Summary of Lecture 1

• “Let there be light” → incident light → reflectivity → reflected light

→ projection → sensitivity → fC(x0, y 0)
• Sampling: fC (x0, y 0) → fC (i, j).
• Quantization: fC(i, j) → fˆC(i, j) ∈ {0, . . . , 255}.
• As a matrix: fˆC(i, j) → fˆgray(i, j) → A(i, j).

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 1
 Images as Matrices

• An image matrix (N × M ):
  



A(0, 0) A(0, 1) A(0, 2) . . . A(0, M − 1) 






  



 A(1, 0) A(1, 1) A(1, 2) . . . A(1, M − 1) 



A= 

 .. 

 

N rows
  

  

  


A(N − 1, 0) A(N − 1, 1) A(N − 1, 2) . . . A(N − 1, M − 1) 


• A(i, j) ∈ {0, 1, . . . , 255}.

• A(i, j):
– “Matrix case:” The matrix element (i, j) with value A(i, j).
– “Image case:” The pixel (i, j) with value A(i, j).
– Will use both terminologies.

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 2
 Simple Processing - Transpose

• The transpose image B (M × N ) of A (N × M ) can be obtained as

B(j, i) = A(i, j)
(i = 0, . . . , N − 1, j = 0, . . . , M − 1).

A B
>> for i = 1 : 512
for j = 1 : 512
B(j, i) = A(i, j); OR >> B = A0 ;
end
end

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 3
 Simple Processing - Flip Vertical

• The vertical flipped image B (N × M ) of A (N × M ) can be obtained

as B(i, M − 1 − j) = A(i, j) (i = 0, . . . , N − 1, j = 0, . . . , M − 1).

A B
>> clear B;
>> for i = 1 : 512
for j = 1 : 512
B(i, 512 + 1 − j) = A(i, j);
end
end

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 4
 Simple Processing - Cropping

• The cropped image B (N1 × N2) of A (N × M ), starting from (n1, n2), can
be obtained as B(k, l) = A(n1 + k, n2 + l) (k = 0, . . . , N1 − 1, l = 0, . . . , N2 − 1).

A B
>> clear B;
>> for k = 0 : 64 − 1
for l = 0 : 128 − 1
B(k + 1, l + 1) = A(255 + k + 1, 255 + l + 1); % n1=n2=255 N1=64,N2=128
end
end

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 5
 Cropping as a Matlab Function

function [B] =mycrop(A, n1, n2, N 1, N 2)

% mycrop.m
% [B] =mycrop(A, n1, n2, N 1, N 2)
% crops the image A from location n1, n2
% with size N 1, N 2

for k = 0 : N 1 − 1
for l = 0 : N 2 − 1
B(k + 1, l + 1) = A(n1 + k + 1, n2 + l + 1);
end
end
function [B] =mycrop(A, n1, n2, N 1, N 2)
% mycrop.m
% [B] =mycrop(A, n1, n2, N 1, N 2)
% crops the image A from location n1, n2
% with size N 1, N 2

B(1 : N 1, 1 : N 2) = A(n1 + 1 : n1 + N 1, n2 + 1 : n2 + N 2);

>> help mycrop
>> B =mycrop(A, 255, 255, 64, 128);

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 6
 Simple Image Statistics - Sample Mean and Sample
Variance

• The sample mean (mA) of an image A (N × M ):

NX
−1 MX
−1
A(i, j)
i=0 j=0
mA = (1)
NM
• The sample variance (σA2 ) of A:
NX
−1 MX
−1
(A(i, j) − mA)2
i=0 j=0
σA2 = (2)
NM
r
• The sample standard deviation, σA = σA2 .

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 7
 Simple Image Statistics - Histogram

Let S be a set and define #S to be the cardinality of this set,

i.e., #S is the number of elements of S .

• The histogram hA(l) (l = 0, . . . , 255) of the image A is defined as:

hA(l) = #{(i, j) | A(i, j) = l, i = 0, . . . , N − 1, j = 0, . . . , M − 1} (3)

• Note that:
255
X
hA(l) = Number of pixels in A (4)
l=0

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 8
 Calculating the Histogram

>> h=zeros(256,1); OR >> h=zeros(256,1);

>> for l = 0 : 255 >> for l = 0 : 255
for i = 1 : N h(l + 1)=sum(sum(A == l));
for j = 1 : M end
if (A(i, j) == l) >> bar(0:255,h);
h(l + 1) = h(l + 1) + 1;
end
end
end
end
>> bar(0:255,h);

