Last Time: Affine Transforms (Linear Spatial Transforms) : Intensity Transformations & Spatial Filtering

Digital Image Processing, 3rd ed.
Gonzalez & Woods

www.ImageProcessingPlace.com
Chapter 3
Intensity Transformations & Spatial Filtering
Last time:
Sxy consists of a subset of m rows
ns, considering the
Affine transforms (linear spatial transforms)
nsformation
1 �
g(x, y) = f (r, c)
mn (r,c)∈Sxy
 
t11 t12 0
[ x y 1 ] = [ v w 1 ] 
 t21 t22 0 
t31 t32 1
© 1992–2008 R. C. Gonzalez & R. E. Woods
Tuesday, February 1, 2011

Gonzalez & Woods
Chapter 3
tsconsists
of a subset
of a of m rows
subset of m rows
y
ering the
IMTRANSFORM
considering
Apply 2-D spatial transformation to image.
the
B = IMTRANSFORM(A,TFORM) transforms the image A according to the 2-D
n
ormationspatial transformation defined by TFORM, which is a tform structure
as returned by MAKETFORM or CP2TFORM. If ndims(A) > 2, such as for
�then the same 2-D transformation is automatically
an RGB 1image,
g(x, y)applied
= 1 f �c)
(r,
to all 2-D planes along the higher dimensions.
g(x,
mn y) = f (r, c)
mn
(r,c)∈S xy
(r,c)∈Sxy
 
t11 t 0 t12  T =
 11 t12 0
y 1x] =y [ 1v w 1v ] wt211 t22 t 0 
t22 0 
2 0 0
[ ]=[ ]  21  1 1 0
t31 t32 1 50 -1 1
t31 t32 1
>> tform = maketform('affine',T);
x = vt11 + wt21 + t31 tform =
y = vt12 + wt22 + t32 ndims_in: 2

ndims_out: 2
forward_fcn: @fwd_affine
inverse_fcn: @inv_affine
tdata: [1x1 struct]
>> IMG2= imtransform(IMG, tform)


1 �
Digital Image
g(x, y) = Processing,f 3rd ed.
(r, c)
mn & Woodsxy
Gonzalez(r,c)∈S

www.ImageProcessingPlace.com T = 
t11 t12 0
Chapter 3  2 0 0
[ x y 1 ] = [ v w 1 ]  t21 t22 0 1 1 0
t31 t32 1 50 -1 1
>> tform = maketform('affine',T);
tform =
x = vt11 + wt21 + t31
ndims_in: 2
y = vt12 + wt22 + t32 ndims_out: 2
forward_fcn: @fwd_affine
inverse_fcn: @inv_affine
tdata: [1x1 struct]
>> IMG2= imtransform(IMG, tform)
NB! IMTRANSFORM assumes

spatial-coordinate conventions for
the transformation TFORM.
Specifically, the first dimension of
the transformation is the horizontal
or x-coordinate, and the second
dimension is the vertical or y-
coordinate. Note that this is the
reverse of MATLAB's array
subscripting convention.

Gonzalez & Woods
Chapter 3
The process of applying a geometric transformation to an image (often

called study) to match it to another image (called template/reference) is
called image registration.

Gonzalez & Woods
Chapter 3

Affine/rigid transformations can be used to perform image registration.

Gonzalez & Woods
Chapter 3

There exist also other types of registration (nonlinear)

-bilinear, bicubic, polynomials
-spline representations
-fully nonlinear (by PDEs)

Gonzalez & Woods
Chapter 3

There exist also other types of registration (nonlinear)

-bilinear, bicubic, polynomials
-spline representations
-fully nonlinear (by PDEs)
For the time being, we will focus only on linear and bilinear
transformations, other registration methods will be considered at the
end of the course (MAT262 only).

Gonzalez & Woods
Chapter 3
Image registration
Linear (affine):
we have 6 free parameter
(v, w) = T (x, y) (x, y) = T −1 (v, w
we need 6 conditions to determine the parameters
(v, w) = T (x, y) (x, y) = T −1 (v, w)
x = c1 v + c2 w + c3 vw + c4
Bilinear: y = c5 v + c6 w + c7 vw + c8
x = c1 v + c2 w + c3 vw + c4
(av + b)(cw + d)
y = c5 v + c6 w + c7 vw + c8
we have 8 free parameter

we need 8 conditions to determine the parameters

Gonzalez & Woods
Chapter 3
A typical method to set up the parameters is through landmarks

also called (tie, control, ... points).
This can be done either manually/automatically.
Some imaging systems include some sensor/markers to help

recognizing the points.
It suffices:
Linear: 6 free parameters 3 couples of points
Bilinear: 8 free parameters 4 couples of points
Each couple of coordinate (x,y) gives 2 conditions.

