Image Processing I (Seminar)

The Parfor-Loop:
The runtime of a for-loop can sometimes take too long to be practical.
This is mainly due to the sequential mechanism used by for-loops.
To counter this, MATLAB provides a parallel processing option that can be
applied to for-loops: Parfor-loops.

A "parfor-loop" is a parallel for-loop, which means that each iteration of the loop
can be completed independently of the other iterations. This means that recursive
algorithms aren't well-suited for parallelization.
___________________________________________________________________________________
e.g.

Variable 'k' is evaluated recursively; its position in the loop's queue must be known.

for i=1:n
k=k+i;
end
____      ____      ____      ____      ____      ____      ____      ____      ____      ____      ____      ____

Variable 'k' is evaluated independently of its position in the loop's queue.

parfor i=1:n
k=i;
end
___________________________________________________________________________________

Utilizing a parallel loop can be a great speed advantage. To demonstrate this, we will time
computationally intensive tasks using a for-loop and a parfor loop.

First, begin by saving this image to your current MATLAB directory:

After this, take a look (just to see what you're working with). This is a good idea to do after every transformation:

imshow(A)

Now we're going to convert the image to grayscale (remember that RGB images are
3-dimensional, and we would like to simplifiy the process by transforming it into 2 dimensions):

B=rgb2gray(A);
imshow(B)

Now, we will create a binary image by using a logical operator to impose a global threshold:

C=B>T;
imshow(C)
(where ' T ' is a scalar threshold value)

This should return a black and white version of the original. Finally, our compuationally-intensive task awaits.
We will design a for-loop to count the total amount of 1's in the binary image 'C'.

k=0;
for i=1:length(C(:,1))
for j=1:length(C(1,:))

(fill in the blank)

end
end

This summation can be completed relatively fast once, but what if we needed to complete this loop...

...>1000 times?

(let n>1000)

for i=1:n

k=0;
for i=1:length(C(:,1))
for j=1:length(C(1,:))

(fill in the blank)

end
end

end

^^^ Using tic & toc, we can find this to be complete in a matter of seconds
See below for timing code:

tic
for i=1:n
k=0;
for i=1:length(C(:,1))
for j=1:length(C(1,:))

(fill in the blank)

end
end

end
toc

___________________________________________________________________________________
Let's compare this to the time required to
complete >1000 runs in-parallel

To begin a parallel for-loop, you must start a parallel pool. This can be done with the button
in the bottom-left corner of the MATLAB window:

Select "Start parallel pool"

Next, select the only other option "Parallel Preferences"

Make sure that 4 workers are specified
!!!
______________________________________________________________________________________________________
Now let's time the same loop, but in parallel:

(let n>1000)

%% (PREVIOUS FOR-LOOP)
tic
for i=1:n
k=0;
for i=1:length(C(:,1))
for j=1:length(C(1,:))

(fill in the blank)

end
end

end
toc

%% (PARFOR-LOOP)
tic
parfor i=1:n
k=0;
for i=1:length(C(:,1))
for j=1:length(C(1,:))

(fill in the blank)

end
end

end
toc

**note the difference in runtime!!**

DON'T FORGET ABOUT THE EXTRA CREDIT OPPORTUNITY!!!!!

Can you build a code that can reliably count     the circles in today's example image?

Here is an example of code that works (this is the same code shown in today's slides):

diam=2;
T=45;
vert_size=43;
horiz_size=22;
n=2;

w=size(A,2);
d=0;
dd=0;
for x  = 2:w
for= diam+1:vert_size
if A(i -diam:i, x - 1)==1
A(i -diam:i, x )=0;
end
end
for = 2:vert_size
if A(i , x )~=0
if A(i -1, x )==A(i , x )
d=d+1;
if d==diam
d=0;
dd=dd+1;
end
else, d=0;
end
end
end
end
dd