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 9
 Example I

A hA (l)

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 10
 Example II

A hA (l)

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 11
 Example III

A hA (l)

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 12
 Point Processing

• We will now utilize a “function” g(l) (l = 0, . . . , 255) to generate a new

image B from a given image A via:
B(i, j) = g(A(i, j)), i = 0, . . . , N − 1, j = 0, . . . , M − 1 (5)

• The function g(l) operates on each image pixel or each image point
independently.
• In general the resulting image g(A(i, j)) may not be an image matrix,
i.e., it may be the case that g(A(i, j)) 6∈ {0, . . . , 255} for some (i, j).
• Thus we will have to make sure we obtain an image B that is an image
matrix.
• The histograms hA(l) and hB (l) will play important roles.

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 13
 Identity point function

• Let g(l) = l (l = 0, . . . , 255).

B(i, j) = g(A(i, j)), i = 0, . . . , N − 1, j = 0, . . . , M − 1
= A(i, j)

• In this case g(A(i, j)) ∈ {0, . . . , 255} so no further processing is necessary

to ensure B is an image matrix.
• Note also that B = A and hence hB (l) = hA(l).

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 14
 Digital Negative

• Let g(l) = 255 − l (l = 0, . . . , 255).

B(i, j) = g(A(i, j)), i = 0, . . . , N − 1, j = 0, . . . , M − 1
= 255 − A(i, j)

B hB (l)

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 15
 Digital Negative Contd.

hA(l) hB (l)

• In this case it is easy to see that hB (255 − l) = hA(l) or hB (g(l)) = hA(l).

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 16
 Calculating the Histogram of Point Processed Images

• For a given point function g(l), “the inverse point function” g −1(k) may
not exist.
• So in general hB (g(l)) 6= hA(l).
• Let Sg−1(k) = {l | g(l) = k, l = 0, . . . , 255} (k = 0, . . . , 255).
• Then:
X
hB (k) = hA(l), k = 0, . . . , 255 (6)
l∈Sg −1 (k)

• You must learn how to calculate and sketch hB (l) given hA(l) and the
point function g(l).

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 17
 Square-root point function

√
• Let g(l) = l (l = 0, . . . , 255).
B(i, j) = g(A(i, j)), i = 0, . . . , N − 1, j = 0, . . . , M − 1
r
= A(i, j)

• In this case g(A(i, j)) 6∈ {0, . . . , 255} and we must ensure B is an image
matrix by further processing.
• We can try “rounding” g(A(i, j)) to integers by defining a new point
function g2(l) = round(g(l)):
B(i, j) = g2(A(i, j)), i = 0, . . . , N − 1, j = 0, . . . , M − 1
r
= round( A(i, j)) (>> B = round(A.ˆ(.5)); )
but that will not exactly yield what we want as we shall see.
• We will return to this example after we learn Contrast Stretching.

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 18
 Contrast Stretching


α1l,





0 ≤ l < a1

g(l) = 

α2(l − a1) + α1a1, a 1 ≤ l < a2 (7)



 α3 (l − a2 ) + (α2 (a2 − a1 ) + α1 a1 ), a2 ≤ l ≤ 255


αi > 1 ⇒ Range Stretching

αi < 1 ⇒ Range Compression

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 19
 Example - Square-root

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 20
 Piecewise Linear, “Continuous” Point Functions

• Let α0 = a0 = 0.







α1l, 0 ≤ l < a1






 α2(l − a1) + α1a1, a1 ≤ l < a 2
g(l) =  ..


(8)



 Pi−1





α i (l − a i−1 ) + ( j=1 αj (aj − aj−1 )), ai−1 ≤ l < ai





..

• |αi| < 1 ⇒ Range Compression.

• |αi| > 1 ⇒ Range Stretching.
• Assuming 0 ≤ g(l) ≤ 255 (l = 0, . . . , 255), affect point function by incor-
porating the round(. . . ) function when necessary, i.e., g2(l) = round(g(l)) :
B(i, j) = g2(A(i, j));

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 21
 A “Discontinuous” Point Function

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 22
 Normalizing an Image

• Let mxA = max(i,j) A(i, j) and mnA = min(i,j) A(i, j).

These can be generated in matlab via:
>> mx=max(max(A));
>> mn=min(min(A));

• The special normalizing point function for A is defined as:

l − mnA
gsA (l) = round( × 255) (9)
mxA − mnA

• Convention:
A ⇒ processing ⇒ B ⇒ “normalize B”⇒ C(i, j) = gsB (B(i, j)).

