancement
DC
20
Histogram Modification
hi (x) hi ( x) and y f ( x) are known. Find ho ( y )
hi ( x)x ho ( y )y. Now as x, y 0
dx 1 d ( f 1 ( y ))
ho ( y ) hi ( x) hi ( f ( y ))
dy dy
x x
x x
x
ho ( y )
y
y y
y f (x)
y y
DC
21
Histogram Modification
Histogram Stretch
Histogram Shrink
Histogram Slide DC
22
Histogram stretching
b) Histogram of
a) Original image original image
c) Image after
histogram stretching
with out clipping
DC
28
Histogram shrink
DC
30
Histogram slide
DC
31
Histogram slide
b) Histogram of
a) Resultant image from sliding the image (a)
histogram down by 50
DC
33
Histogram Equalization
Algorithm :
Step 1 : For a N M image of G gray levels, create an array H of
length G. Initialized with 0 values i.e. H [i ] 0 for i 0,1,.., G.
Step 2 : From image histogram scan every pixel and increment the
relevant number of H - if pixel p has intensity g p , perform
H[g p ] H[g p ] 1
Step 3 : Form the cumulative image histogram H c , H c [0] H [0]
H c [ p ] H c [ p 1] H [ p ], p 1,2,..., G 1
G 1
Step 4 : Set T [ g ] round H c [ g ]
NM
step 5 : Rescan the image and write an output image with gray
levels g q , setting g q T [ g p ].
DC
34
Histogram Equalization
• Consists of four steps:
1. Find the running sum of the histogram
values
2. Normalize the values from step (1) by
dividing by the total number of pixels
3. Multiply the values from step (2) by the
maximum gray level value and round
4. Map the gray level values to the results
from step (3) using a one-to-one
correspondence
DC
35
Histogram Equalization
Example:
3-bits per pixel image – range is 0 to 7.
Given the following histogram:
Number of Pixels
Gray Level Value (Histogram values)
0 10
1 8
2 9
3 2
4 14
5 1
6 5
7 2
DC
36
Histogram Equalization
1) Create a running sum of the histogram values.
This means the first value is 10, the second is
10+8=18, next 10+8+9=27, and so on. Here we
get 10, 18, 27, 29, 43, 44, 49, 51
2) Normalize by dividing by the total number of
pixels. The total number of pixels is:
10+8+9+2+14+1+5+2 = 51 (note this is the last
number from step 1), so we get: 10/51, 18/51,
27/51, 29/51, 43/51, 44/51, 49/51, 51/51
3) Multiply these values by the maximum gray
level values, in this case 7, and then round the
result to the closest integer. After this is done we
obtain: 1, 2, 4, 4, 6, 6, 7, 7
DC
37
Histogram Equalization
4) Map the original values to the results from step
3 by a one-to-one correspondence. This is done
as follows:
Original Gray Histogram
Level Value Equalized Values
0 1
1 2
2 4
3 4
4 6
5 6
6 7
7 7 DC
38
Histogram Equalization Examples
1.
DC
39
Histogram Equalization Examples
2.
DC
41
Linear Contrast Enhancement
y
255
0 x
Bmin Bmax
DC
42
Example: Enhancement with
different transform functions
0.5
Original image f ( x) x
DC
43
Geometric Enhancement
Techniques
DC
44
Geometric Enhancement Techniques
+ + + + + y
+ + + + +
+ + + + +
Pixel (i, j ) x ( x, y )
+ + + + + Image
+ + + + +
Template/M
ask
is the pixel brightness value of the image and t (m, n) is the template
entry at that location.
DC
45
Local Enhancement
g ( x, y ) A( x, y )[ f ( x, y ) m( x, y )] m( x, y )
M
where A( x, y ) k ; 0 k 1
( x, y )
DC
46
Mean Value Smoothing
M N
1
r (i, j )
MN
f (i m, j n)
m 1 n 1
1 M N
Let (i, j)
MN m 1 n 1
f (i m, j n) then
• A noisy image:
g ( x , y ) f ( x, y ) n ( x, y )
• Averaging K different noisy images:
M
1
g ( x, y )
K
g ( x, y )
i 1
i
DC
48
Image Averaging
• As K increases, the variability of the
pixel values at each location decreases.
DC
49
Image Enhancement in the spatial domain
a b
c d
e f
DC
51
Noise Reduction
DC
52
Mean Filter
+ + + + + j
+ + + + +
i 1 1 1
+ + + + +
+ + + + + 1 1 1
+ + + + + 1 1 1
Image
Template/Mask
DC
53
Mean Filter
DC
54
Mean Filter
DC
55
Noise Reduction
DC
56
Noise Removal
Before After
DC
57
General Templates
8-neighbor case
4-neighbor case
DC
58
Removal of Small Components
DC
59
Removal of Small Components
Before After
DC
60
Smoothing
i m j m
In[r i, c j ]
i m j m
Out[r , c]
(2m 1)(2m 1)
2
d ( x xc ) 2 ( y yc ) 2
DC
63
Median Filtering
DC
64
Median Filtering
Before After DC
65
Median Filtering
DC
66
Median Filtering
Before After
DC
67
Median Filter
High-Boost Filtering
DC
71
Frequency Domain Technique
Position invariant operator: h x, y Image: f x, y
g ( x, y ) f ( x, y ) h ( x, y )
Fourier transform of g ( x, y ), h( x, y ) and f ( x, y )
are G (u , v), H (u , v) and F (u , v)
The convolution theorem :
G (u , v) H (u , v) F (u, v)
Then take inverse
g ( x, y ) G 1H (u , v) F (u , v)
DC
72
Frequency Domain Technique
DC
73
Frequency Domain: Lowpass filtering
DC
74
Frequency Domain: Lowpass filtering
Ideal filter:
DC
75
Frequency Domain: Lowpass filtering
H(u,v)
v
1
u v
D(u,v)
u D0
DC
76
Frequency Domain: Lowpass filtering
Butterworth lowpass filter:
DC
77
Frequency Domain: Lowpass filtering
Butterworth lowpass filter:
DC
78
Frequency Domain: High pass filtering
Ideal filter:
DC
79
Frequency Domain: High pass filtering
Ideal filter:
H(u,v)
v 1
u
v
D(u,v)
u D0
DC
80
Frequency Domain: High pass filtering
Ideal filter:
DC
81
Frequency Domain: High pass filtering
Butterworth high pass filter:
DC
82
Frequency Domain: High pass filtering
Butterworth high pass filter:
DC
83
Frequency Domain: Band pass filtering
Band-pass filters reduce signal frequencies outside of a
range (band) of interest.
DC
84
Frequency Domain: Band pass filtering
This example shows two applications of the Difference of
Gaussians approach for band-pass filtering.
DC
85
Two-pass combined method
Published in Defence Science Journal, 2007
ORIGINAL IMAGE
FREQUENCY
DOMAIN SPATIAL DOMAIN
IMAGE ENHANCEMENT
ENHANCED IMAGE
DC
86
Example
DC
87
Example
DC
88
Homomorphic Filtering
DC
89
Homomorphic Filtering
DC
90
Homomorphic Filtering
DC
91
Homomorphic Filtering
DC
92
Homomorphic Filtering
Adaptive Gaussian Filter:
DC
93
Homomorphic Filtering
Human and building thermal IR image: (a) original image, (b) adaptive
histogram equalisation image, (c) Gaussian filter image by 5×5 window
with σ = 1.5, (d) proposed adaptive Gaussian method without
homomorphic filter, (e) homomorphic filter image without adaptive
Gaussian method, and (f) proposed method. DC
94
Homomorphic Filtering
DC
95
Homomorphic Filtering
DC
96
Homomorphic Filtering
DC
97
Color
Image Enhancement
DC
98
Enhancement of Color Images
DC
99
Enhancement of Color Images
Histogram modification can be performed
on color images, but doing it on each color
band separately can create relative color
changes
DC
101
Histogram Equalization of Color
Images (contd.)
Note: In this case the red band gives the best results
This will depend on the image and the desired result DC
102
Histogram Equalization of Color
Images
Typically the most important color band is
selected, and this choice is very much
application-specific and will not always
provide us with the desired result.
DC
104
Image Transformation
DC
1
Mathematical Transformation
• Why
– To obtain a further information from the signal that
is not readily available in the raw signal.
