Assignment 2 Solution

EIE 529 - Digital Image Processing, Fall 2020-21

Department of Electronic and Information Engineering,
The Hong Kong Polytechnic University

Q1: Image Restoration (25 marks)

Solution
1) We first take the 2D Fourier transform on both sides of 𝑔(𝑥, 𝑦) = ℎ(𝑥, 𝑦) ∗ 𝑓(𝑥, 𝑦) +
𝜂(𝑥, 𝑦), and the image degradation model in the frequency domain is expressed as the
following expression:
𝐺(𝑢, 𝑣) = 𝐻(𝑢, 𝑣)𝐹(𝑢, 𝑣) + 𝑁(𝑢, 𝑣),
where 𝐹(𝑢, 𝑣) is the spectrum of the original image, 𝐻(𝑢, 𝑣) is the frequency response of the
point spread function, and 𝑁(𝑢, 𝑣) is the noise spectrum. By using the definition of 𝑃𝑔𝑔 (𝑢, 𝑣),
we have:
𝑃𝑔𝑔 (𝑢, 𝑣) = 𝔼[|𝐺(𝑢, 𝑣)|2 ] = 𝔼[𝐺(𝑢, 𝑣)𝐺 ∗ (𝑢, 𝑣)]
∗
= 𝔼[(𝐻(𝑢, 𝑣)𝐹(𝑢, 𝑣) + 𝑁(𝑢, 𝑣))(𝐻(𝑢, 𝑣)𝐹(𝑢, 𝑣) + 𝑁(𝑢, 𝑣)) ]
= 𝔼[𝐻(𝑢, 𝑣)𝐹(𝑢, 𝑣)𝐻 ∗ (𝑢, 𝑣)𝐹 ∗ (𝑢, 𝑣) + 𝑁(𝑢, 𝑣)𝑁 ∗ (𝑢, 𝑣)]
= |𝐻(𝑢, 𝑣)|2 𝔼[𝐹(𝑢, 𝑣)𝐹 ∗ (𝑢, 𝑣)] + 𝔼[𝑁(𝑢, 𝑣)𝑁 ∗ (𝑢, 𝑣)]
= |𝐻(𝑢, 𝑣)|2 𝑃𝑓𝑓 (𝑢, 𝑣) + 𝑃𝜂𝜂 (𝑢, 𝑣).

(5 marks)
2) Note that 𝑓(𝑥, 𝑦) has a size of 𝑀 × 𝑁 (𝑥 = 0, ⋯ , 𝑀 − 1, 𝑦 = 0, ⋯ , 𝑁 − 1) where 𝑀, 𝑁 are
even numbers, and 𝑔(𝑥, 𝑦) = 𝑓(𝑥, 𝑦) + 2𝑓(𝑥 − 1, 𝑦) + 𝑓(𝑥 − 2, 𝑦) + 𝜂(𝑥, 𝑦):
a) We first obtain the 𝑀 × 𝑁 2D-DFT on both size of 𝑔(𝑥, 𝑦) = 𝑓(𝑥, 𝑦) + 2𝑓(𝑥 −
1, 𝑦) + 𝑓(𝑥 − 2, 𝑦) + 𝜂(𝑥, 𝑦) as follows:

2𝜋 2𝜋
𝐺(𝑢, 𝑣) = 𝐹(𝑢, 𝑣) + 2𝑒 −𝑗 𝑀 𝑢 𝐹(𝑢, 𝑣) + 𝑒 −𝑗 𝑀 2𝑢 𝐹(𝑢, 𝑣) + 𝑁(𝑢, 𝑣)
2𝜋 2𝜋
= 𝐹(𝑢, 𝑣) (1 + 2𝑒 −𝑗 𝑀 𝑢 + 𝑒 −𝑗 𝑀 2𝑢 ) + 𝑁(𝑢, 𝑣).