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 23
 More Range Stretching/Compression

• In some cases we wish to view a matrix (image matrix or otherwise)

having a wide range of values.
• Let

A(0, 0) + 106 i = j = 0



B(i, j) = 
 A(i, j)
 otherwise

where A is an image matrix (A(i, j) ∈ {0, . . . , 255}).

• Normalizing B using C(i, j) = gsB (B(i, j)) will show an image (C) that is
completely black except for the pixel (0, 0) which will have the value
255.
• This loses all the information in C about A.
• The next two slides consider an example which addresses the problem
by point processing. The resulting point function is sometimes called
an emphasis/de-emphasis function.

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 24
 Example

•A good example of the “wide range” problem happens when we want

to show the 2-D Discrete Fourier Transform (DFT) of an image.
• >> F = round(fftshift(abs(fft2(A))));
• Let B(i, j) = log10(F (i, j) + 1) and C(i, j) = gsF (F (i, j)), D(i, j) = gsB (B(i, j)).

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 25
 Example - contd.

• Specific range stretching/compression might yield even better results. This is an effective tool to
visualize things quickly. Different problems can be tackled in the same spirit by utilizing point
functions based on log10 (. . .), log10 (log10 (. . .)), e... and so on.

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 26
 “Slicing”

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 27
 Thresholding

• Binary thresholding (T = 128)




 0 A(i, j) ≤ T
B(i, j) = 
  255 A(i, j) > T

• We will see more sophisticated thresholding later.

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 28
 Summary

• In this lecture we learnt about some simple structural processing tech-

niques like transposing, flipping and cropping.
• We learnt about simple statistics like sample mean and sample variance.
• Most importantly we learnt about histograms.
– Histograms tell us how the values of individual pixels in an
image are “distributed”.
– An image may have the same histogram as another image.
• We learnt about many point processing techniques. These will be-
come very handy as we gain more experience with images. Make sure
you read the textbook [1] (Chp. 7, pp. 233-241) for more examples.
• We learnt more about matlab, how to write functions, faster executing
scripts, etc.

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 29
 Homework Rules
• Homeworks will be read as opposed to check-marked.
• Remember that I have been doing this for a while. I usually can tell what kind of processing an
image underwent.
• Label all axis for plots and bars. Put a descriptive title on plots, bars and images.
• Be neat and always show the original image in a homework for easy comparison to processed
images.
• All utilized matlab scripts must be handed in with the homework.
• Try to conserve paper by organizing what you hand in.
• Write general matlab scripts and functions so that you can reuse them in the future. Try to make
them “bullet-proof” so that you avoid mistakes.
• Matlab scripts/functions must be commented with descriptive help information.
• When showing images, use the image, colormap(gray(256)), and axis(’image’) commands. There
is a command in matlab (imagesc) that does some automatic normalization. Do not use the
command imagesc.
• Remember the best way to view the notes and to work in matlab is after you have set your desktop
to 24 or 32 bits color depth. Otherwise you may see artificial artifacts.
• When instructed to give “brief comments”, “brief explanations” or “brief comparisons”, etc. please
be brief and to the point.

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 30
 Homework II

1. Flip your image horizontally and diagonally. Print the results.

2. Crop a portion of your image that mostly shows your head. Print the result and parameters used.
3. Calculate the mean, variance and histogram of your image. Show the histogram as a bar plot
next to your image (use the subplot command). Indicate the size of your image.
4. Compute the square-root of your image. Round it using the round function. Show the result
and its histogram. Now normalize the rounded image. Show the result and its histogram. Briefly
compare it to your original, unprocessed image.
5. Compute the square-root of your image. Normalize the result and print it together with its
histogram. Briefly compare it to the output of (4.) and to your original image. Which pixel value
ranges got stretched/compressed?
6. Compute B(i, j) = log10 (A(i, j) + 1) where A is your image. B is a general matrix and not an
image matrix. Generate an image matrix from B by normalizing it. Show the normalized image
next to the original. Briefly describe the changes. Anything that can be seen better? Which pixel
value ranges got stretched/compressed?
7. For an image A with histogram hA (l), sketch the histogram hB (l) for (a) B(i, j) = g1 (A(i, j))
and (b) B(i, j) = g2 (A(i, j)) where hA (l), g1 , g2 are as shown in Figure 1.

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 31
 Figure 1: Homework question 7

References

[1] A. K. Jain, Fundamentals of Digital Image Processing. Englewood Cliffs, NJ: Prentice Hall, 1989.

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 32