1

Lecture 3
Image Enhancement

Instructor: Dr. Yaser Esmaeili

Email: [email protected]

Department of Computer Science and Software Engineering

Concordia University

Slides modified from materials provided by Drs. Yiming Xiao & Tien Bui
2

Image Enhancement: improving the quality of images

(make it better) or modifying image to bring out hidden
features. Two kinds: spatial domain and frequency
domain.

Image Restoration: correcting images subjected to

noise, blurs, distortions, atmospheric effects, etc
 Digital Image
Enhancement

In Spatial domain
(Transform on pixels) In Frequency domain
(Fourier transforms)

Intensity Spatial Filtering

Transformation (asked & decided by
(Pixel Processing) neighbor pixels)

Image Smoothing Sharpening

Negative Filters
Etc. Filters

Linear- Log-
Transform Transform
ation ation
Laplacian-
Median Etc. based Filters
Histogram filter (2nd derivative)
Equalization Average Gradient-
filter based Filters
(1st derivative)
4

Enhancement
equalization
Histogram

https://pyimagesearch.com/
Denoising

Restoration
J. V. Manjon, et al. Journal of Magnetic Resonance Imaging, 2010.
5

Image Enhancement I:
Pixel-wise Operations
6

Image Enhancement: Operating on the Pixels

• Make image “better” for a specific application
– The idea of “better” is somewhat subjective

• We distinguish two domains:

– Spatial or Pixel domain:
– Frequency Domain:

• For this section: Pixel Domain

– Operations on single pixel at a time
– Operations on groups of pixels (neighborhoods)

Pixel
4-Neighbors 8-Neighbors
 7

Simplest form of processing: Point Processing

Pixel
T

255
No change
S=T(r)

Image “negative”: s=L-1-r

Thresholding

Black 0 r 255
8

Negative of an image

Original Negative
9

Intensity Transformation: Contrast Stretching

8-bit image, low contrast

(r2,s2)

(r1,s1)

original [rmin, rmax]

(r1, s1) = (rmin, 0)

(r2, s2)=(rmax, L-1)
rmin = rmin of original image

(r1, s1) = (m, 0)

(r2, s2)=(m, L-1)
stretching thresholding
10

The locations of and control the shape of the

transformation function.
a) If and the transformation is a linear
function: no change in intensity.
b) If = m, and the transformation is
a thresholding function: a binary image.
c) In general, (why?)

m
11

Intensity Transformation: Intensity Slicing

The aim is to highlight specific range of intensity of interest.

original processed
12

Simplest form of processing: Point Processing

Pixel
T

255 Common Examples:

• Dynamic Range Compression

s
Gamma Correction

Narrow range of “dark” gets

mapped to broad range of “gray”
0 Black r 255
13

Origins of gamma correction:

• Nonlinear response of CRT’s and imagers

• To correct for this in image display, the images or

commands to the CRT are “pre-distorted”

Luminance

0 Applied/Measured Voltage (U)

14

Gamma Correction
15
16

Fracture

original γ=0.6

γ=0.4 γ=0.3
17

original γ=3

γ=4 γ=5
18

Intensity Transformation: Bit-plane Slicing

Ex. In 8-bit image, grey level 194 is (1 1 0 0 0 0 1 0).

Grey levels (0-127) have the 8th bit = 0, and levels
(128-255) have the 8th bit = 1 (counting bits from right to left)

1
1
0
0
0
0
1
0
194
19

Intensity Transformation: Bit-plane Slicing

Bit-8 Bit-7

Bit-6 Bit-5 Bit-4

Bit-3 Bit-2 Bit-1

20

Histogram Processing

• Distribution of gray-levels can be judged by

measuring a Histogram
15

9
7
5
Histogram:

Graylevel
21

Histogram Processing
22

Histogram Processing:

For B-bit image,

• Initialize 2B counters with 0

• Loop over all pixels x,y
• When encountering gray level f(x,y)=i, increment the
counter number i

• With proper normalization, the histogram can be interpreted

as an estimate of the probability density function (pdf) of the
underlying random variable (the gray-level)

• You can also use fewer, larger bins to trade off amplitude
23

Histogram Processing

One-to-one mapping: no information destruction

24

Example:
25
Take a break!
26
27

Histogram Equalization
• The general idea: map the histogram of the
given image to a flat histogram by performing
a nonlinear operation on the gray value at
each pixel.

• Nonlinear Transformation: s = T(r)

• Questions:
– What is the right transformation?
– How do we find it, given a particular image.

• Analysis for the continuous grayscale first

28

Histogram Equalization
• Consider the histogram of the given (continuous grayscale)
image as a pdf p(r), where r is in the interval [0,L-1].eg L=256
– Recall that as a pdf we have

– Any pixel operation T(r) should map [0,L-1] to [0,L-1]

• Desired properties of T(r):

– Keep the black/white order (T(r) should be monotonic increasing)
– T(r) should be single valued (one-to-one), hence invertible.

