Description
1. (20 points.) First, write a few functions that will help with the rest of the homework. Turn in your code listings
for each function. Each function should only be a couple lines long.
(a) (5 points.) A function called fftshow that displays the log of the magnitude of the 2-D DFT of an image
and plots it so that the DC element is at the center of the frequency-domain image, and the result is scaled
to use the full grayscale range. Use the functions fft2 and fftshift.
(b) (5 points.) A function called boxfilt that takes as input a grayscale image im, an odd box width w and
an odd box heighth, and returns an ideal low-pass box frequency-domain filter the same size as im that
that passes the w frequencies around DC along the columns of the image and the h frequencies around
DC along the rows of the image. The DC element should be at the upper left corner of the output. You can
assume the image also has odd dimensions.
(c) (5 points.) A function called freqflt that applies a frequency-domain filter to an image by multiplying
the filter with the image’s 2D DFT, taking the inverse 2D DFT, and returning the real part of the result as a
uint8. The function should assume that the frequency-domain filter has the same size as the input image.
Note there is a function called ifftshift that moves the DC element back to the upper left corner, if you
need it.
(d) (5 points.) A function called gaussfilt that takes as input a grayscale image im and a 2×2 covariance
matrix C. The filter should return a low-pass frequency-domain Gaussian filter the same size as im centered
at DC with the given covariance matrix. This one’s a little tricky. Remember the formula for a 2D Gaussian
distribution (e.g., equation 2-123), and how you used meshgrid to construct a related image in HW 1. We
want the maximum element of the filter to be 1, so you only need to worry about the exp(···) part, not the
scaling factor in front. As above, the DC element should be at the upper left corner of the output.
2. (30 points.) Consider the input image tiger.png. In each of the following subproblems, provide a result (either
in the spatial domain or frequency domain as indicated) and explain why the result looks the way it does. For
frequency-domain results, use fftshift to make sure DC is in the middle of the image for easier visualization.
(a) (4 points.) Use fftshow to display the 2D DFT of the image.
(b) (4 points.) Construct a box filter with [w,h] = [31, 31]. Look at it in the frequency domain with
fftshift.
(c) (4 points.) Apply the above box filter to the input image. What does the result look like in the spatial
domain?
(d) (4 points.) Construct a box filter with [w,h] = [91, 91]. Look at it in the frequency domain with
fftshift, and apply it to the input image. How does the output image differ from part (c)?
(e) (4 points.) Construct a box filter with [w,h] = [91, 31] and apply it to the input image. How does the
result differ from parts (c) and (d)?
(f ) (5 points.) Construct a box filter with [w,h] = [91, 399] and apply it to the input image. What is the
interpretation of this filter? How does the result differ from part (e)?
(g) (5 points.) Construct a box filter with [w,h] = [599, 31] and apply it to the input image. What is the
interpretation of this filter? How does the result differ from part (e)?
More fun on the next page −→
3. (30 points.) We’ll again consider the input image tiger.png, but now look at the results of Gaussian filtering
instead of box filtering. In each of the following subproblems, provide a result (either in the spatial domain or
frequency domain as indicated) and explain why the result looks the way it does.
(a) (4 points.) Construct a Gaussian filter using gaussfilt with C =
·
100 0
0 100 ¸
and look at it in the frequency domain with fftshift. How does it compare to the box filter in Problem 2b?
(b) (4 points.) Apply the above Gaussian filter to the input image. What does the result look like in the spatial
domain?
(c) (4 points.) What does the above result look like in the frequency domain?
(d) (4 points.) Construct a Gaussian filter using gaussfilt with C =
·
900 0
0 900 ¸
and apply it to the input
image. How does the result differ from part (b)?
(e) (4 points.) Construct a Gaussian filter using gaussfilt with C =
·
900 0
0 100 ¸
and apply it to the input
image. Explain the results.
(f ) (5 points.) Now let R =
·
cos 30◦ −sin 30◦
sin 30◦
cos 30◦
¸
and D =
·
900 0
0 100 ¸
. Construct a Gaussian filter using
gaussfilt with C = R
>DR. What does this filter look like in the frequency domain?
(g) (5 points.) What is the result of applying this filter to the input image? Carefully explain the results.
4. (20 points.) Consider the high-boost spatial-domain filter given by 1
16
−1 −2 −1
−2 28 −2
−1 −2 −1
.
(a) (5 points.) In Matlab, display the magnitude of the 2D 256×256 DFT of this filter, centering so the DC term
is at the center of the image. Interpret what you see.
(b) (5 points.) Apply the filter to the image tiger.png and verify that this is indeed an edge-enhancing filter.
(c) (10 points.) With paper and pencil, explicitly compute the frequency response H(u, v) of the high-boost
filter. You should get a closed-form answer involving trig functions.