Description
Important!!! You may use imagesc() on this assignment!
(1) Use the data in medfilt_problem_dat.mat (variable will be called dat when you read it in)
for this exercise and it looks like Figure 1.
Figure 1: Image you’ll be using for median filter problem.
(a) (2 points) Write the code for a median filter using a 3 × 3 window, using zero padding.
Compare your result to what you get using medfilt2 in MATLAB. To check this, take the
difference of the two filtered images and inspect the values of the difference. Viewing the
images may not be sufficient, so check the values carefully. Are the images the same? Why
or why not?
1
(b) (1 point) You may use medfilt2 for this part. What is the minimum filter size that removes
all of the “salt and pepper” in the image?
(c) (2 points) I generated the noise in this image by randomly selecting pixels and adding noise
to a 5×5 window around that pixel. For an isolated piece of this “salt” or “pepper” (meaning
it isn’t overlapping or near another chunk of noise), what is the minimum window size that
guarantees it will be removed? You should derive this analytically. Did your solution match
what you found in the previous part? Why or why not?
(2) (5 points) Linear Hough transform. For this problem you’ll be working with the image shown
in Figure 2, contained in dat_hough.mat. The goal is to apply the linear Hough transform to bring
out the edges of the black shape. It is up to you to process the image to remove as much of the
artifact as possible, estimate the edges and generate a binary image (edges have a value of 1 and
everything else is 0). Then you apply the Hough transform as we did in class and you may use
the matlab function we used in class. You may not use the built in MATLAB Hough transform
function.
Please display the following images: Image after you removed the artifact (as much as possible),
image after edge detection (not yet binary), the binarized edge image and the result from applying
the Hough transform. You must use approaches learned in this class for each step and explain
why you chose the approaches. Display the first three images in a 3 × 1 panel plot and the Hough
transform result in a single panel plot that uses the imfuse() MATLAB function to combine the
Hough lines and original image (we did this in class).
Figure 2: Image you’ll be using for linear Hough Transform problem.
(3) (5 points) The Circle Hough transform. The circle Hough is very similar to the linear Hough,
but instead of constructing lines through each edge point, we construct circles centered at those
points. If the proper radius is chosen, the constructed circles will overlap at the center of the
circles you are trying to identify. See class notes for more details. Circles require three parameters:
x-coordinate of the center, y-coordinate of the center and the radius. So, unlike the linear Hough,
there are 3 parameters to search over instead of 2. You will create a 3D accumulator matrix (xcoordinate, y-coordinate and radius) that counts how often a constructed circle passed through an
(x,y) coordinate for a given radius. I will give you three radius values to try (48, 50 and 52). Your
original image is 540 × 720, so your accumulator matrix is 540 × 720 × 3. Use the img_circle.mat
2
file for this problem, which is also shown in Figure 3. There are two circle sizes in the image, but
your goal is to find the smaller circles. I have already binarized it, so no preprocessing is required
and you may simply start with the Hough transform.
Figure 3: Image you’ll be using for the circle Hough Transform problem.
No need to write a separate function, as I did for the Linear Hough in class. Here are the steps
that you will need to code. For each pixel that lands on an edge with coordinates, (i, j):
1. Construct a circle centered at that point. The following will build this circle
(a) For each radius, r, (use r=48, 50 and 52) do the following
(b) For each angle, t from t=0-359 do the following
i. x = i − r × cos(t ∗ pi/180) is the x-coordinate of a point on a circle centered at your
edge point
ii. y = j − r × sin(t ∗ pi/180) is the y-coordinate of a point on a circle centered at your
edge point
(c) Record this circle point in your accumulator matrix. If you named the accumulator
matrix accum, you would add a count to the accum(round(x), round(y), k) position,
where k refers to 1, 2 or 3, which is the radius you’re currently on. k = 1 would refer to
the radius of length r = 48, etc.
2. To identify the best radius, find the one that has the largest accumulation value. For example,
max(max(accum(:,:,1)) will be the max bin for the radius of 48, assuming you stored your
accumulation values for the 48 voxel radius in this part of the array. The largest max tells
you which radius is best. What was the best radius?
3. Now that you’ve found the best radius, you can see if it did a good job finding the smaller circle
centers. As with the linear Hough, you need to threshold the accumulator matrix. Unlike the
linear Hough, you don’t need to do anything to transform it back to the original image, you
simply need to view the two images together. Display the accumulator matrix for the best
radius, unthresholded, and also use imfuse to fuse the original image with your estimated
centers (thresholded accumulator matrix) and display using imagesc. All the smaller circles
should have a point in the center and you shouldn’t have points elsewhere if all went well! If
you didn’t get the centers or if you have many extra points that aren’t in the centers, try a
different threshold.
3