• Raw Signal
– Normally the time-domain signal
• Processed Signal
– A signal that has been "transformed" by any of the
available mathematical transformations
• Fourier Transformation
– The most popular transformation
DC
2
Frequency Analysis
• Frequency Spectrum
– Be basically the frequency components (spectral
components) of that signal
– Show what frequencies exists in the signal
DC
3
Stationarity of Signal
• Stationary Signal
– Signals with frequency content unchanged
in time
– All frequency components exist at all times
• Non-stationary Signal
– Frequency changes in time
– One example: the “Chirp Signal”
DC
4
Stationarity of Signal
2 Hz + 10 Hz + 20Hz 3 6 0 0
2 5 0 0
Magnitude
Magnitude
1 4 0 0
Stationary 0 3 0 0
-1 2 0 0
-2 1 0 0
-3 0
0 0 .2 0 .4 0 .6 0 .8 1 0 5 1 0 1 5 2 0 2 5
0.0-0.4: 2 Hz + 1 2 5 0
Magnitude
0.4-0.7: 10 Hz +
0 .8
0 .6 2 0 0
0.7-1.0: 20Hz
Magnitude
0 .4
0 .2 1 5 0
Non- 0
Stationary -0 .2
-0 .4
1 0 0
-0 .6 5 0
-0 .8
-1 0
0 0 .5 1 0 5 1 0 1 5 2 0 2 5
DC
5
Chirp Signals
Frequency: 2 Hz to 20 Hz Different in Time Domain Frequency: 20 Hz to 2 Hz
1 150 1 150
0.8 0.8
0.6 0.6
Magnitude
Magnitude
Magnitude
100 100
0.2 0.2
0 0
-0.2 -0.2
50 50
-0.4 -0.4
-0.6 -0.6
-0.8 -0.8
-1 0 -1 0
0 0.5 1 0 5 10 15 20 25 0 0.5 1 0 5 10 15 20 25
Fourier
f(x)
Transform
Inverse Fourier
f(x)
Transform
DC
7
Fourier Transform
DC
8
2-D Fourier Transform
DC
9
Discrete Fourier Transform
DC
10
Discrete Fourier Transform
(1)
(2)
DC
11
Discrete Fourier Transform
(3)
(4)
(5)
(6)
DC
12
Discrete Fourier Transform
DC
13
Result Fourier Transform
DC
14
Result Fourier Transform
DC
15
Properties of Fourier Transform
Multiplication by N Multiplication by N
DC
16
Properties of Fourier Transform
DC
17
Properties of Fourier Transform
DC
18
Properties of Fourier Transform
DC
19
Properties of Fourier Transform
DC
20
Properties of Fourier Transform
Graphical
illustration of
convolution.
The shaded
areas indicate
regions where
the product is
not zero
DC
21
Properties of Fourier Transform
DC
22
Properties of Fourier Transform
DC
23
Properties of Fourier Transform
DC
24
Properties of Fourier Transform
DC
25
Morphology
DC
1
Morphology
DC
2
Structuring Elements
DC
3
Common Structuring Elements
DC
4
Operations
• Intersection
A B { p | p A and p B}
• Union
A B { p | p A or p B}
• Complement
A { p | p and p A}
Ω is a universal binary image (all 1s)
DC
5
Translation
• A : binary image
• p : pixel
• Translation of A by p is an image
• Shifts the origin of A to p
Ap {a p | a A}
DC
6
Binary Dilation
A (a1 , a2 ,..., a N ) E N and B (b1 , b2 ,..., bN ) E N then Dilation is defined as
A B c E N | c a b for some a A and b B
B
A A B
Here A {(0,1), (1,1), (2,1), (2,2), (3,0)}, B {(0,0), (0,1)}
Then A B {(0,1), (1,1), (2,1), (2,2), (3,0), (0,2), (1,2), (2,2), (2,3), (3,1)}
DC
7
Properties of Dilation
1. A B B A (Commutative)
2. ( A B ) C A ( B C ) (Associative)
3. A B Ab
bB
4. ( A) x B ( A B ) x
5. [ A B1 ... ( Bm ) x ... Bn ] ( A B1 ... Bn ) x
6. ( A) x ( B ) x A B
7. A B A D B D
8. ( A B ) C ( A C ) ( B C )
9. A ( B C ) ( A B ) ( A C )
10. ( A B) C ( A C ) ( B C ) DC
8
Examples
(b) Structuring
element
(a) Original (c) Dilation
A B DC
9
Dilation
DC
10
Binary Erosion
A (a1 , a2 ,..., a N ) E N and B (b1 , b2 ,..., bN ) E N then Erosion is defined as
AB x E N | x b A for every b B
B
A AB
Here A {(1,0), (1,1), (1,2), (1,3), (1,4), (1,5), (2,1), (3,1), (4,1), (5,1)},
B {(0,0), (0,1)} Then AB {(1,0), (1,1), (1,2), (1,3), (1,4)} DC
11
Properties of Erosion
1. AB Ab
bB
DC
12
Erosion
(b) Structuring
element
(a) Original (c) Erosion
DC
13
Erosion and Dilation
DC
14
Complement of A and Reflection of B
A B' DC
15
'
AB
DC
16
A B '
DC
17
Opening
DC
18
Example
(b) Structuring
element
(a) Original (c) Opening
A B DC
19
Example
(b)
(a) (c)
A B ( A B)B
• Fill holes smaller than the structuring
element
DC
22
Example
(b) Structuring
element
(a) Original (c) Closing
A B DC
23
Example
DC
26
HIT and MISS Transformation
Let A is a set and B is a structuring element, which is composed of X
and background. The Hit and Miss transformation of A by B is
A * B ( AX ) [ A(W X )] where X is a set and W is a small window.
We can generalize, A * B ( AB1 ) ( AB2 )
Example: Objective is to find the location of one of the shapes, say X
A X Y Z
Origin
Z
W W-X
(b) Local
background of X
Y with respect to W
X
(a)
DC
27
AX
A
(c) (d)
AW X
AX A[W X ]
(e) (f)
DC
28
Boundary Extraction
( A) A AB
Origin
(b) B
(a) A
DC
30
Region Filling
Example
Origin
(b) B
(a) A (a) A
DC
31
(d) X 0 (e) X1 (f) X2
DC
33
Road Detection from Remote Sensing
Images (<2m)
DC
34
.
.
Road Detection from Remote Sensing Images (<2m)
Enhancement technique: For each pixel (x, y) in the image f
1 5
1. Compute the mean gray value, d ( x, y ) | f d ( x, y )[i ] S d [i ] |, d 1,..., 4
5 i 1
5
1
2. Compute the standard deviation, (x, y)
d
5 i1
( fd (x, y)[i] Sd[i] d (x, y)[i])2 , d 1,...,4
15
7. Compute the standard deviation, d (x, y) (gd (x, y)[i] Sd [i] d (x, y)[i])2 , d 1,...,4
5 i1
(a) (b)
DC
36
Build-Up area Analysis: Enhancement
DC
37
Shape Features
DC
1
Shape Features
• Organization of points in space
• Not related to any intensity or color features
• Common Features
– Centroid
– Area
– Perimeter
– Radius
– Orientation
– Major and Minor Axes
– Elongatedness
– Aspect Ratio
– Bounding box
– Moment
– Compactness
– Complexity of object (No. of perimeter pixels/No. of object pixels) DC
2
Centroid
N
1
x
N
x
i 1
i
N
1
y
N
y
i 1
i
DC
3
Area
• Number of pixels in a component
• Gives an indication of size
• Easy to compute
A 1
( x , y )R
where R is a given region
DC
4
Perimeter
• Boundary points of the object
• Length of the perimeter is the number of
pixels on the boundary of the object
• Length of the perimeter is an important
measure of the shape
• First order of approximation
– Number of pixels on the boundary
• More accurate approximation
– Horizontal and vertical moves : 1 unit
– Diagonal move: 1.4 unit DC
5
Circularity
• Degree of similarity with a circle
2
P
Circularity
A
• Based on continuous planer shapes
• Make adjustments for digital shapes
DC
6
Circularity
R
Circularity
R
R :Mean radial distance
r :Standard deviation of radial distance
• Radial distance computed as the distance
between the centroid and a boundary point
DC
7
Circularity
1 k 1
R ( xi , yi ) ( x, y )
k i 0
( x , y ) ( x, y )
1
1 k 1 2
2
r i i R
k i 0
DC
9
Compactness
DC
10
Compactness
Calculate the "compactness" of each of the following
'islands' using the above definition
DC
11
Compactness
Solution: The results for area ¸ perimeter for the six shapes are shown
below:
Shape A B C D E F
Area 12 28.26 3 16 16 4
Perimeter 16 18.84 8 16 32 8
(cm)
Area ¸ 0.75 1.5 0.375 1 0.5 0.5
Perimeter
Staircase A B C D E F
2
(Re gion _ border _ length)
Compactness
Area
DC
14
Compactness
Proposed Method: Published in DSJ, 2008
Center
(a) Object
(b) Fitted Circle
Matching_Area
Non_Matching_area_due_to_fitted_circle
DC
16
Result of Proposed definition of general
measure for rectangle, circle and ellipse
DC
18
Euler’s Number
Connected Component: A connected component of a set is a
subset of maximal size such that any two of its points can be
joined by a connected curve lying entirely within the subset.
Euler Number E C H
C - Connected component
H Number of holes
DC
20
Moments
M pq x p q
y f ( x , y ) dx dy
x y
where
1 if ( x, y ) is in the object
f ( x, y )
0 otherwise
DC
21
Moments
• Defines moments of order p+q
• Changes with location of the object
• Discrete case
M pq x y f ( x, y )
p q
x y
DC
22
Moments
M 00 x 0 y 0 f ( x, y ) f ( x, y ) Area of the object
x y x y
M 10 M 01
x y
M 00 M 00
DC
23
Central Moments
• Moments change with the location of the object
• Not useful for comparing shapes
pq ( x x ) p ( y y ) q f ( x, y )
x y
10 ( x x )1 ( y y )0 f ( x, y )
x y
xf ( x, y ) x f ( x, y ) 00 f ( x, y ) Area
x y x y x y
2
30 M 30 3x M 20 2 x M10
12 M12 2 yM11 x M 02 2 y 2 M10
21 M 21 2 x M11 yM 20 2 x 2 M 01
2
03 M 03 3 yM 02 2 y M 01
DC
25
Central Moments
• Centroid forms the origin
1 1 2M 11
• Orientation of on object is given by tan ( )
2 M 20 M 02
Alternate Method: (D.Chaudhuri and A. Samal in Pattern
Recognition,Vol. 40, pp. 1981-1989, 2007)
( xib , yib ), i 1,2,..., n Border points of the object. Angle of the major
axis with horizontal axis. ( xc , yc ) Centriod of the object. Then equation of the
line passing through centroid and makes an angle be tan x y yc tan xc 0.