Gonzalez & Woods
x = v + c2 w + c3 vw + c4
c1www.ImageProcessingPlace.com
y = c5 v + c6 w + c7 vw + c8
Chapter 3
Intensity Transformations
(av + b)(cw + d) & Spatial Filtering
(x1 , y1 ), (x2 , y2 ), (x3 , y3 ), (x4 , y4 )
(v1 , w1 ), (v2 , w2 ), (v3 , w3 ), (v4 , w4 )
Solve the two linear systems
         
v1 w1 v1 w1 1 c1 x1 v1 w1 v1 w1 1 c5 y1
 v2 w2 v2 w2 1  c2   x2   v2 w2 v2 w2 1  c6   y2 
         
  = ,   = 
 v3 w3 v3 w3 1  c3   x3   v3 w3 v3 w3 1  c7   y3 
v4 w4 v4 w4 1 c4 x4 v4 w4 v4 w4 1 c8 y4
(note that they have the same matrix of coefficients)
to find the parameters of the transformation.

Gonzalez & Woods
Chapter 3

Gonzalez & Woods
Chapter 3
We could use more points (overdetermined system)

in which case we have a Least Squares linear problem, that can be
solved by standard methods (normal equations, QR, SVD, etc.).

Gonzalez & Woods
Chapter 3
We could use more points (overdetermined system)

in which case we have a Least Squares linear problem, that can be
solved by standard methods (normal equations, QR, SVD, etc.).
These transformations might also be combined to work on

subregions of the images, which then have to be patched together.

(x1 , y1 ), (x2 , y2 ), (x3 , y3 ), (x4 , y4 )
Digital
(v1 ,Image
w1 ),Processing, 3rd ed.
(v2 , w2 ), (v 3 , w3 ), (v4 , w4 )
Gonzalez & Woods
Solve the two linear systems www.ImageProcessingPlace.com
Chapter 3
      

v1 w1 v1 w1 1 c1 x1 v1 w1 v1 w1 1
 v w v w 1  c   x   v w v w 1 
 2 2 2 2  2   2   2 2 2 2 
Chapter
 3: Intensity   =  and
transformations  , spatial
 filtering 
 v3 w3 v3 w3 1   c3   x3   v3 w3 v3 w3 1 
v4 w4 v4 w4 1 c4 x4 v4 w4 v4 w4 1
Generic form:
g(x, y) = T [f (x, y)]
new intensity value at (x,y) is a function of the

old intensity value at
(x,y), or Sxy

Gonzalez & Woods
Chapter 3
Example of spatial filtering:

blurring by 3x3 neighborhood
Create a 3x3 mask, and run it over the whole image to generate the new
image (can also use conv2)
Here I have
run the mask
only on the
left half of
© 1992–2008 the picture.
R. C. Gonzalez & R. E. Woods

Gonzalez & Woods
Chapter 3
Example of spatial filtering:

blurring by 3x3 neighborhood
Create a 3x3 mask, and run it over the whole image to generate the new
image (can also use conv2)
Here I have
Example of a blurred
run the mask image by a 3x3 mask
only on the
left half of
© 1992–2008 the picture.

Gonzalez & Woods
Chapter 3
Simplest example:
1-pixel neighborhood (the pixel itself), in which case we operate
directly on the intensity of the pixel.
This is called point processing as opposed to neighborhood

processing.
Typical operations we will perform:

Image enhancement
contrast stretching
Is the task of manipulating an image so
thresholding that the result is more suitable for specific
applications.
Mostly visual (subjective), no general
quantitative assessment.
Gonzalez & Woods
Chapter 3
Contrast stretching (sigmoid):
low intensity areas are mapped

to even lower
high intensity areas are mapped
to even higher
medium intensity areas are
“stretched”.

Gonzalez & Woods
Chapter 3

to even lower
to even higher
“stretched”.
© 1992–2008 R. C. Gonzalez & R. E. Woods Original contrast stretched

Gonzalez & Woods
Chapter 3

to even lower
to even higher
“stretched”.
© 1992–2008 R. C. Gonzalez & R. E. Woods Original contrast stretched

Gonzalez & Woods
Chapter 3
Thresholding: we choose a threshold parameter k: intensities below

this parameters are set to 0, those above are set to 1.
This produces a binary image.

