Assignment 1

Addis Ababa University
Department of Computer Science

Postgraduate Program
Digital Image Processing (COSC 6012)
Assignment 1: Filtering May 2019
1. Smoothing images with a Gaussian filter
The 2D Gaussian function

G(x,y) = exp(−(x2+ y2) /2σ2)
= exp(−x2/2σ2 ) · exp(−y2/2σ2)
= G(x) ·G(y)
constitutes a very convenient low-pass (smoothing) filter. Notice that we have constructed G(x,y) as
the product of two 1D Gaussians G(x) and G(y) showing separability. This is a very desirable property,
since it allows fast implementation of filtering by two consecutive 1D convolutions. The fact that the
same parameter has been used for both the x and the y standard deviations entails the circular
symmetry of G(x,y). This results in the filter being isotropic, meaning that all directions in the image
are treated in the same way.
a) Separability
Make a 2D smoothing filter of size 9 x 9 and with a standard deviation of 1.5 by sampling the
Gaussian function f(x,y).
std=1.5; % standard deviation

[x,y]=meshgrid(−4:4,−4:4); %sampling grid
g=exp(−(x.*x + y.*y)/(2*std*std)); %2D smoothing filter
g=g/sum(sum(g)); % sum of weights equals one
Sample the 1D Gaussian functions f(x) and f(y) to make the 1D filters of size 1x9 and 9x1, respectively.
x1=−4:4; %sampling grid

gx= exp(−(x1.*x1 )/(2*std*std)); %smoothing filter in the x-direction
gx=gx/sum(gx); % sum of weights equals one
gy=gx’; %smoothing filter in the y=direction (transpose of gx)
subplot(3,1,1); stem(gx); %show the 1D smoothing filter
Verify the separable property of the Gaussian filter by computing gg=gx*gy. Use
"gg=conv2(gx,gy);" for the computation. Show the filter weights for g (2D filter) and for gg
(convolution of gx and gy). Compare the weights!
How do you use the separable property when filtering an image?
Page 1 of 4
b) Low-pass filtering (smoothing)
Load the image “blood1.tif” and filter the image by using the 2D filter g.
f=imread('blood1.tif');
f=double(f(1:256,1:256));
y=conv2(f,g,’valid’); %filter the image
subplot(2,2,1); imshow(f/255); % original image
subplot(2,2,2); imshow(y/255); % smoothed image
How many products are required when filtering the image using the 2D filter?
Now, use filter separability to filter the image by means of the two 1D filters gx and gy ((f*gx)*gy).
Control (by computing the sum of the squared differences) that you obtain the same result you
obtained when you filtered the image with the 2D filter g.
How many multiplications are used when filtering the image using the 1D filters?
Make low-pass filters gx and gy of larger sizes (1x15, 15x1, std=3) and filter the image “blood1.tif”.
Explain how the smoothing effect varies when the size of the filter is changed.
c) Noise reduction through smoothing
Add Gaussian noise to the image “blood1.tif”.
ff=imnoise(f/255,’gaussian’,0,0.01).*255; %add noise
Filter the image by using gx and gy.
What is the result of the smoothing operation?
Extract the subpart of the original image and of the noisy image that also appear in the filtered
image.
Compare using the sum of the squared errors:

noisy image vs. original image
filtered image vs. original image
Which one is closer to the original? Why?
Try reducing the variance of the noise to 0.001. Does filtering still help? Why?
Page 2 of 4
2. Derivative filters and the extraction of edges
The partial derivatives of the Gaussian filter G(x,y) can be used (but for a sign change) as derivative
(edge) filters in the x and y directions of the image.
-∂/∂x{G(x,y)} = −∂/∂x{G(x)} G(y)
= −∂/∂x{exp(−x2/2σ2 )} exp(−y2/2σ2)
= x/σ2 exp(−x2/2σ2 ) exp(−y2/2σ2)
= −G'(x) G(y)
-∂/∂y{G(x,y)} = −G(x) ∂/∂y {G(y)}

= exp(−x2/2σ2) ∂/∂y{exp(−y2/2σ2 )}
= y/σ2 exp(−y2/2σ2 ) exp(−x2/2σ2)
= −G(x) G'(y)
where G'(x) is the derivative of the 1D Gaussian function G(x).

r
The gradient of the image is defined as: ∇I = (∂I / ∂x, ∂I / ∂y ) . The magnitude of the gradient is large
at the location of edges in the image.
a) Derivative (edge) filtering
By sampling the 1D Gaussian function G(x) and its derivative G'(x) and using separability, we can
construct derivative filters along the x and y directions of the image.
std=1.5; %standard deviation
x2=−4:4; %sample grid

dx= x2/(std*std).*exp(−(x2.*x2)/(2*std*std)); %derivative in the x-direction
dy=dx’; %derivative in the y-direction
x1=−4:4; %sample grid

gx= exp(−(x1.*x1 )/(2*std*std)); %smoothing filter in the x-direction
gx=gx/sum(gx); % sum of weights equals one
gy=gx’; %smoothing filter in the y-direction (transpose of gx)
subplot(3,1,1); stem(gx); %1D smoothing filter

subplot(3,1,2); stem(dx); %1D derivative filter
Filter the image “blood1.tif” in the x-direction ((f*dx)*gy) and in the y-direction ((f*dy)*gx).
Show the original image and the filtered images (x and y derivatives).
r
Compute the magnitude of the gradient: ∇I = (∂I / ∂x) 2 + (∂I / ∂y ) 2 and display the resulting
(edge-filtered) image.
Page 3 of 4
b) Edge filtering in the presence of noise
Add Gaussian noise to the image “blood1.tif”.
ff=imnoise(f/255,’gaussian’,0,0.01).*255; %add noise
Compute the magnitude of the gradient.
Show the edge filtered image.
Is edge filtering sensitive to noise?
Does the fact that derivation is performed using derivatives of Gaussians help?
Page 4 of 4

Assignment 1

Uploaded by

Copyright:

Available Formats

Assignment 1

Uploaded by

Document Information

Original Description:

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

Assignment 1

Uploaded by

Copyright:

Available Formats

Addis Ababa University

Department of Computer Science

1. Smoothing images with a Gaussian filter

The 2D Gaussian function

std=1.5; % standard deviation

x1=−4:4; %sampling grid

How do you use the separable property when filtering an image?

c) Noise reduction through smoothing

Add Gaussian noise to the image “blood1.tif”.

ff=imnoise(f/255,’gaussian’,0,0.01).*255; %add noise

Filter the image by using gx and gy.

What is the result of the smoothing operation?

Compare using the sum of the squared errors:

Which one is closer to the original? Why?

-∂/∂y{G(x,y)} = −G(x) ∂/∂y {G(y)}

where G'(x) is the derivative of the 1D Gaussian function G(x).

a) Derivative (edge) filtering

std=1.5; %standard deviation

x2=−4:4; %sample grid

x1=−4:4; %sample grid

subplot(3,1,1); stem(gx); %1D smoothing filter

Add Gaussian noise to the image “blood1.tif”.

ff=imnoise(f/255,’gaussian’,0,0.01).*255; %add noise

Compute the magnitude of the gradient.

Show the edge filtered image.

Is edge filtering sensitive to noise?

You might also like

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.