As a result, the frequency response of the image degradation process is given by

2𝜋 2𝜋 2𝜋 2𝜋 2𝜋 2𝜋
𝐻(𝑢, 𝑣) = 1 + 2𝑒 −𝑗 𝑀 𝑢 + 𝑒 −𝑗 𝑀 2𝑢 = 𝑒 −𝑗 𝑀 𝑢 (2 + 𝑒 𝑗 𝑀 𝑢 + 𝑒 −𝑗 𝑀 𝑢 ) = 2𝑒 −𝑗 𝑀 𝑢 (1 +
2𝜋𝑢 1
cos ( )). Therefore, the inverse filter is given by 𝑅1 (𝑢, 𝑣) = 𝐻(𝑢,𝑣) =
𝑀
1
2𝜋
2𝜋
, and the restored image by using the inverse filter is given as follows:
−𝑗 𝑢
2𝑒 𝑀 (1+cos( 𝑢))
𝑀

1
 𝐺(𝑢, 𝑣)
𝐹̂1 (𝑢, 𝑣) = 𝐺(𝑢, 𝑣)𝑅1 (𝑢, 𝑣) = .
2𝜋 2𝜋𝑢
−𝑗 𝑢
2𝑒 𝑀 (1 + cos ( 𝑀 ))

(6 marks)
b) The frequencies for which the restored image cannot be estimated using the inverse
𝑀
filtering technique are 𝑢 = and 𝑣 = 0,1, ⋯ , 𝑁 − 1.
2
(4 marks)
c) Since 𝑝(𝑥, 𝑦) ∗ 𝑓(𝑥, 𝑦) = −𝑓(𝑥, 𝑦 − 1) + 2𝑓(𝑥, 𝑦) − 𝑓(𝑥, 𝑦 + 1) (See lecture notes
on Image Enhancement), the frequency response of 𝑝(𝑥, 𝑦) can be derived using
2𝜋 2𝜋
similar methods in 2), which is given by 𝑃(𝑢, 𝑣) = 2 − 𝑒 −𝑗 𝑁 𝑣 − 𝑒 𝑗 𝑁 𝑣 = 2 −
2𝜋𝑣
2cos ( ). As a result, the image restoration filter based on the regularization approach
𝑁
is given as:

𝐻 ∗ (𝑢, 𝑣)
𝑅2 (𝑢, 𝑣) =
|𝐻(𝑢, 𝑣)|2 + 𝜆|𝑃(𝑢, 𝑣)|2
2𝜋 2𝜋𝑢
𝑒 𝑗 𝑀 𝑢 (1 + cos ( 𝑀 ))
= 2 2.
2𝜋𝑢 2𝜋𝑣
2 (1 + cos ( 𝑀 )) + 2𝜆 (1 − cos ( 𝑁 ))

The restored image is given as:

2𝜋 2𝜋𝑢
𝐺(𝑢, 𝑣) ⋅ 𝑒 𝑗 𝑀 𝑢 (1 + cos ( 𝑀 ))
𝐹̂2 (𝑢, 𝑣) = 𝐺(𝑢, 𝑣)𝑅2 (𝑢, 𝑣) = 2 2.
2𝜋𝑢 2𝜋𝑣
2 (1 + cos ( 𝑀 )) + 2𝜆 (1 − cos ( 𝑁 ))

(6 marks)
d) The frequency/frequencies for which the restored image cannot be estimated using the
2𝜋𝑢 2𝜋𝑣
regularization approach are the ones with cos ( ) equals to -1 and cos ( ) equal to
𝑀 𝑁
𝑀
1 (so that the denominator is zero), which is 𝑢 = and 𝑣 = 0.
2
(4 marks)

Q2: Image Coding and Compression (25 marks)

Solution
1) Note that the distortion of the Lloyd-max quantizer D is a function of {𝑟𝑘 }𝐿𝑘=1 and {𝑑𝑘 }𝐿𝑘=0 .
a) The optimal values of 𝑑0 and 𝑑𝐿 to minimize the distortion should be 𝑑0 = 𝑎 and 𝑑𝐿 =
𝑏. This can be proved by contradiction.
(2 marks)

