LECTURE 3

Chapter 3: Enhancement in the spatial domain

3.3 Histogram Processing

Histogram
h(rk ) = nk

• rk : kth gray level

• nk : number of pixels of gray level rk

Normalization ⇒ discrete PDF

p(rk ) = nk /n

• n: total number of pixels

L−1
X
p(rk ) = 1
k=0

Histogram equalization: 3.3.1

Histogram specification: 3.3.2

3.3 Histogram Processing Lecture 3 (page 2)

Examples
3.3 Histogram Processing Lecture 3 (page 3)

3.3.1 Histogram Equalization

First consider continuous functions and transformations of

the form

s = T (r), r ∈ [0, 1]

and assume that

(a) T (r) single-valued, monot. increasing for r ∈ [0, 1]
(b) T (r) ∈ [0, 1] for r ∈ [0, 1] Why?
Inverse transformation

r = T −1(s), s ∈ [0, 1]
3.3 Histogram Processing Lecture 3 (page 4)

View gray levels as random variables

• pr (r): continuous PDF of r
• ps(s): continuous PDF of s
If T −1(s) satisfies condition (a) then
¯ ¯
dr ¯¯¯
¯
¯
¯
ps(s) = pr (r) ¯¯¯¯
¯
ds ¯

Consider the transformation function

Z
r
s = T (r) = 0 pr (w) dw

RHS is the cumulative distribution function (CDF) of r,

and satisfies conditions (a) and (b). From Leibniz’s rule...

ds dT (r)
=
dr dr
d ½Z r ¾
= pr (w) dw
dr 0
= pr (r)
¯ ¯
dr ¯¯¯ ¯
¯
¯
⇒ ps(s) = pr (r) ¯¯¯ ¯
¯
ds ¯ ¯
¯
¯
1 ¯¯¯¯
¯
¯
= pr (r) ¯
¯
¯
¯
pr (r) ¯¯
¯

= 1, s ∈ [0, 1]
3.3 Histogram Processing Lecture 3 (page 5)

Z
r
Thus for T (r) = 0 pr (w) dw, ps(s) is always uniform

Now consider discrete values...

Recall
nk
p(rk ) = , k = 0, 1, 2, . . . , L − 1
n

The discrete version of

Z
r
s = T (r) = 0 pr (w) dw
is
k
X
sk = T (rk ) = pr (rj )
j=0
nj
k
X
= , k = 0, 1, 2, . . . , L − 1
j=0 n

and is called histogram equalization

NB: This will not produce a uniform histogram, but will tend
to spread the histogram of the input image

Advantages:
• Gray-level values cover entire scale (contrast enhancement)
• Fully automatic
3.3 Histogram Processing Lecture 3 (page 6)

Example 3.3: Histogram equalization

3.3 Histogram Processing Lecture 3 (page 7)

Example 3.3: Histogram equalization...

Transformation functions
3.3 Histogram Processing Lecture 3 (page 8)

3.3.2 Histogram Matching (Specification)

• Some applications: hist. equalization not best approach

• So, generate processed image with specified histogram

Development of the method

Again consider continuous gray levels r and z...

• pr (r): continuous PDF for input image
• pz (z): specified (desired) cont. PDF for output image

Let s be a random variable where

Z
r
s = T (r) = 0 pr (w) dw

and define z as a random variable where

Z
z
G(z) = 0 pz (t) dt = s,

thus
G(z) = T (r)
and
z = G−1(s) = G−1 [ T (r) ]
3.3 Histogram Processing Lecture 3 (page 9)

Procedure for continuous formulation

(1) Obtain T (r)

(2) Obtain G(z)
(3) Obtain G−1
(4) Pixels of input image → T → G−1 → Pixels of output
image

• Seldom possible to obtain analytical expressions for T (r)

and G−1
• Easier for discrete case, but only approximation for desired
histogram

Discrete formulation
Recall histogram equalization...
k
X
sk = T (rk ) = pr (rj )
j=0
k
X nj
= , k = 0, 1, 2, . . . , L − 1
j=0 n

For histogram specification, we have

k
X
vk = G(zk ) = pz (zi) = sk , k = 0, 1, 2, . . . , L − 1
i=0
3.3 Histogram Processing Lecture 3 (page 10)

Discrete formulation for histogram specification

k
X
sk = T (rk ) = pr (rj )
j=0
k
X nj
= , k = 0, 1, 2, . . . , L − 1
j=0 n

k
X
vk = G(zk ) = pz (zi) = sk , k = 0, 1, 2, . . . , L − 1
i=0

zk = G−1 [ T (rk ) ] = G−1 [ sk ], k = 0, 1, 2, . . . , L − 1

r0 −→ T −→ s0
r1 −→ T −→ s1
.. .. ..
rL−1 −→ T −→ sL−1

z0 −→ G −→ s0 = v0
z1 −→ G −→ s1 = v1
.. .. ..
zL−1 −→ G −→ sL−1 = vL−1
3.3 Histogram Processing Lecture 3 (page 11)

(1) Obtain histogram of given image;

(2) Precompute: rk −→ sk ; (3) Obtain G from given pz (z)
(4) Precompute: sk −→ zk . Use iterative scheme...
Let zk = ẑ for each k, where ẑ is the smallest integer in
the interval [0, L − 1] so that
(G(ẑ) − sk ) ≥ 0, k = 0, 1, 2, . . . , L − 1.
For k = k +1, start with ẑ = zk and incr. in integer values
(5) For each pixel, use (2) and (4): rk −→ sk −→ zk

• NB: None of the values of pz (zi) can be zero. Why?

3.3 Histogram Processing Lecture 3 (page 12)
3.3 Histogram Processing Lecture 3 (page 13)

• Histogram specification: trial-and-error process

3.3 Histogram Processing Lecture 3 (page 14)

3.3.3 Local Enhancement

• Previous methods (3.3.1 and 3.3.2) were global
• Define square or rectangular neighbourhood (mask) and
move the center from pixel to pixel
• For each neighbourhood...
• Calculate histogram of the points in the neighbourhood
• Obtain histogram equalization/specification function
• Map gray level of pixel centered in neighbourhood
• Can use new pixel values and previous histogram to calculate
next histogram

Example 3.5: Enhancement using local histograms

3.3 Histogram Processing Lecture 3 (page 15)

3.3.4 Use of Hist. Statistics for Image Enhancement

The nth moment of r (discrete) about its mean is defined as

L−1
X
µn(r) = (ri − m)np(ri)
i=0
where m is the mean value of r:
L−1
X
m= rip(ri)
i=0

Note that µ0 = 1 and µ1 = 0, and that µ2 is the variance

σ 2(r):
L−1
X
µ2(r) = (ri − m)2p(ri)
i=0

Mean: measure of average gray level

Variance: measure of average contrast

Let (x, y) be the coordinates of a pixel in an image, and Sxy

denote a subimage centered at (x, y):
X
mSxy = rs,tp(rs,t)
(s,t)∈Sxy
X ½ ¾
σS2 xy = rs,t − mSxy p(rs,t)
(s,t)∈Sxy
3.3 Histogram Processing Lecture 3 (page 16)

Example 3.6: Enhancement based on local statistics



 E · f (x, y) if mSxy ∈ [0, k0MG] AND σSxy ∈ [k1DG, k2DG]
g(x, y) = 
 f (x, y) otherwise

MG: Global mean

DG: Global standard deviation
3.3 Histogram Processing Lecture 3 (page 17)

Example 3.6: Enhancement based on local statistics

E = 4.0; k0 = 0.4; k1 = 0.02; k2 = 0.4; (3×3) local region