Now the perpendicular distance from the border points
pi ( xib xc ) sin ( yib yc ) cos , i 1,2,.., n Sum square perpendicular distance
2
P xib xc sin yib yc cos
n
i 1
2 xib xc yib yc
n
P
Now for minimum 0 tan 2 i 1
x x y y
n
b 2 b 2
i c i c
i 1 DC
26
Moments
• Row moments
1
pp ( x x ) f ( x, y)
p
Ax
• Column moments
1
pp ( x y) p f ( x, y)
Ay
• Eccentricity (elongatness)
1
(M 20 M 02 4M11
) 2
e
M 00
DC
27
Elongatedness
Elongatedness is a ratio between the length and width of the region bounding
rectangle. This is the rectangle of minimum area that bounds the shape.
Alternate Method:
• Elongatedness can be evaluated as a ratio of the region area and the square
of its thickness.
area
elongatedness 2
2d DC
28
Elongatedness
Another Method: (D. Chaudhuri, IEEE Trans. on Systems, Man and Cybernetics,
Vol. 27, No. 5, pp. 871-877, 1997)
DC
29
Bounding Box
DC
30
Bounding Box
• Gives a rough idea of the location of the
object
• Smallest rectangle to enclose the shape
• Extremal points
– Leftmost, rightmost, topmost and bottommost
– Lie on the bounding box boundary
– May have up to 8 extremal points
– Occur in opposite pairs
DC
31
Bounding Box
• Each point of opposite extremal point pairs
define an axis
• Properties of axes are useful
– Lengths
– Orientations
• Need to make corrections to the lengths
– Error due to sampling
DC
32
Bounding Rectangle
Orientation of on object is given by
1 2 M 11
tan 1 ( )
2 M 20 M 02
Defining ( x, y ) x cos y sin ( x, y ) x sin y cos
search for the minimum and maximum of and over all
boundary points ( x, y ). The values of min , max , min , max
then define the bounding rectangle. l1 ( max min )
and l2 ( max min ) are its length and width.
b
Bounding rectangle can not
represent elogatedness DC
33
Bounding Rectangle
New Proposed Method: (D.Chaudhuri and A. Samal in Pattern
Recognition,Vol. 40, pp. 1981-1989, 2007) n
2 xib xc yib yc
i 1
tan 2
2 yib yc 2
n
xib xc
• Find Centroid and orientation of the object. i 1
(a) Example (b) Centroid and board of (c) Major and Minor
object axes
DC
36
Bounding Rectangle
(a) (b)
DC
37
Best-fitted Rectangle
Bisection method for finding best-fitted rectangle based on area minimization
(Published in Machine Vision and Applications, Springer, 23(6), 2011.)
1. First find the minimum bounding rectangle (by our previous method), which
we call Upper Estimated Rectangle (UPER)
2. Find Under Estimated Rectangle (UNER) by least square method.
3. Find the optimal rectangle by bisection method using corresponding vertices
of UPER and UNER.
H G
L K
D C
A B
I J
E F
DC
41
Optimal Rectangle Fit
DC
42
Optimal Rectangle Fit
DC
43
Optimal Rectangle Fit
DC
44
Optimal Rectangle Fit
DC
45
Circular Fit
• In the literature various algorithms are available. Journal: Pattern
Recognition, Pattern Recognition Letters, IEEE Trans. on System, Man and
Cybernetics, IEEE Trans. on Image Processing etc.
Assume that the area of the object and fitting circle area are same.
Let A be the area of the object and r be the radius of the fitting circle.
A
Therefore, A r r
2
1 n 1 n
The centroid is xc xi and yc yi , where n is the number of
n i 1 n i 1
border points of the object and xi , yi be the position of those border
points of the object.
DC
46
Circular Fit
Published in CVGIP
xi , yi be the border points of the object, i 1,2,.., n. Let R
be the raidus and x, y be the center of the fitting circle. Now
find R and x, y ?
n 2
i 1
n 2
e( R, x, y )
R 2 xi x yi y
2 2
J ( R , x, y )
2
i 1
DC
47
Circular Fit
Other proposed Method: (Published in Pattern Recognition Letters, 20010)
1 n 1 n
1. Compute centroid x xi ; y yi ; where ( xi , yi ) are the
n i 1 n i 1
border points of the object
n
3. Total error for all border points will be E xi 2 yi 2 r 2
i 1
4.
E
r
1 n 2
0 gives r xi yi 2
2
n i 1
DC
48
Circular Fit
DC
49
Circular Fit
DC
50
Circular Fit
DC
51
Elliptic Fit
Proposed Method: (Published in Defence Science Journal, Sep 2008)
xi yi are n border points of the object. Let A be the area of the object.
1 n 1 n
Centroid x, y xi , yi . Let be angle of major axis i.e.
n i 1 n i 1
n
2 xi x yi y
orientation of the object. tan 2 i 1
and
x x y
n n
2 2
i i y
i 1 i 1
n
2
n n
xi x yi y cos 2 2 xi x yi y sin 2 0
2
i 1 i 1 i 1
Now the area of the fitted ellipse is ab, where a and b are major
and minor axes. Now the equation of the ellipse is
x2 y2
2
2 1, with center at (0,0). We assume that A ab. Now
a b
A
b . Putting the value b in equation of ellipse we get
a
2 x 2b 4 A2b 2 A2 y 2 0. Find b 2 and correspondingly find a. DC
52
Elliptic Fit
Other proposed Method: (Published in Pattern Recognition Letters, 2010)
1 n 1 n
1. Find centroid of the object : x xi ; y yi
n i 1 n i 1
where ( xi , yi ) are n border points of the object
n
2 xi x yi y
i 1
2. Find orientation of the object : tan 2
x x
n
i
2
yi y 2
i 1
x2 y 2
3. Equation of the ellipse origin at (0,0) : 2 2 1
a b
xi n 2 2 2
yi
4. Total error for all border points : E 1 2 2
i 1 a b
DC
53
Elliptic Fit
S S
5. 0 and 0 gives
a b
2
n n n
2 2
x i yi 4 4
x i y i
i 1 i 1 i 1
a2 n n n n
xi 2
y i xi yi
4 2 2
i
y 2
i 1 i 1 i 1 i 1
2
n n n
2 2
x i yi 4 4
x i y i
i 1 i 1 i 1
b2 n n n n
xi 4
yi xi yi
2 2 2
i
x 2
i 1 i 1 i 1 i 1
DC
54
Elliptic Fit
DC
55
Elliptic Fit
DC
56
Edge Detection
DC
1
Importance
DC
2
Edge Basics
DC
3
Gradients
• Intensity discontinuity in the image
• Attributes
– Location (x,y)
– Magnitude
– Orientation
f
G x x
G[ f ( x, y )] f
G y
y
DC
4
Gradients
f f
1 2
x y
1 2
Orientation of the edge : tan ( )
1
DC
6
Edge Operators
• Discrete Case
f f ( x x, y ) f ( x, y )
1 lim
x x 0 x
f f ( x, y y ) f ( x, y )
2 lim
y y 0 y
1 f ( x 1, y ) f ( x, y )
2 f ( x, y 1) f ( x, y )
DC
7
Edge Operators
Three techniques for detecting edges:
• By subtracting a smoothed image from its original.
• By calculating spatial derivatives.
• By using edge detecting template operator.
j
Pixel values of a sub-area of an image
i
w1 w2 w3 i -1, j -1 i -1, j i -1, j +1
Z1 Z2 Z3
w4 w5 w6 i , j -1 i,j i , j +1
Z4 Z5 Z6
w7 w8 w9 i +1, j -1 i +1, j i +1, j +1
Z7 Z8 Z9
Template weights 8-neighbors of (i , j ) pixel
9
Image
Response : Z wi Z i
i 1
DC
8
Edge Operators
The gradient of f ( x, y ) at co - ordinate ( x, y ) is defined as the vector
2 2
f / x f f
f and f
f / y x y
The magnitude can be approximated at point Z 5 in a number of ways :
f Z 5 Z8 2 Z 5 Z 6 2 Also f Z5 Z8 Z5 Z 6
1 0 0 1
0 -1 -1 0
DC
9
Edge Operators
• Gradient is estimated in eight direction for a 3 3 mask.
• The convolution result of greatest magnitude indicates the gradient
direction.