2
 b) To minimize 𝐷 by optimizing the variables {𝑟𝑘 } and {𝑑𝑘 } , we need to take the
derivative of 𝐷 with respect to these variables. In particular, to show the first expression,
𝜕𝐷
we need to derive 𝜕𝑑 and set it to zero, i.e.:
𝑘

𝜕𝐷
= (𝑑𝑘 − 𝑟𝑘 )2 𝑝(𝑑𝑘 ) − (𝑑𝑘 − 𝑟𝑘+1 )2 𝑝(𝑑𝑘 ) = 0 ⇒ −2𝑑𝑘 𝑟𝑘 + 2𝑑𝑘 𝑟𝑘+1 + 𝑟𝑘2 − 𝑟𝑘+1
2
=0
𝜕𝑑𝑘
𝑟𝑘 + 𝑟𝑘+1
⇒ 2𝑑𝑘 (𝑟𝑘+1 − 𝑟𝑘 ) − (𝑟𝑘 + 𝑟𝑘+1 )(𝑟𝑘+1 − 𝑟𝑘 ) = 0 ⇒ 𝑑𝑘 = ,𝑘
2
= 1, ⋯ , 𝐿 − 1.
𝜕𝐷
To obtain the second expression, we need to derive 𝜕𝑟 and set it to zero, i.e.:
𝑘

𝑑𝑘 𝑑𝑘 𝑑𝑘 𝑑𝑘
𝜕𝐷
= ∫ −2(𝑥 − 𝑟𝑘 )𝑝(𝑥)𝑑𝑥 = 0 ⇒ ∫ 𝑥𝑝(𝑥)𝑑𝑥 = ∫ 𝑟𝑘 𝑝(𝑥)𝑑𝑥 ⇒ ∫ 𝑥𝑝(𝑥)𝑑𝑥
𝜕𝑟𝑘 𝑑𝑘−1 𝑑𝑘−1 𝑑𝑘−1 𝑑𝑘−1
𝑑
𝑑𝑘 ∫𝑑 𝑘 𝑥𝑝(𝑥)𝑑𝑥
𝑘−1
= 𝑟𝑘 ∫ 𝑝(𝑥)𝑑𝑥 ⇒ 𝑟𝑘 = 𝑑𝑘 , 𝑘 = 1, ⋯ , 𝐿.
𝑑𝑘−1 ∫𝑑 𝑝(𝑥)𝑑𝑥
𝑘−1

(8 marks)
2) Given that a coding system with an information source that generates 5 symbols
{𝑎1 , 𝑎2 , 𝑎3 , 𝑎4 , 𝑎5 }, with the probabilities of {0.1,0.15,0.25,0.20,0.30}, respectively.

a) The entropy of the information source can be obtained via the entropy formula:

5
𝐻 = −∑ 𝑃(𝑎5 )log 2 𝑃(𝑎5 )
𝑗=1
= −(0.1 × log 2 0.1 + 0.15 × log 2 0.15 + 0.25 × log 2 0.25
+ 0.2 × log 2 0.2 + 0.3 × log 2 0.3) = 2.2282 (bits).
(3 marks)
b) The Huffman coding process is shown below. The codewords the coding system would
generate if the following input sequence is received {𝑎1 , 𝑎2 , 𝑎4 , 𝑎5 , 𝑎3 }, is given by
101100110001. The average number of bits for coding this input sequence (i.e., the
3+3+2+2+2
average number of bits for coding a symbol) is = 2.4 bits/per symbol.
5

3
 (6 marks)
c) The arithmetic coding process is shown below, and the final message symbol in the
input sequence narrows the range to [𝟎. 𝟎𝟏𝟗𝟖𝟐𝟓, 𝟎. 𝟎𝟐𝟎𝟎𝟓), i.e., any number within
this subinterval (e.g., 0.02000) can be used to represent the message.

(6 marks)
For your information, the value having the shortest binary form in the range
[0.019825,0.02005) is “.00000101001” in binary form, which is 0.02 in decimal.