Gonzalez & Woods
Chapter 3
Thresholding: we choose a threshold parameter k: intensities below

this parameters are set to 0, those above are set to 1.
This produces a binary image.
Thresholded image (setting to 1 the

values above 150, 0 the others)

(v1 , w1 ), (v2 , w2 ), (v3 , w3 ), (v4 , w4 )
Gonzalez & Woods
         
v1 w1 v1 w1 1 c1 x1 v1 w1 v1 w1 1
www.ImageProcessingPlace.com c5 y1
 v2 w2 v2 w2 1  c2     v v2 w2 1     
    x2   2 w
Chapter 32   c6   y2 
  = ,   = 
 v3 w3 v3 w3 1  c3   xTransformations
Intensity 3   v3 w3& v w
Spatial
3 3 1   c7 
Filtering  y3 
v4 w4 v4 w4 1 c4 x4 v4 w4 v4 w4 1 c8 y4
Generic form:
Other maps of the type s=T(r)
g(x, y) = T [f (x, y)]
Negative map: s = (L − 1) − r
s = c log(1 + r)
useful when one is interested in gray/white areas and the black areas
are dominating

(v1 , w1 ), (v2 , w2 ), (v3 , w3 ), (v4 , w4 )
Gonzalez & Woods
         
v1 w1 v1 w1 1 c1 x1 v1 w1 v1 w1 1
 v2 w2 v2 w2 1  c2     v v2 w2 1     
    x2   2 w
Chapter 32   c6   y2 
  = ,   = 
Spatial
3 3 1   c7 
v4 w4 v4 w4 1 c4 x4 v4 w4 v4 w4 1 c8 y4
Generic form:
g(x, y) = T [f (x, y)]
s = c log(1 + r)
are dominating

(v1 , w1 ), (v2 , w2 ), (v3 , w3 ), (v4 , w4 )
Gonzalez & Woods
         
v1 w1 v1 w1 1 c1 x1 v1 w1 v1 w1 1
 v2 w2 v2 w2 1  c2     v v2 w2 1     
    x2   2 w
Chapter 32   c6   y2 
  = ,   = 
Spatial
3 3 1   c7 
v4 w4 v4 w4 1 c4 x4 v4 w4 v4 w4 1 c8 y4
Generic form:
g(x, y) = T [f (x, y)]
s = c log(1 + r)
are dominating

(v1 , w1 ), (v2 , w2 ), (v3 , w3 ), (v4 , w4 )
Gonzalez & Woods
         
v1 w1 v1 w1 1 c1 x1 v1 w1 v1 w1 1
 v2 w2 v2 w2 1  c2     v v2 w2 1     
    x2   2 w
Chapter 32   c6   y2 
  = ,   = 
Spatial
3 3 1   c7 
v4 w4 v4 w4 1 c4 x4 v4 w4 v4 w4 1 c8 y4
Generic form:
g(x, y) = T [f (x, y)]
s = c log(1 + r)
are dominating

y = c5 v + c6 w + c7 vw + c8
(av + b)(cw + d) Gonzalez & Woods
(x1 , y1 ), (x2 , y2 ), (x3 , y3 ), (x4 ,www.ImageProcessingPlace.com
y4 )
(v1 , w1 ), (v2 , w2 ), (v3 , w3 ), (v4 , w4 ) Chapter 3
e two linear systems Intensity Transformations & Spatial Filtering
        
1 v1 w1 1 c1 x1 v1 w1 v1 w1 1 c5 y1
v2 w2 1  c2   x2   v2 w2 v2 w2 1  c6   y2 
2         
 = ,   = 
3 v3 w3 1  c3   x3   v3 w3 v3 w3 1  c7   y3 
4 v4 w4 1 c4 x4 v4 w4 v4 w4 1 c8 y4
form:
g(x, y) = T [f (x, y)]
s = (L − 1) − r
Log map: s = c log(1 + r)
maps lower intensities to a

wider range, while white
intensities are
mapped to a smaller range

y = c5 v + c6 w + c7 vw + c8
(av + b)(cw + d) Gonzalez & Woods
(x1 , y1 ), (x2 , y2 ), (x3 , y3 ), (x4 ,www.ImageProcessingPlace.com
y4 )
(v1 , w1 ), (v2 , w2 ), (v3 , w3 ), (v4 , w4 ) Chapter 3
e two linear systems Intensity Transformations & Spatial Filtering
        