Sobel Derivative Operator:
f ( Z 7 2 Z 8 Z 9 ) ( Z 1 2 Z 2 Z 3 ) ( Z 3 2 Z 6 Z 9 ) ( Z 1 2 Z 4 Z 7 )
Sobel Template Operator:
1 2 1 -1 -2 -1 1 0 -1 -1 0 1
0 0 0 0 0 0 2 0 -2 -2 0 2
-1 -2 -1 1 2 1 1 0 -1 -1 0 1
0 1 2 0 -1 -2 2 1 0 -2 -1 0
-1 0 1 1 0 -1 1 0 -1 -1 0 1
-2 -1 0 2 1 0 0 -1 -2 0 1 2
DC
10
Edge Operators
Prewitt Derivative Operator:
f ( Z 7 Z 8 Z 9 ) ( Z 1 Z 2 Z 3 ) ( Z 3 Z 6 Z 9 ) ( Z 1 Z 4 Z 7 )
Prewitt Template Operator:
1 1 1 -1 -1 -1 1 0 -1 -1 0 1
0 0 0 0 0 0 1 0 -1 -1 0 1
-1 -1 -1 1 1 1 1 0 -1 -1 0 1
0 1 1 0 -1 -1 1 1 0 -1 -1 0
-1 0 1 1 0 -1 1 0 -1 -1 0 1
-1 -1 0 1 1 0 0 -1 -1 0 1 1
DC
11
Edge Operators
Robinson Template Operator:
1 1 1 -1 -1 -1 1 1 -1 -1 1 1
1 -2 1 1 -2 1 1 -2 -1 -1 -2 1
-1 -1 -1 1 1 1 1 1 -1 -1 1 1
1 1 1 1 -1 -1 1 1 1 -1 -1 1
-1 -2 1 1 -2 -1 1 -2 -1 -1 -2 1
-1 -1 1 1 1 1 1 -1 -1 1 1 1
DC
12
Edge Operators
Kirsch Template Operator:
3 3 3 -5 -5 -5 3 3 -5 -5 3 3
3 0 3 3 0 3 3 0 -5 -5 0 3
-5 -5 -5 3 3 3 3 3 -5 -5 3 3
3 3 3 3 -5 -5 3 3 3 -5 -5 3
-5 0 3 3 0 -5 3 0 -5 -5 0 3
-5 -5 3 3 3 3 3 -5 -5 3 3 3
DC
13
Edge Operators
1 2 1 1 0 -1
0 0 0 2 0 -2
-1 -2 -1 1 0 -1
0 -1 2 2 -1 0
1 0 -1 -1 0 1
-2 1 0 0 1 -2
DC
14
Laplacian Edge Operator
• Mathematical approximation
• Gives only magnitude
• No orientation
f f 2 2
L ( x, y ) 2 2
x y
DC
15
Laplacian Edge Operator
-1 0 -1 -1 -1 1
0 4 0 -1 8 -1
-1 0 -1 -1 -1 1
DC
16
Finding Boundaries
DC
17
Problem in Binary Image Edge
Operation
(a) Two circular shape objects (b) Edges of these circles are
one pixel apart from each other connected after edge operation
DC
18
Problem in Binary Image Edge Operation
Our proposed method: (published in Pattern Recognition, 2007)
We assume that the center pixel in case is a 1 and the other pixels
in the 22-sub-template may appear in any combinations of 0 and
1. Four possible types of configurations are possible in these sub-
templates.
Type 1: All three-pixel values are 0.
Type 2: Two of them are 0 and one pixel value is 1.
Type 3: Two of them are 1 and one pixel value is 0.
Type 4: All three-pixel values are 1. DC
19
Problem in Binary Image Edge Operation
Our proposed method: (published in Pattern Recognition, 2007)
Region of interest
Region of interest
DC
22
Problem in Binary Image Edge Operation
DC
23
Edge thresholding
DC
24
Notion of edge
Intensity function Sharp Change
The profile that an edge (transition from dark to
light) is modeled as a smooth, rather than as an
abrupt, change of gray value
DC
25
Gaussian Mask
We know the equation of Gaussian function is :
DC
26
Gaussian Mask
DC
29
Gaussian Mask
DC
30
Gaussian Mask
DC
32
Marr-Hildreth Operator
• Based on the zero-crossing of the second
derivative
• Agree with the fact that edge corresponds to
an abrupt change in the image function
• First derivative of the image should have an
extremum at edge position in the image
• Second derivative should be zero at the same
position
• Much easier to find a zero-crossing position
than an extremum
DC
33
Marr-Hildreth Operator
DC
34
Marr-Hildreth Operator
DC
35
Convolution
Continuous Case
h ( x, y ) f ( x , y ) * g ( x , y )
h ( x, y ) f ( x' , y ' ) g ( x x' , y y ' )dx' dy '
Discrete Case
n m
h(i, j ) f (k , l )g (i k , j l )
k 1 l 1
DC
36
Convolution
h(i, j ) Ap1 Bp2 Cp3 Dp4 Ep5 Fp6 Gp7 Hp8 Ip9
DC
37
Marr Hildreth Operator
Step 1 : I G * I
'
x 2 y2
-
G ( x, y ) e 2 2
2 2
2 2
2
x y
I I (G * I ) ( G ) * I
'' 2 ' 2 2
DC
38
Marr Hildreth Operator
r2 r2
2 1 2
Let r x y so G (r ) e
2 2 2 2
Now G(r ) re 2
2
r2
1 r 2 2
and the second derivative G(r ) 2 2 1e 2
Now returning to the original co - ordinates x, y and introducing
a normalizing multiplicative coefficient c, we get a convolution
mask of a LoG operator :
x2 y 2
x y
2 2 2
h( x, y ) c e 2 2
, where c normalizes the sum of
4
mask elements to zero.
DC
39
G 2
17X17 Mask
DC
41
Laplacian of Gaussian Mask
5X5 Mask DC
42
Blurring
+ve: white
-ve: black
+ve: white
-ve: black
DC
46
Canny Edge Detection
1. Smooth the image with a Gaussian filter to reduce noise and unwanted
details and textures g (m, n) G (m, n) * f (m, n) where
1 m2 n2
G exp 2
2 2
2. Compute gradient of g (m, n) using any of the opertors (Roberts, Sobel) of
get M (m, n) g m2 (m, n) g n2 (m, n) and (m, n) tan 1[ g n (m, n) / g m (m, n)]
M (m, n) if M (m, n) T
3. Threshold M : M T (m, n)
0 otherwise
where T is so chosen that all edge elements are kept while most of the
noise is supressed
DC
47
Canny Edge Detection
4. Supress non - maximum pixels in the edges in M T obtained above to thin
the edge ridges. To do so, check to see whether each non - zero M T (m, n) is
greater than its two neighbors along the gradient direction (m, n). If so, keep
M T (m, n) unchanged, otherwise, set it to 0
DC
48
Canny Edge Detection
Non-Max Suppression:
Consider the three edge points. Assume point (x,y) as the
point having the largest gradient of edge.
Check for the edge points in the direction perpendicular to
the edge and verify if their gradient is less than (x,y).
If the values are less than (x,y) gradient, then we can
suppress those non-maxima points along the curve.
DC
49
Canny Edge Detection
Hysteresis Thresholding:
If the gradient at a pixel is :
– Above “High” declare it as an ‘edge pixel.’
– Below, “Low” declares it as a ‘non-edge pixel.’
– Between “low” and “high.”
•Consider its neighbors iteratively then declare it an “edge
pixel” if it’s connected to an “edge pixel” or via pixels
between “low” and “high.”
DC
50
Canny Edge Detection
DC
51
Canny Edge Detection
(a) (b)
(c)
(d)
DC
53
Problems with Edge Detection
• Arbitrary thresholding
• Thick edges
– Use Edge thinning
• Broken edges
– Use some parametric methods
• False edges
– Filtering
DC
54
Segmentation
DC
1
Segmentation
DC
2
Goals
• Change of representation
– Pixels to regions
– Higher level units
DC
3
Bottom Up Approaches
DC
4
Histogram Based methods
T T1 T2
g (i, j )
N for T f (i, j ) T
N 1 N
0 Otherwise DC
7
Iterative Threshold Selection
DC
8
Automatic thresholding
{(r , c) | Image(r , c) i}
P(i )
RC
DC
10
Automatic thresholding
First Method
The mixture probability density function is given by
P( x ) P1 p1 ( x ) P2 p 2 ( x )
For Gaussian case
P1 ( x 1 ) 2 P2 (x 2 )2
P( x ) exp exp
2 1 2 1 2 2 2 2
2 2
P1 and P2 are the a priori probabilities of the two levels
P1 P2 1
Let us assume that the background represents as dark region and
object represent as bright region. So in this case 1<2 and we may
define a threshold, T, such that all pixels with gray levels below T are
considered background and the opposite as object DC
11
Automatic thresholding
The probability of classifying an object point as a background point is
T
E1 (T ) p
2 ( x )dx
P1 p1 (T ) P2 p 2 (T )
DC
12
Automatic thresholding
Now putting
1 (T 1 ) 2
p1 (T ) exp and
2 1 2 1
2
1 (T 2 ) 2
p 2 (T ) exp
2 2 2 2
2
Taking log we get
AT BT C 0
2
B 2( 1 2 2 1 )
2 2
A 1 2
2 2
where
2 P1
C 1 2 2 1 2 1 2
2 2 2 2 2 2
ln
1 P2
DC
13
Automatic thresholding
If the variance are equal, a single threshold is sufficient:
1 2 2
P2
T ln
2 1 2 P1
1 2
T
2
DC
14
Automatic thresholding
Otsu’s Method:
Let [1,2,.....,L], L gray levels in the image. The number of pixels at level i is
denoted by ni and the total number of pixels by N =n1+n2+......+nL.
