Region Processing

Question 1

Describe Convolution

Answer 1

• t(x,y) template
• at each image LOCATION
• SUM of products
• output VALUE

Question 2

How to normalize result of convolution template?

Answer 2

-normalize output directly (255*9/9)
or
-convolve with [1/9 1/9 1/9 …]

Question 3

Why is the distributivity of convolution useful?

Answer 3

efficiency gain

compute composite filter + convole > convolve,+ convole

Question 4

If convolving with a nxn template requires n^2 multiplications and additions per output pixel, how many does convolving with two nx1?

Answer 4

2n

Question 5

What are 3 applications of convolution?

Answer 5

• Smoothing (noise reduction)
• Sharpening (edge enhancement)
• Template matching (finding objects)

Question 6

Simplest way to smooth?

Answer 6

average neighbour values

Question 7

If you have a neighbourhood of m pixels to smooth what values would you put in the convolutional filter?

Answer 7

1/m

Question 8

Way to avoid smoothing all pixels?

Answer 8

Conditional smoothing

Replace pixel with smoothed version only if their difference is less than δ

Question 9

How does median smoothing work?

Answer 9

Pick median of set

Question 10

What does Gaussian smoothing help prevent?

Answer 10

ringing artefact (echoes of boundary)

Question 11

In summary which smoothing techniques (convolution) exist?

Answer 11

• average neighbourhood values
• conditional
• median
• gaussian

Question 12

What is the definition of an edge?

Answer 12

An edge, is an extended,
significant,
local change in image intensity

Question 13

Edge detection methods?

Answer 13

• Roberts Cross
• Prewitt/Sobel
• Canny/Deriche
• Laplacian of Gaussian
• Template matchingg

Question 14

What’s a different more common name for edge enhancement?

Answer 14

Sharpening

Question 15

What is the roberts/cross process of edge enhancement?

Answer 15

• enhance the edges of an image
• enhance the ones orthogonal to it
• combine these
• RESULT: edge magnitude image, tangent of orientation image

Question 16

What is the problem with roberts cross edge detector?

Answer 16

Noise sensitivity -> corruption of edge strength

Question 17

What edge enhancement operators are less sensitive to noise than Roberts Cross?

Answer 17

Prewitt/Sobe

Question 18

What is the issue with both Roberts/Cross and Prewitt/Sobe?

Answer 18

Scale σ

Question 19

What happens as the scale parameter is increased?

Answer 19

• all objects consired at first

- smaller objects discarded

Question 20

Process of edge enhancement with scale?

Answer 20

• smooth image (Gaussian template +σ)

- differentiation or double differentiation

Question 21

What are the requirements for the Canny/Deriche Edge Detector?

Answer 21

• accurate localisation

- single response to an edge

Question 22

Two definitions of “local”

Answer 22

• close PIXEL distance

- OBJECT

Question 23

Which operator uses the Second Derivative?

Answer 23

Marr-hildreth

Question 24

What does the Second Derivative improve?

Answer 24

Localizaiton of edge pixel

Question 25

When looking at double derivative graph, where is the edge location?

Answer 25

Where the graph crosses the axis (“zero-crossing”)

Question 26

What template does Marr-Hildreth use?

Answer 26

Laplacian of Gaussian

Question 27

What is template matching?

Answer 27

technique for measuring similarities-> FIND things

Question 28

How do we measure dissimilarity of template g and and image f?

Answer 28

sum(f-g)^2