1 v1 w1 1 c1 x1 v1 w1 v1 w1 1 c5 y1
v2 w2 1  c2   x2   v2 w2 v2 w2 1  c6   y2 
2         
 = ,   = 
3 v3 w3 1  c3   x3   v3 w3 v3 w3 1  c7   y3 
4 v4 w4 1 c4 x4 v4 w4 v4 w4 1 c8 y4
form:
g(x, y) = T [f (x, y)]
s = (L − 1) − r
Log map: s = c log(1 + r)
maps lower intensities to a

wider range, while white
intensities are
mapped to a smaller range

        
1 1 c1 x1 Digital
v1 w1 Image
v1 w1 Processing,
1 c5 3rd yed.
1
1      v  c   y 
2  c2   x2   2 w2 v2 w 2 1 &Woods
Gonzalez 6   2 
 = ,     =  
3 1  c3   x3   v3 3 3 1
w3 vwww.ImageProcessingPlace.com
w   c 7   y 3 
4 1 c4 x4 v4 1
w4 v4 w4 Chapter c38 y4
g(x, y) = T [f (x, y)]
s = (L − 1) − r
Power-law (gamma) map: gamma=0.2
s = c log(1 + r)
s = crγ
These subsume the log and

inverse log, but are more
powerful, as just varying the
gamma one gets a whole family
of transformations.
gamma=3

        
1 1 c1 x1 Digital
v1 w1 Image
v1 w1 Processing,
1 c5 3rd yed.
1
1      v  c   y 
2  c2   x2   2 w2 v2 w 2 1 &Woods
 = ,     =  
3 1  c3   x3   v3 3 3 1
w   c 7   y 3 
4 1 c4 x4 v4 1
g(x, y) = T [f (x, y)]
s = (L − 1) − r
s = c log(1 + r)
s = crγ

of transformations.
gamma=3

        
1 1 c1 x1 Digital
v1 w1 Image
v1 w1 Processing,
1 c5 3rd yed.
1
1      v  c   y 
2  c2   x2   2 w2 v2 w 2 1 &Woods
 = ,     =  
3 1  c3   x3   v3 3 3 1
w   c 7   y 3 
4 1 c4 x4 v4 1
g(x, y) = T [f (x, y)]
s = (L − 1) − r
s = c log(1 + r)
s = crγ

of transformations.
gamma=3

Gonzalez & Woods
Chapter 3

Gonzalez & Woods
Chapter 3
Piecewise linear transformations:
Contrast stretching by
piecewise linear transformations.
(r1, s1)
(r2, s2)
Control points

Gonzalez & Woods
Chapter 3

Gonzalez & Woods
Chapter 3
r1=0=s1
r2=L-1=s2
Identity

Gonzalez & Woods
Chapter 3

Gonzalez & Woods
Chapter 3
r1=r2
s1=0, s2=L-1
Thresholding

Gonzalez & Woods
Chapter 3
r1=r2
s1=0, s2=L-1
Thresholding
Other combinations give different effects on the image.

Gonzalez & Woods
Chapter 3
Other combinations give different effects on the image.

Gonzalez & Woods
Chapter 3
Intensity level slicing:

to highlight a specific range of intensities.
Most popularly implemented as:

showing white in one range and black in the rest
showing white in one range and graytone in the rest

Gonzalez & Woods
Chapter 3

Gonzalez & Woods
Chapter 3
Bit slicing:
2^7 2^6 2^5 2^4 2^3 2^2 2^1 2^0
0 1 0 1 0 0 1 0
82=64+16+2

Gonzalez & Woods
Chapter 3
The reconstructed image is obtained as

2^0*bit_plane1 + 2^1*bit_plane2 + …. + 2^7*bit_plane8
bit plane 1
The most
significant planes
are often those with
high bit number
bit plane 8
Gonzalez & Woods
Chapter 3

Gonzalez & Woods
Chapter 3

4 x4 v4 w4 v4 w4 1 c8 y4
Gonzalez & Woods
g(x, y) = T [f (x, y)] Chapter 3

s = (L − 1) − r
Histogram processing
s = c log(1 + r)
Histogram: is a discrete function that counts how many pixels have a
given intensity
s = cr γvalue.
h(rk ) = nk Number of pixels that have such

intensity value
k-th intensity value

4 x4 v4 w4 v4 w4 1 c8 y4
Gonzalez & Woods
g(x, y) = T [f (x, y)] Chapter 3

s = (L − 1) − r
Histogram processing
s = c log(1 + r)
Histogram: is a discrete function that counts how many pixels have a
given intensity
s = cr γvalue.
h(rk ) = nk Number of pixels that have such

intensity value
k-th intensity value

6 7 8
Digital Image Processing, 3rd (av
ed. + b)(cw + d)
+ b)(cw + d) (x
Gonzalez & Woods
1 , y1 ), (x2 , y2 ), (x3 , y3 ), (x4 , y4 )
y2 ), (x3 , y3 ), (x4 , y4 ) (v1 , w1 ), (v2 , w2 ), (v3 , w3 ), (v4 , w4 )