ni L
pi , pi 0 pi 1
N i 1
Now suppose there are two classes C0 and C1. By a threshold at level k C0
contains pixels with levels [1,.......,k] and C1 contains pixels with levels
[k+1,.......,L]. Then the probability of class occurrence and the class mean
levels, respectively, are given by,
k
0 P (C0 ) pi (k ),
i 1
L
1 P (C1 ) pi 1 (k )
i k 1 DC
15
Automatic thresholding
k k
(k )
0 iP (i | C0 ) ipi / 0 ,
i 1 i 1 (k )
L T (k )
L
1 iP(i | C1 ) ip i 1
i k 1 i k 1 1 (k )
k L
p, T ipi
k
where (k ) i ( k ) ip i and
i 1 i 1 i 1
0 0 1 1 T , 0 1 1
DC
16
Automatic thresholding
In order to evaluate the goodness of the threshold (at level k),
we shall introduce the following discriminant criterion
measures (or measure of class separatibility) used in the
discriminant analysis:
B 2
T B2 2
2 , 2 , 2
w w T
where w 2 0 0 2 1 12 ,
L
T 2
(i T ) pi
2
and
i 1
B 0 ( 0 T ) 1 ( 1 T ) 0 1 ( 1 0 )
2 2 2 2
DC
17
Automatic thresholding
The class variances are
k k
0 (i 0 ) P ( i | C 0 ) (i 0 ) 2 p i 0
2 2
i 1 i 1
L L
1
2
(i 1 ) P (i | C 1 )
2
(i 1 ) 2 pi 1
i k 1 i k 1
B ( k *) max B ( k )
2 2
1 k L DC
18
Automatic thresholding
Fig: (a) Original ENVISAT SAR image. (b) Segmentation of original image using single threshold. (c)
Segmentation of original image using the proposed segmentation technique. (d) Features enhanced
image. (e) Segmentation of enhanced image using a single threshold. (f) Segmentation of enhanced D C
image using proposed technique. 19
Automatic thresholding
(a) Original RADARSAT image. (b) Features enhanced image. (c) segmented image using
proposed algorithm. (d) Thinned image. (e) Image after branch removal. (f) Final linked image. DC
20
Ohlander’s Approach
• Recursive algorithm
• Compute the modes of the histogram for
the whole image
– Derive the clusters
• Compute the histogram for the clusters
– Derive the sub-clusters if needed
• Histogram has multiple modes
• Use a mask to determine the region of
interest
DC
21
Region Based Approach
Regions:
• Each region should be uniform, and
homogeneous with respect to some
characteristic
– Color, gray scale, texture etc
• Region interiors should be simple without many
small holes
• Adjacent regions should have significantly
different values/characteristics
• Boundaries of the regions should be smooth
• Regions should cover the image
DC
22
Region Based Segmentation
The objective of region based segmentation is a partition an
image into regions
Basic Formulation:
0 0 5 6 7 A A B B B
1 1 5 8 7 A A B B B
0 1 6 7 7 A A B B B
2 0 7 6 6 A A B B B
0 1 5 6 5 A A B B B
A A A A A
A A A A A
A A A A A
A A A A A
A A A A A
(c) At T = 8 DC
24
Region Growing Segmentation
Region Growing:
DC
26
Region Growing Segmentation
Select the initial
seed point
Neighbors are 8-
Select next
neighbors of seed
seed pixel
pixel
Satisfying pixels
Compare neighbors to
are assigned to
the seed pixel with
the region and are
homogeneity criterion
the new neighbors
DC
28
Region Growing Segmentation
DC
29
Region Growing Segmentation
DC
30
Region-Oriented Segmentation
DC
31
Region-Oriented Segmentation
DC
32
Double Thresholding for Region
Growing
• Select two thresholds T1 and T2.
• Partition the image into three regions based on
these threshold values.
- R1 containing all pixels with gray values
below T1. R2 containing all pixels with gray
values between T1 and T2. R3 containing all
pixels with gray values above T2.
• Visit each pixel assigned to region R2. If the pixel
has a neighbor in region R1, then reassign the
pixel to region R1.
• Repeat step 3 until no pixels are reassigned
• Reassign any pixels left in region R2 to region R3 DC
33
Region Based Segmentation
Region Splitting and Merging
R
R21 R22
R1 R R2R
23 24 R1 R2 R3 R4
R3 R4
R21 R22 R23 R24
(a) Partitioned Image (b) Corresponding Quad tree
Region Merging:
DC
36
Edge Based Segmentation
DC
37
Edge Based Segmentation
Edge detecting an image significantly reduces the
amount of data and filters out useless information, while
preserving the important structural properties in an
image.
DC
38
Edge Based Segmentation
DC
39
Edge Based Segmentation
Edge Relaxation:
A
D
DC
41
Edge Based Segmentation
Example
C
A
B
(a) First Step (b) Second Step
DC
43
Boundary Detection
Approach
Segmentation : Find the pixels (inside) which
belong to an object
Approach 2
Directly find the regions from the intensity image
(Canny edge detector and linker)
Approach 3
Link edges to find the boundary points of regions
DC
45
Fitting Models to Segments
Least Square Method
• For determining the parameters of the best
mathematical model fitting observed data
j 1
Max-Error Criteria:
n
MAXE max f ( xi ) yi . NOTE : this measure depends only on the
i 1
worst fit point, but RMS depends on the fit of all of the points. DC
47
Fitting a Straight Line
y y Circular fit is
unbiased
Model is y c1 x c0
Linear fit is
Error is e j y y j
biased
0 x
0 x
DC
48
Fitting a Straight Line
n
LSE E (c1 ,c 0 ) (c1 xi c0 yi ) 2 : Two parameters (c0 , c1 )
i 1
E E
Best fit means error will be minimum i.e. 0 and 0
c0 c1
E n n
Now 0 xi c1 nc0 yi 0
c 0 i 1 i 1
E n 2 n n
and 0 xi c1 xi c0 xi yi 0
c1 i 1 i 1 i 1
By solving we get
n n n n n n n
i yi xi xi yi
x 2
n xi yi xi yi
c0 i 1 i 1 i 1 i 1
2
and c1 i 1 i 1 i 1
2
n
n
n
n
n xi2 xi n xi2 xi
i 1 i 1 i 1 i 1 DC
49
Split-and-merge technique
Result
DC
50
Split-and-merge technique
Result
DC
52
Bi-Entropy Segmentation
DC
53
Bi-Entropy Segmentation
DC
54
Bi-Entropy Segmentation
a. QuickBird image, b
histogram, c by Liscano and
Wong, d Chaudhuri and
Agarwal, e Arifin and Asano, f K-
means algorithm, g Otsu
method, h Kapur et al., i the
proposed algorithm, j K. Vennila
DC
et al. and k Suresh Manic et al 55
Bi-Entropy Segmentation
DC
1
Pattern Recognition
DC
2
Pattern Recognition
Applications
Automatic inspection
Speech recognition
Document processing (character recognition)
Medical diagnostics
Remote Sensing
Computer vision
Personal identity verification, etc
DC
3
Model for Classification
DC
4
PR Model
DC
5
PR General Model
Measurement Feature
Space Space
M FD Decision
Space
DC
6
Feature Extractor
DC
7
Feature Extraction
Binary (Shape)
Feature Extraction
Preprocessing
Sensor for each channel
Color Feature
Problem Extraction
Domain DC
8
Shape Features
• Organization of points in space
• Not related to any intensity or color features
• Common Features
– Centroid
– Area
– Perimeter
– Radius
– Orientation
– Major and Minor Axes
– Elongatedness
– Aspect Ratio
– Bounding box
– Moment
– Compactness
– Complexity of object (No. of perimeter pixels/No. of object pixels) DC
9
Gray level & Statistical Features
• Common Features
– Maximum gray value of object
– Minimum gray value of object
– Mean gray value of object
– Standard deviation of object gray value
– Mean gray level in background
– Difference of mean gray values of object and background
– Contrast ratio
– Bright pixel ratio (No. of pixels within 10% of max
gray/No. of object pixels)
DC
10
Texture Based Features
What is Texture?
• No formal definition of texture exists
• Visual texture is something that everyone
understands, but it is difficult to define
- “Despite its importance in image data, a formal approach or precise
definition of texture does not exist.” (Haralick, 1979)
DC
12
Why should we study Texture?
• Texture analysis has been used in various
application:
- Industrial Inspection: Automated inspection problems, applications
include defect detection in images of textiles and automated inspection of
carpet wear and automobile paints
Fig: Texture pairs with identical second order statistics. The bottom halves of
the images consist of texture tokens that are different from the ones in the
top half. (a) Humans cannot perceive the two regions without careful
scrutiny, (b) The two different regions are immediately discriminable by
humans. DC
13
Applications
DC
14
Applications
DC
15
Applications
Medical Image Analysis
The processing of the ultrasound images of the heart using textural features.
(a) A 128 432 ultrasound image from the left ventricle of the heart. (b) The
fractal dimension used as the texture feature. (c) Image segmentation from a k-
means clustering algorithm. The white region is cluster 2 which corresponds to
the blood. DC
16
Applications
Document Processing
(a) (b)
(a) And (b) are examples of aerial images using SAR sensors
DC
20
Modeling Texture
• Structural Approaches
– Organize texture as a repeated pattern of a primitive
element (texel). Deal with the arrangement of image
primitives, such as the description of texture based on
regularly spaced parallel lines.
• Statistical Approaches
– View texture as a quantitative measure of the
arrangement of the intensity. Characterized of textures as
smooth, coarse, grainy and so on.