• Question: Given image with histogram p(r), what does

histogram of s=T(r) look like in general?
29

Histogram Equalization
• Consider continuous values of intensity
• r = gray level of image to be equalized in the range [0, L-1]
0 : black
L-1: white
Goal is to design a transformation

Assumption about T(r) :

(a)T(r) is single valued and monotonically increasing in the
interval
(b)

Condition (a) assures that

30

• From Probability theory: (See Review Slides p. 36)

• We consider Cumulative Distribution Function (CDF) as a

Transformation

• If T(r) is just a CDF or the integral of the input pdf

then applying results in an image whose pdf is uniform.

(For PDF and CDF see pages 32-34 of the Review Slides)
31

Discrete Case

max intensity
new value input value
32

Example

s0 = 1
s1 = 3
s2 = 5
s3 = 6
s4 = 6 8 grey levels
s5 = 7 n = MxN
s6 = 7
s7 = 7

Example: 3-bit image, L= 8, image size = n = 64x64 = 4096, intensity levels [0,7].
Therefore, p0(r0) = 790/4096 = 0.19, etc…

Note that the gray levels are integers so the values of the equalized histogram
are: s0 = 1, s1 = 3, s2 = 5, s3 = s4 = 6, s5 = s6 = s7 = 7.
33
34

Histogram Equalization (examples)

35
36

Generalizations

• Adaptive Histogram Equalization

• Histogram Matching
– Instead of flattening the histogram, we
want to make the histogram of A look like
the histogram of another image B.

– Procedure (given two images):

• Find transformation (T and U) that flatten
images A, B
• Required transformation is U-1(T(A))
37

Histogram Matching
• Rather than uniform we want to match as closely as
possible to a desired, given pdf.
• r = pixel value before matching
• z = pixel value after matching
• We can compute from the given image, and is given.
What is the Transformation from r to z ?
• In the continuous case:

• Can form Look-up table

38

Approach

Method:
39
Approach

(L-1)

original target

transform Result
40
original target
Approach

transform Result

Method:

s0 = 1, s1 = 3, s2 = 5, s3 = s4 = 6, s5 = s6 = s7 = 7.

G0 = G1 = G2 = 0, G3 = 1 (round-off), G4 = 2, G5 = 5, G6 = 6, G7 =7.
41

Histogram Matching (example)

target
original

transform Result
42

Histogram Matching (example)

43

Comparison of Histogram Equalization and Histogram Matching

Histogram
Equalization

More details
44

Histogram Matching (example)

target
histogram

Transform
Original
histogram

Result

original
45

Local Enhancement using Local Histogram

• Define a “window” as a rectangle neighborhood.
• Move centre of windows across the image. At each
location, compute histogram in the window and do
histogram equalization.

scikit-image.org
46

Local Enhancement using Local Histogram

https://www.geeksforgeeks.org/adaptive-histogram-equalization-in-
image-processing-using-matlab/

• Ridge artifacts due to window selection

• Some constraints may be needed to ensure results
 Local Enhancement using Local Histogram

Using histogram statistics for image enhancement

49

Multi-Image “Averaging”
• It may be possible to obtain multiple noise-
corrupted images of a scene and average them
to obtain a less noisy result.
• It is important that there be no motion between
the frames, or that the frames be aligned first.

Given:

Compute:
50

Simple Averaging
Example: Additive noise =
Gaussian with zero
mean & standard
deviation of 64 grey
Original add noise levels

k=8 k=16
Averaging k times

k=64 k=128
51

Character of image noise via averaging

Large N for averaging is desirable but costly

52

Optimal Multi-image “Averaging”

Given:

Compute:

Given the noise statistics of the measured images, what should these
coefficients be to yield the “best” results?

Inversely proportional to the variance of noise in each image

53

MRI scan reconstruction from multiple sensor data

Adaptive averaging for image reconstruction

weights

(Hamilton et al. Europe PMC, 2017 )

54

Summary
• Point processing
✓ Contrast stretching
✓ Intensity slicing
✓ Gamma correction
✓ Bit-plane slicing
• Histogram processing
✓ Histogram equalization (global & local)
✓ Histogram matching
• Multi-image averaging
✓ Simple averaging
✓ Optimal/adaptive averaging
55

Reading from textbook

Chapter 3: Page 120-153

56

Food for thought

• For the bit map, how will the image histogram change if we
remove the lowest-order bit?
• When we preform histogram equalization, why it is very rare that a
perfectly flat histogram can be obtained for the resulting image?
• Why histogram matching often won’t result in the exact histogram
as the template image?
• In what scenario is local histogram equalization beneficial?
• If we want to improve image contrast, will gamma correction be a
good choice? compared to histogram equalization?
• When using averaging of N images to reduce noise, if the standard
deviation of the Gaussian noise is S, what is the standard deviation
of the noise in the noise-reduced image?

https://www.youtube.com/watch?v=-RDYHXwJ1LY