CSD 212

Digital Image Processing

Image Enhancement :
Gray-Level Transformations
Part-2
 Gray-Level Transformations
●
Histogram processing
– What is a histogram ?
●
Probability theory refresher
– What can it achieve ?
 Gray-Level Transformations
●
Histogram processing
– What can it achieve ?
 Gray-Level Transformations
●
Histogram processing
– What can it achieve ?
 Gray-Level Transformations
●
Histogram processing
– What can it achieve ?
 Gray-Level Transformations
●
Histogram processing
– What can it achieve ?
 Gray-Level Transformations
●
Histogram processing
– What is a histogram ?
●
Probability theory refresher (very brief)
– Probability mass/density function
– Random variables
– Random experiments
– Sample space, events, probability functions
– Transformation of a random variable
 Gray-Level Transformations
●
Random experiment = an experiment whose
outcome isn’t certain
– Flip of a coin
– Throw of a die

●
Sample space = set of all possible outcomes of a
random experiment
– Coin flip : { head, tail }
– Die throw : { 1, 2, 3, 4, 5, 6 }
 Gray-Level Transformations
●
Random experiment
– Select a pixel location in an image
●
Sample space
– Set of all pixel locations
{ (1,1), (1,2), … , (1,8),
(2,1), (2,2), … , (2,8),
…
(8,1), (8,2), … , (8,8) }
 Gray-Level Transformations
●
Event is a subset of a sample space
– Event of getting a head on a coin toss = { head }
– Event of getting an even number on a die roll = ?
= { 2, 4, 6 }
– Event of getting a pixel in the left half of the image

●
Event space = set of all possible events
 Gray-Level Transformations
●
Probability function P(.)
– A probability function on sample space Ω
assigns every event A ⊆ Ω
a number ∈ [0,1] s.t.
● P (Ω) = 1
●
P (A U B) = P (A) + P (B) when A ∩ B = 0
– P(A) is the probability that event A occurs
– Examples
● Probability function for coin toss = ?

● Probability function for die roll = ?

 Gray-Level Transformations
●
Random variable = a function
– X:Ω→R
– Domain = sample space
– Range = set of real numbers
 Gray-Level Transformations
●
Random variable is an abstraction
– We don't care about outcomes (directly)
– We only care about values mapped to outcomes
– Examples
●
We don't care about die value, just even or odd
●
We don't care about pixel location, just pixel intensity !
 Gray-Level Transformations
●
Defining events using random variables
– Example: Sum of a pair of dice
– Sample Space: Ω = { (i,j) : 1 <= i,j <= 6 }
– Probability function: P ( (i,j) ) = 1/36
– Random variable: X ( (i,j) ) = i + j
– Event { X=a } = { w ∈ Ω : X(w)=a }
●
P (X = 5)
= P ( { (1,4), (2,3), (3,2), (4,1) } )
= P ((1,4)) + P ((2,3)) + P ((3,2)) + P ((4,1))
= 4 / 36
 Gray-Level Transformations
●
Defining events using random variables
– Example:
Intensity at a randomly-chosen pixel in 8x8 image f(x,y)
– Sample Space: Ω = { (x,y) : 1 <= x,y <= 8 }
– Probability function: P ( (x,y) ) = 1/64
– Random variable: X ( (x,y) ) = f (x,y)
– Event { X > 100 } = { w ∈ Ω : X(w) > 100 }
 Gray-Level Transformations
●
Discrete random variable (RV)
– Maps outcomes to values in a countable set
– Probability mass function p : R → [0,1]
● Probability that the RV takes a value 'a'
● p (a) = P (X=a) = ∑w: X(w) =a P(w)
– Cumulative distribution function F : R → [0,1]
● Probability that the RV takes a value less than x
● F (a) = P (X <= a) = ∑w: X(w) <= a P(w)
● F (a) = P (X <= a) = ∑b : b <= a P(X=b)
 Gray-Level Transformations
●
Continuous random variable (RV)
– Maps outcomes to values in an uncountable set
●
Sample space itself is uncountable
– e.g., continuous image function X(.) defined on [0,1]x[0,1]
– Probability density function p : R → R
●
p (x) is NOT probability (p(x) can be > 1)
● Events, of interest, are subsets of the form { x : a < x < b }
● p (a < x < b) = ∫a<x<b p (x) dx
– Cumulative distribution function F : R → [0,1]
●
Probability that the RV takes a value less than c
– p (x < c) = ∫x<c p (x) dx
– p (x < c) = ∫w: X(w) <= c P (w) dw
 Gray-Level Transformations
●
Histogram equalization : Motivation
 Gray-Level Transformations
●
Histograms are related to photograph exposure
– Aperture size (amount of light per unit time)
– Shutter speed (time of exposure)
 Gray-Level Transformations
●
Histograms are related to photograph exposure
– Aperture size
– Shutter speed
 Gray-Level Transformations
●
Histogram equalization
– Assume continuous distributions (and image)
– Given histogram p (a)
– Design function T (a) = b
such that
p (b) = uniform distribution
– Assume 0 < a < 1, 0 < b < 1
– Then, we want p (b) = 1