Chapter 3
w2 ), (v3 , w3 ), (v4 , w4 ) Solve the two linear systems
      
v1 w1 v1 w1 1 c1 x1 v1 w1 v1 w1 1
 v w v w 1  c   x   v w v w 1 
It is
 typical to 
 scale nk by MN, the       2 
 number
  of pixels.
 The  corresponding
 =  number represents
2 2 2 2 2 2 2the 2 2
,  
x1probability of avpixel
1 whaving
1 v1the v13 intensity.
w1given w3 cv53 w3 1   y1c3   x3   v3 w3 v3 w3 1 
x 2 
 v w v w v14 w4 cv4 w4 1 y c4
 2 2 2  
2 6
x4
2
v4 w4 v4 w4 1
,   = 
x3   v3 w3 1  form:
v3 w3Generic c7   y3 
x4 v4 w4 v4 w4 1 c8 y4 g(x, y) = T [f (x, y)]
s = (L − 1) − r
) = T [f (x, y)] s = c log(1 + r)
(L − 1) − r s = crγ
�
h(rk ) = nk nk = M N
c log(1 + r) k
s = crγ nk �
� p(rk ) = p(rk ) = 1
MN
nk = M N k
k Probability that a random pixels has intensity rk .
k © 1992–2008
�
) = 1
p(r k
NTuesday, February 1, 2011
Gonzalez & Woods
Chapter 3

Last Time: Affine Transforms (Linear Spatial Transforms) : Intensity Transformations & Spatial Filtering

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Last Time: Affine Transforms (Linear Spatial Transforms) : Intensity Transformations & Spatial Filtering

Uploaded by

Copyright:

Available Formats

Digital Image Processing, 3rd ed.

Gonzalez & Woods

© 1992–2008 R. C. Gonzalez & R. E. Woods

Tuesday, February 1, 2011

x = vt11 + wt21 + t31 tform =

y = vt12 + wt22 + t32 ndims_in: 2

>> IMG2= imtransform(IMG, tform)

Tuesday, February 1, 2011

>> tform = maketform('affine',T);

>> IMG2= imtransform(IMG, tform)

NB! IMTRANSFORM assumes

© 1992–2008 R. C. Gonzalez & R. E. Woods

Tuesday, February 1, 2011

The process of applying a geometric transformation to an image (often

© 1992–2008 R. C. Gonzalez & R. E. Woods

Tuesday, February 1, 2011

The process of applying a geometric transformation to an image (often

Affine/rigid transformations can be used to perform image registration.

© 1992–2008 R. C. Gonzalez & R. E. Woods

Tuesday, February 1, 2011

The process of applying a geometric transformation to an image (often

Affine/rigid transformations can be used to perform image registration.

There exist also other types of registration (nonlinear)

© 1992–2008 R. C. Gonzalez & R. E. Woods

Tuesday, February 1, 2011

The process of applying a geometric transformation to an image (often

Affine/rigid transformations can be used to perform image registration.

There exist also other types of registration (nonlinear)

Tuesday, February 1, 2011

we have 8 free parameter

Tuesday, February 1, 2011

A typical method to set up the parameters is through landmarks

Some imaging systems include some sensor/markers to help

Each couple of coordinate (x,y) gives 2 conditions.

© 1992–2008 R. C. Gonzalez & R. E. Woods

Tuesday, February 1, 2011

(note that they have the same matrix of coefficients)

to find the parameters of the transformation.

© 1992–2008 R. C. Gonzalez & R. E. Woods

Tuesday, February 1, 2011

© 1992–2008 R. C. Gonzalez & R. E. Woods

Tuesday, February 1, 2011

We could use more points (overdetermined system)

© 1992–2008 R. C. Gonzalez & R. E. Woods

Tuesday, February 1, 2011

We could use more points (overdetermined system)

These transformations might also be combined to work on

© 1992–2008 R. C. Gonzalez & R. E. Woods

Tuesday, February 1, 2011

Solve the two linear systems www.ImageProcessingPlace.com

new intensity value at (x,y) is a function of the

© 1992–2008 R. C. Gonzalez & R. E. Woods

Tuesday, February 1, 2011

Example of spatial filtering:

Tuesday, February 1, 2011

Example of spatial filtering:

Tuesday, February 1, 2011

This is called point processing as opposed to neighborhood

Typical operations we will perform:

Contrast stretching (sigmoid):

low intensity areas are mapped

© 1992–2008 R. C. Gonzalez & R. E. Woods

Tuesday, February 1, 2011