• Syntactic Approaches
– Analyze texture as a grammar of primitive elements and
their organization
DC
21
Statistical Approaches
Gray Level Co-occurrence Matrices (GLCM)
DC
22
Order of Texture
DC
23
Framework for the GLCM
0 0 1 1
Example: Test image gray levels
0 0 1 1
0 2 2 2
2 2 3 3
Framework matrix:
DC
24
Framework for the GLCM
Fill in the matrix framework for the east spatial relationship:
1350 (NW) 900 (N) 450 (NE)
2 2 1 0
0 2 0 0 00 (W) 00 (E)
0 0 3 1
0 0 0 1 450 (SW) 1350 (SE)
900 (S)
DC
26
Framework for the GLCM
Expressing the GLCM as a probability
Vi , j
Normalization equation : Pi , j N 1
where Vi , j is the
V
i , j 0
i, j
j 0 i 0 y 2 j y 2 Pi, j; d ,
j 0 i 0
1. Con(d , ) Contrast
L -1 L 1 2
i j Pi, j; d ,
i 0 j 0
2. E (d , ) Energy
L -1 L 1 2
Pi, j; d ,
i 0 j 0
DC
28
Texture Parameters
3. H (d , ) Entropy
L -1 L 1
- P i, j; d , logP i, j; d ,
i 0 j 0
4. C d , Correlation
L 1 L 1
i j Pi, j; d ,
i 0 j 0
x y
x y
5. Ld , Local Homogeneity
L -1 L -1
1
Pi, j; d , It should be
j 0 1 i - j
2
i 0
noted that most
6. SE Sum Entopy - p sum k log psum k cases d = 1 and
k
= 00, 450, 900 and
where psum k Pi, j; d , with i j k
i j
1350.
DC
29
Texture Segmentation
• Use the texture energy measures to
– Segment
– Cluster
DC
30
Classifier
DC
31
Classifier
• Number of classes : m
• Size of the feature vector : d
• K : Knowledge about the class
• Classification
– Using nearest mean
– Maximum a posteriori probability
– Feed forward artificial neural network
DC
32
Pattern Matching (by Minimum
Distance Classifier)
M number of class : Wi , i 1,..., M and N i is
the number of pattern vector from class Wi
d i ( X ) X mi , i 1,..., M
1
where mi
Ni
X , i 1,2,..., M
X Wi
Euclidean distance : d i ( X ) X mi X mi
2 T
X X 2 m X m mi
T T
i
T
i
2
Classification : X Wi if d i ( X ) d j ( X ) for j i
2
DC
33
Pattern Matching (by Correlation)
y
N
x Origin t
K
J (s, t)
M s
W(x-s, y-t)
f(x, y)
DC
34
Pattern Matching (by Correlation)
Correlation between f ( x, y ) and w( x, y ) is
C ( s, t ) f ( x, y ) w( x s, y t ),
x y
x y x y
DC
35
Nearest Mean Classifier
• Summarize the sample data using a mean vector
(use centroid)
ni
1
xi
ni
x
j 1
i, j
DC
36
Nearest Mean Classifier
DC
38
Nearest Mean Classifier
DC
39
Nearest Mean Classifier
• Problems
– When the structure of
the classes is complex
– Class is separated
(multimodal)
– Mean vector is not
representative
– Separate the class (2)
• Difficult in higher
dimensions
– Problems may remain
DC
40
Clustering
DC
41
Pattern and pattern Class
DC
42
Clustering
DC
43
Clustering
XXXXXX
XXXXX
XXX
XXX
Intra Cluster
XXXXXXX
XXXXXXXXX XXX
XX
DC
44
What is Cluster?
Everiit documents some of the definitions of a cluster as follows:
• A cluster is a set of entities which are alike and entities from different
clusters are not alike.
DC
45
Applications of Clustering
Document Processing
Man-Machine interfaces
Remote Sensing data analysis
K ! i 1 i
DC
47
Clustering
DC
48
Clustering
DC
49
Similarity Measures (Metric)
1/ 2
2
n
Euclidean Distance : d E X , Y xi yi
i 1
n
City - Block Distance : d C X , Y xi yi
i 1
n
Chessboard Distance : d M X , Y max xi yi
i 1
n
1
New Distance : d N X , Y xi XY yi XY xi yi
n 2 i 1,i i XY
n
2
Published in Pattern Recognition 1992, Vol. 25 DC
50
Decision Theoretic Methods
X x1 , x2 ,..., xn N - Dimensional pattern vector.
T
DC
51
Tree of Classification
CLASSIFICATION
INTRINSIC EXTRINSIC
(UNSUPERVISED) (SUPERVISED)
NON-HIERARCHICAL
HIERARCHICAL (PARTITIONAL)
DC
52
Clustering by Density Estimation
and Mode Seeking
Densely populated regions are usually viewed as
clusters.
By finding highest density point called mode.
The nearest point of each mode can be considered
as a cluster center.
Partition is obtained by assigning each pattern to
its closest cluster center.
DC
53
Density Estimation
Kn
fn ( X )
nVn
Kn is the number in a region Vn and n is the total number of
patterns.
Note:
In the Parzen window approach Vn is a function of n.
In nearest Neighbor approach Kn is a function of n.
For convergence Vn is inversely proportional to n in both
the methods and Kn is usually taken as proportional to n.
DC
54
Parzen Density Estimator
1 n
x xi
f n ( x)
nhn
K where x, xi , i 1,..., n be a sequence
i 1 hn
of independent and identically distributed random variables.
{hn } is a sequence of positive constant satisfying hn 0 and
nhn as n and K ( y ) is a function on such that
Sup K ( y ) ,
y
K ( y) dy
lim y K ( y ) 0
y K(y)dy 1
1/ q
ln
hn ; q is the dimension and ln is the length of the MST
n
1
if x j xij 1 j 1,2,..., q
K(X ) 2 q
0 otherwise
DC
57
Hierarchical Clustering
Begin with Proximity matrix
Graphical representation is the main feature
HIERARCHICAL
AGGLOMERATIVE DIVISIVE
DC
58
Agglomerative Method
Algorithm:
Compute similarity matrix and initialize each
pattern in a cluster.
DC
59
Agglomerative Method
DC
60
Dendrogram for Single-link Algorithm
F G
F2
B
A C D E
F1
(C,D)
(D,F)
Similarity
(A,C)
A B C D E F G DC
61
Dendrogram for Complete-link Algorithm
F G
F2
B
A C D E
F1
(A,G)
(D,G)
Similarity
(A,B)
A B C D E F G DC
62
Divisive Method
Reverse of Agglomerative method.
DC
64
Mathematical Construction
Suppose S X 1 ,..., X n q be given set and has some partition
into K clusters C1 ,..., C K such that cluster C K has nK patterns and
K
each pattern is in exactly one cluster. So ni n
i 1
nK
1
Mean vector is X ( K )
nK
i i is the ith pattern
X (K )
i 1
, X (K )
X
nK T
C K is e X
2
K i
(K )
X (K )
i
(K )
X i( K )
i 1
K
So the squared error for the clustering is Ei2 ei2
i 1 DC
65
Partitional Method (K-Means)
Basic Steps:
1. Find the seed point of each K cluster.
2. Find a similarity measure.
3. Cluster the data on the basis of seed points and
similarity measure.
4. Re-compute the cluster centers.
DC
66
Cluster Center or Seed Point
Astrahan Method:
1. Compute density for each data unit.
2. Order the data units by density.
3. Choose the highest density point as the first
seed point.
4. Subsequent seed points are chosen in order of
decreasing density subject to the condition that
each new seed point is at least a minimum
distance from all previously chosen points.
DC
67
Cluster Center or Seed Point
Ball and Hall Method:
1. Overall mean vector of data set is considered
as the first seed point.
DC
68
Cluster Center or Seed Point
Our Method: (Problem of choosing 1 out of n0 data
units) (Published in IEEE Trans. SMC, Vol. 24, 1994)
1. Compute density for each datum.
2. Rearrange the mi in decreasing order. Let L be the ordered
list. Let i1.
3. Choose the datum X that tops the list L as the ith seed
point. If i = K go to step 6.
4. Count the number of data in the current set S. If the number
is less than n0 –1 then go to step 6. Else, from the current set
S find the n0 –1 nearest neighbors of the data X. Delete X and
theses n0 –1 neighbors from L and S to get the list of L and S
for next iteration.
5. Make i i +1 and go to step 3.
6. Stop DC
69
K-Means Algorithm
1. Select an initial partition of n objects into K
clusters.
2. Assign each pattern to its closest center.
3. Re-compute new cluster centers.
4. If convergence criterion is satisfies, then stop;
else, go to step 2.
xi C j iff d 2 ( xi , c j ) d 2 ( xi , c p ), p j , p k , j k .
DC
71
Crisp Data & Hard Boundaries
x
xiC j
i, p
c j, p ,1 p d
Cj
Objective function to be minimized
k
J(X,C) d 2 ( xi , c j )
j 1 xiC j
DC
72
The KMeans Algorithm( X , k , )
X : Given set of patterns
k : Number of predetermined clusters
: Termination parameter
K-Means Algorithm
Initialize prototypes C {c1 , c2 , , ck }
Jm 0
Repeat
c previous c; J mprevious J m
for each x j X
for each c j C
2
compute d x j c j
2
i, j
K=6 K=5 DC
74
Seed-based Clustering Technique
DC
76
Split-and-Merge Technique
MacQueen’s K-Means Method with Coarsening and
Refining Parameters:
DC
79
How ISODATA works:
A second iteration is performed with the new cluster
centres
DC
80
How ISODATA works:
Along the way…
DC
81
Parameters
DC
82
When choosing parameters:
Experiment!