– Consider a monotonically increasing

intensity-transformation function T(.)
 Gray-Level Transformations
p(a)
●
Histogram equalization
T(a)

p(b)
– Let b = T(a) a

– Mass conservation implies: p(a) da = p(b) db

– Integrate left hand side from 0 to a'
– Integrate right hand side from 0 = T (a') = b'
– Left hand side = CDFa (a')
– Right hand side = CDFb (b')
 Gray-Level Transformations
●
Histogram equalization
– If we want p (b) = 1, then CDFb (b') = b'
– So, CDFa (a') = CDFb (b') = b'
– So, T (a) = CDFa (a)
– So, transformation function =
CDF of intensities in original image
 Gray-Level Transformations
●
Histogram equalization example (3-bit image)
Rounding required,
if output needs to
be 3 bit

= CDF(r) * 7
Gray-Level Transformations
Gray-Level Transformations
 Gray-Level Transformations
●
Why is the output histogram NOT exactly uniform ?
– Because of discretization of space and intensity
– Theoretically
●
Contradicts the assumption of continuous distributions
(i.e., continuous spatial domain, continuous intensities)
●
We used that to derive transformation for histogram equalization
– Practical problem (1 of 2)
●
Imagine 8-bit images → 256 intensities
●
Imagine an input image containing only 100 different intensities
● Uniform histogram must contain some pixels for each intensity
● How can we map 100 values to 256 values ?
– No systematic way to do that !
 Gray-Level Transformations
●
Why is the output histogram NOT exactly uniform ?
– Because the set of intensities in digital images is discrete
– Practical problem (2 of 2)
●
Imagine 8-bit images → 256 intensities
●
Imagine an input image containing 256 different intensities and
2560 pixels
●
Uniform histogram must contain an equal number of (10) pixels
for each intensity
●
If some intensity “a” in input image has 20 pixels,
then intensity b=T(a) in output image will have >= 20 pixels
●
How can we decide which 10 of those >=20 pixels should be
assigned intensity “b” ?
– No systematic way to do that !
 Gray-Level Transformations
●
Histogram matching
– Sometimes, we don't want to equalize
– But, we want to match the histogram of one image to
another chosen histogram (NOT uniform)
●
e.g., applications in medical imaging
– In that case, CDFa (a) = CDFb (b) implies
b = CDFb-1 (CDFa (a))
 Gray-Level Transformations
●
Histogram matching
 Gray-Level Transformations
●
Limitations of histogram equalization
– Images has lots of dark pixels or / and lots of bright pixels
– Then the “middle” intensities don't get contrast enhanced
– How can we get better output ? e.g., top right image

???

Hist. Eq.
 Gray-Level Transformations
●
Limitations of histogram equalization
– Images has lots of dark pixels or / and lots of bright pixels
– Then the “middle” intensities don't get contrast enhanced

Hist. Eq.
 Gray-Level Transformations
●
Adaptive histogram equalization (AHE)
– Localized analysis
– Algorithm:
For each pixel “p”
●
Construct a window of size NxN around the pixel
●
Perform (partial) histogram equalization within that window
(1) Compute histogram within window
(2) Compute CDF within window
(3) Map the intensity of the center pixel “p” based on the CDF
– Window size N is a user-defined parameter
● User-defined parameter is also called “free” parameter
● How to choose N ?
 Gray-Level Transformations
●
Adaptive histogram equalization (AHE)
 Gray-Level Transformations
●
Adaptive histogram equalization (AHE)
– Variations in local histograms
 Gray-Level Transformations
●
Input image Global Hist. Eq.

● On next slide:
AHE with square windows of widths 25, 64, 100, 200
AHE
AHE

Input Image Hist. Eq.

 Gray-Level Transformations
●
Adaptive histogram equalization (AHE)
 Gray-Level Transformations
●
Limitations of adaptive histogram equalization
– Small window size → bring out more detail
– Small window size → amplify noise in constant patches

●
How to try to get the best of both worlds ?
– We want to enhance texture, without amplifying noise
 Gray-Level Transformations
●
Contrast-limited adaptive histogram equalization
(CLAHE)
– Principle: Contrast amplification in the vicinity of a given
pixel value is given by:
●
Slope of the transformation function
= Slope of the CDF
●
But CDF slope depends on histogram (PDF) value
 Gray-Level Transformations
●
Contrast-limited adaptive histogram equalization
(CLAHE)
– Limits the amplification by:
(1) Clip the histogram at a predefined value (free parameter)
(2) Redistribute the mass uniformly throughout the range
(3) Then, compute the CDF
AHE
CLAHE Gray-Level Transformations
 Gray-Level Transformations
●
(CL)AHE is computationally expensive
– Computes PDF (+ CDF) at window around at every pixel
●
Approximations to (CL)AHE for speedup
(1) Split image into MxM non-overlapping tiles
●
e.g., M = 9
(2) Compute CDF for each tile
(3)Bilinearly interpolate transformation for all pixels based
on distances from tile centers
 Gray-Level Transformations
●
Approximations to CLAHE for speedup
(1) Split image into MxM non-overlapping tiles
(2) Compute CDF for each tile
(3)Bilinearly interpolate transformation for all pixels based
on distances from tile centers