DC
83
Output
Disadvantages of ISODATA
5 1
Center Pixel
6 8
7
DC
86
Split-and-Merge Technique
Strip-based Clustering Technique:
Find the strips of finite width at different directions
around the centroid of the data.
Find smallest density strip and if it’s percentage
satisfied the splitting restriction then split.
Sparsely populated regions need not always be
present at centroid of the data. So apply modified K-
means (K = 2) method with restriction imposed.
Merging should be occurring in the border region of
two clusters. Find the merging circle set and
calculate the density for each cluster within circle. If
it is satisfied the merging restriction then merge. DC
87
Split-and-Merge Technique
DC
88
Split-and-Merge Technique
DC
93
Multi-seed Clustering Technique
DC
95
Fuzzy Clustering Technique
DC
96
Fuzzy Clustering Technique
• Crisp (ordinary) Cluster
• Fuzzy Cluster
DC
97
Fuzzy Clustering Algorithm
Fuzzy C-means clustering – Each data point belongs to each
cluster to a degree that is specified by a fuzzy membership
grade.
, .
DC
100
FCM Algorithm Continue…
Step 5: Update the membership values for each data point
DC
101
Remote
Sensing
DC
1
Orbit
An orbit definition is a path one object in space uses to move
around another object in space. The path is predictable
because it repeats regularly. Objects in space include
satellites, planets, moons, space stations, and more.
DC
2
Orbit
Two types orbit: 1. Synchronous orbit and 2. Geostationary orbit
Synchronous orbit is a type of orbit in which the period of
time it takes a satellite to orbit a body is equal to the body's
rotational speed and is orbiting in the same direction as its
rotation.
Geostationary orbit is a common
kind of circular synchronous orbit.
Geostationary orbit involves a satellite
orbiting above the Earth’s
equator while its orbital period is equal
to Earth's rotational period or day.
Satellites following a geostationary
orbit will usually orbit above the
Earth's equator at ~36,000 kilometers
and follow a circular orbit, maintaining
this distance. DC
3
Sun-Synchronous Satellite
Sun-synchronous
satellite is an orbit a
satellite takes around
the Earth where the
movement and orbit
of the satellite always
appear the same
from the viewpoint of
the Sun. The satellite
essentially follows an
orbital path around
the boundary
between the
nighttime and
daytime sides of
Earth. DC
4
Geo-stationary Satellite
Satellites in geostationary orbit
(GEO) circle Earth above the
equator from west to east following
Earth’s rotation – taking 23 hours 56
minutes and 4 seconds – by
travelling at exactly the same rate
as Earth. This makes satellites in
GEO appear to be ‘stationary’ over
a fixed position. In order to perfectly
match Earth’s rotation, the speed of
GEO satellites should be about 3
km per second at an altitude of 35
786 km. This is much farther from
Earth’s surface compared to many
satellites.
DC
5
Remote Sensing
Remote sensing is the process of acquiring data/information
about objects/substances not in direct contact with the sensor,
by gathering its inputs using electromagnetic radiation or
acoustical waves that emanate from the targets of interest. An
aerial photograph is a common example of a remotely sensed
product.
Types of remote sensing: Active
Passive
DC
6
Applications of Remote Sensing
• Meteorology - Study of atmospheric temperature, pressure, water
vapor, and wind velocity.
• Oceanography - Measuring sea surface temperature, mapping
ocean currents, and wave energy spectra and depth sounding of
coastal and ocean depths.
• Glaciology- Measuring ice cap volumes, ice stream velocity, and sea
ice
distribution.
• Geology- Identification of rock type, mapping faults and structure.
• Geodesy- Measuring the figure of the Earth and its gravity field.
• Topography and cartography - Improving digital elevation models.
• Agriculture - Monitoring the biomass of land vegetation.
• Forest- monitoring the health of crops, mapping soil moisture.
• Botany- forecasting crop yields.
• Hydrology- Assessing water resources from snow, rainfall and
underground aquifers.
DC
7
Applications of Remote Sensing
• Disaster warning and assessment - Monitoring of floods and
landslides, monitoring volcanic activity, assessing damage zones from
natural disasters.
• Planning applications - Mapping ecological zones, monitoring
deforestation, monitoring urban land use.
• Oil and mineral exploration- Locating natural oil seeps and slicks,
mapping geological structures, monitoring oil field subsidence.
• Military- developing precise maps for planning, monitoring military
infrastructure, monitoring ship and troop movements.
• Urban- determining the status of a growing crop.
• Climate- the effects of climate change on glaciers and Arctic and
Antarctic
Regions.
• Sea- Monitoring the extent of flooding.
• Rock- Recognizing rock types.
• Space program- is the backbone of the space program.
• Seismology: as a premonition.
DC
8
Principles and process of
remote sensing
1. Energy Source or Illumination
(A)
2. Radiation and the
Atmosphere (B)
3. Interaction with the Target (C)
4. Recording of Energy
(Electromagnetic
Radiation) by the Sensor (D)
6. Transmission, Reception,
and Processing (E)
7. Interpretation and Analysis
(F)
8. Application (G)
DC
9
DC
10
Spectral Ranges Commonly
Used in Remote Sensing
• Gamma Rays <0.30 nm: This range is completely absorbed by the upper
atmosphere and not available for remote sensing.
• X-Rays 0.03-30.0 nm: This range is completely absorbed by the
atmosphere and not employed in remote sensing.
• Ultraviolet: 0.03-0.40 μm: Hard UV 0.03-0.3 μm: This range is completely
absorbed by the atmosphere and not employed in remote sensing.
Photographic UV 0.30-0.40 μm: This range is not absorbed by the
atmosphere and detectable with film and photo detectors but with severe
atmospheric scattering.
• Visible Light: This narrow band of electromagnetic radiation extends from
about 400 nm (violet) to about 700 nm (red). It’s Available for remote
sensing the Earth, can be imaged with photographic film. Violet: 400-430
nm, Indigo: 430-450 nm, Blue: 450-500 nm: water increasingly absorbs
EM radiation at longer wavelengths, band 1 provides the best data for
mapping depth detail of water covered areas. It is also used for soil,
vegetation discrimination, forest mapping.
DC
11
Spectral Ranges Commonly
Used in Remote Sensing
• Green: 500-570 nm: The spectrum corresponds to the chlorophyll
absorption of healthy vegetation and is useful for mapping detail such as
depth or sediment in water bodies. Cultural features such as roads and
buildings also show up well in this band.
• Yellow: 570-590 nm, Orange: 590 - 610 nm
• Red: 610-700 nm: Chlorophyll absorbs these wavelengths in healthy
vegetation. Hence, this band is useful for distinguishing plant species, as well
as soil and geologic boundaries.
• Infrared: 0.7-300 μm: This region is sensitive to plant water content, which
is a useful measure in studies of vegetation health. This band is also used for
distinguishing clouds, snow, and ice. NIR: 0.7-1.5 μm, SWIR: 1.5-3 μm,
MWIR: 3-8 μm, LWIR: 8-15 μm, FIR:>than 15 μm.
• Microwaves (Radar) 1 mm-1 m: Microwaves can penetrate clouds, fog,
and rain. Radar is the active form of microwave remote sensing. Radar
images are acquired at various wavelength bands.
• Radio and TV Waves: 10 cm-10 km: The longest-wavelength portion of the
electromagnetic spectrum. DC
12
Spectral Ranges Commonly Used in
Remote Sensing
• Major ranges utilized for earth resources sensing are between
0.4 and 12 m (visible/infrared range).
DC
13
Characteristics
• Acquired from aircraft or spacecraft platforms
• Available in digital format
• Scanning Micro-densitometer is used for digital conversion
• Major characteristic of the image is the wavelength band it
represents.
• Passive sensor: some images depends on the solar
radiation in ultraviolet, visible and near-to-middle infrared
range wavelengths.
• Others are measurements of the spatial distribution of
energy emitted by earth surface, so called Thermal infrared
wavelength range.
• Active sensor: Microwave band of wavelengths.
DC
14
Characteristics
Resolution in satellite images:
• Spatial Resolution: Described by pixel size. It is defined in
terms of the Instantaneous Field Of View (IFOV) of the sensing
element.
• Spectral Resolution: Number and width of spectral bands of
electromagnetic energy detectable by the sensor.
• Temporal resolution: Measure of the repeat cycle or frequency
with which a senor revisits the same part of the earth’s surface.
• Radiometric Resolution: It is ability of the sensor to sense
small changes in reflected energy from the target. More the
number of quantization levels by which the reflected energy is
represented, higher the radiometric resolution.
DC
15
Characteristics
• Spatial and spectral resolutions determine the data volume
provided by sensor
• Example: Landsat Thematic Mapper has 7 wavelength bands
with 8-bit radiometric resolution. 6 of which have 30m spatial
resolution and one of which has 120m spatial resolution
(thermal band, long wavelength, long aperture is required to
collect sufficient signal energy to maintain the radiometric
resolution).
• 185 Km 185 Km image represents 2.37 million pixels in
thermal band and 38 million pixels for other 6 bands.
• At 8 bits per pixel a complete 7 band image is composed of
1.848 GB.
DC
16
Characteristics of Landsat TM Satellite
•LANDSAT-1 launched 1972 and lasted
until 1978.
•LANDSAT-2 launched 1975
•Three more satellites were launched in
1978, 1982, and 1984 (LANDSAT-3, 4,
and 5 respectively).
•LANDSAT-6 was launched on October
1993 but the satellite failed to obtain
orbit.
•LANDSAT-7 launched in 1999
•Only 7 and 5 are still working
DC
17
Characteristics of Landsat TM Satellite
DC
18
LANDSAT TM: applications
Band Nominal Spectral applications
location
1 Blue Water body penetration, soil-water
discrimination, forest type mapping, cultural
feature ID
2 Green Green reflectance peak of veg, for veg ID and
assessment of vigor, cultural feature ID
3 Red Chlorophyll absorption region, plant species
differentiation, cultural feature ID
4 Near infra red Veg types, vigor and biomass content,
delineating water bodies, soil moisture
assessment
5 mid infra red (1.55- Veg moisture, soil moisture, diff of soil from
1.75 mm) clouds
6 Thermal infra red Veg stress analysis, soil moisture, thermal
mapping
7 mid infra red(2.08- Discriminating mineral and rock types, veg
2.35 mm) moisture DC
19
LANDSAT TM
•An example: August 14, 1999 (left) and October 17, 1999
(right) images of the Salt Lake City area
• differences in color due to growing season
DC
20
SPOT Satellite
•Launched by France
• Stands for Satellite Pour
l'Observation de la Terre
•Operated by the French
Space Agency, Centre
National d'Etudes Spatiales
(CNES).
DC
21
Characteristics of SPOT Satellite
Altitude 832 Km
SPOT
DC
23
IKONOS data
High resolution satellite developed
by Space Imaging, launched 1999
Has sun-synchronous orbit and
crosses equator at 10:30 AM
Ground track repeats every 11
days
Extremely flexible: it can quickly
stabilize itself after pointing at a
new target, allowing it to follow
wavy features.
The entire spacecraft moves, not
just the sensor
11 km by 11 km image size, but
user specified strips and mosaics
can be ordered DC
24
IKONOS data
IKONOS collects panchromatic band (.45 to
.90 mm) at 1 m resolution
Collects four multispectral bands at 4 m
resolution
Bands include blue (.45 to .52 mm) , green
(.51 to .60 mm) , red (.63 to .70 mm), near IR
(.76 to .85 mm)
Radiometric resolution is 11 bits, or 2048
values
DC
25
IKONOS data
Here is 1m IKONOS view of suburbs, near winter Olympics
DC
26
IKONOS data
1m IKONOS view of Dubai
DC
27
IKONOS data
1m IKONOS pan image of Rome
DC
28
IRS-1C/1D Satellite
The first satellite, IRS 1C was launched on 28 December
1995 on a Russian Molniya – M (Blok - 2BL) rocket. It
operated until it was decommissioned on 21 September
2007 after serving for 11 years and 8 months.
IRS-1C
Multi-spectral Panchromatic
Image Image
DC
31
CARTOSAT Series Satellite
DC
32
CARTOSAT – 1 Satellite
DC
33
CARTOSAT – 1 Satellite
Launch date 5 May 2005
Launch vehicle PSLV- C6
Orbit 618 km Polar Sun
Synchronous
Payloads PAN FORE, PAN - AFT
Spatial Resolution PAN-F: 2.5 m x 2.78 m
PAN-A: 2.22 m x 2.23 m
Orbit Period 97 min
Number of Orbits Per day 14
Local time of equator 10:30 am
crossing
Repetivity 126 days
Revisit 5 days DC
34
CARTOSAT - 1 Satellite
DC
35
CARTOSAT - 3 Satellite
DC
36
Microwave Remote Sensing Image Data
Altitude 800 Km
Wavelength 0.233 m
Incidence angle 20o
Swath width 100 Km
Range resolution 25 m
Azimuth resolution 25 m
Spectral resolution 5 bit
DC
38
Seasat SAR Image
DC
39
Band Combinations
3,2,1
4,3,2
5,4,3
DC
40
Feature space image
• A graphical
representation of
the pixels by
plotting 2 bands
Band 4
vs. each other
• For a 6-band
Landsat image,
there are 15 Band 3
feature space
images DC
41
Remote Sensing
What is remote sensing image processing?
Pre-processing
Enhancement
Classification
Interpretation DC
42
Remote Sensing
What are some commonly used techniques for remote
sensing image enhancement?
Histogram equalization
Contrast stretching
Spatial filtering
Image fusion
Supervised
Unsupervised
DC
43
Remote Sensing
What are the applications of remote sensing image
processing?
DC
49
Spectral Ranges Commonly Used in
Remote Sensing
• Major ranges utilized for earth resources sensing are
between 0.4 and 12 m (visible/infrared range)
DC
50
Spectral Reflectance of Earth
60 Surface Material
3
%
50
Reflectance
40
30
20
2
10
1
0
0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.4 2.6
Wavelength m
Green
Blue
Red
DC
52
Classification of Remote Sensing Data
Classification
Supervised Unsupervised
(Clustering)
DC
54
DC
55
Image Classification
• Why classify?
• Make sense of a landscape
– Place landscape into categories
(classes)
• Forest, Agriculture, Water, etc
• Classification scheme = structure
of classes
– Depends on needs of users
DC
56
What is a Classified Image
DC
57
Classification of Remote Sensing Data
DC
58
Training Areas
DC
59
Classification of Remote Sensing Data
Parallelepiped Classification Algorithm:
Number of spectral bands is n. mck is mean value
of the training data obtained for class c in band k
out of M possible classes. Sd ck is the standard
deviation of the training data class c of band k .
Let Gijk be the gray value of the (i, j ) th pixel of
band k . Now Gijk in class c iff
mck Sd ck Gijk mck Sd ck
where c 1,2,..., M and k 1,2,..., n DC
60
Classification of Remote Sensing Data
Minimum distance to Mean Classification Algorithm
Number of spectral bands is n. mck is mean value
of the training data obtained for class c in band k
out of M possible classes. Suppose X ( x11 ,..., x1n )
is the unknown pixel to be classified. The Euclidean
distance between X and M c is
d ( X, M c ) x11 mc1 x12 mc 2 ... x1n mcn
2 2
2 1/ 2
e = “I have a cold”
D = “runny nose,” “watery eyes,” “coughing”
P(D|e)P(e)
P(e|D)=
P(D)
So Bayes’ theorem is “diagnostic”.
DC
62
Bayes’ Theorem
D = Data, e = some event
P(D|e)P(e)
P(e|D) =
P(D)
P(e) is called the prior probability of e. Its what we know
(or think we know) about e with no other evidence.
P(D|e) is the conditional probability of D given that e
happened, or just the likelihood of D. This can often be
measured or computed precisely – it follows from your
model assumptions.
P(e|D) is the posterior probability of e given D. It’s the
answer we want, or the way we choose a best answer.
You can see that the posterior is heavily colored by the
prior, so Bayes’ has a GIGO liability. e.g. its not used to
test hypotheses DC
63
Bayesian Classifier
DC
64
Bayesian Classifier
Simple Example
• Classifying Cars into two classes of a scene
– Red
– Black
• Inspection using IR camera
• Use only one feature (intensity)
• Training :
– Use many samples from the two classes
– Measure the feature values
DC
65
Bayesian Classifier
Example:
DC
66
Bayesian Classifier
Example:
DC
67
Bayesian Classifier
Example
: Class of red car
2 : Class of black
x : Observed feature from scene (scalar)
p(x | ) :Probability of the feature value of a car x
given that the car is red
p( x | ) P( )
i 1
i i DC
69
Classification of Remote Sensing Data
Maximum Likelihood Classification Algorithm
Baye’s Classification:
DC
70
Classification of Remote Sensing Data
Maximum Likelihood Classification Algorithm
P( X / wi ) is the posterior probability i.e. for X in
multi - spectral space a set of probabilities can be
computed that give relative likelihood that pixel belongs
to each available class. Baye's theorem
P( wi / x) P ( x / wi ) P( wi ) / P( X)
P( wi ) is a priori probability and P( X) is probability density
x wi if P( x / wi ) P( wi ) P ( x / w j ) P( w j ) j i
g i ( x) ln[ P( x / wi ) P( wi )] ln P( x / wi ) ln P( wi )
x wi if g i ( x) g j ( x) j i DC
71
Classification of Remote Sensing Data
Multivariate Normal Class Models:
1
1 / 2
P( x / wi ) (2 ) N / 2
i exp ( x mi ) i ( x mi )
1
2
where mi and i are the mean vector and covariance
matrix of the data in class wi .
1 1
1
g i ( x) ln P( wi ) ln i ( x mi ) i ( x mi )
2 2
Equal probable case
1
g i ( x) ln i ( x mi ) i ( x mi )
DC
72
Classification of Remote Sensing Data (Multi-seed
Technique) (Published in Proc. CODIS 2004 and IEEE Trans. GRS, Sep
2008)
Processing Steps:
Training Sample Selection
Test for homogeneity
If heterogeneous then form heterogeneous sample as
the union of homogeneous regions
Statistical parameters extraction from each
homogeneous region
Minimum Distance Logical Based Classifier
Majority must be granted logic
Minimum of minimum distance granted logic
DC
73
Classification of Remote Sensing Data
DC
84
Precision and Recall
DC
85
Precision and Recall
DC
86
Accuracy Assessment
DC
87
Accuracy Assessment
DC
88
Accuracy Assessment
DC
89
Accuracy Assessment
DC
90
Accuracy Assessment
DC
91