Summary of Lecture #1

● Convolution describes Linear Time Invariant systems

● Sampling and reconstruction are LTI systems
● Imaging devices often are LTI systems
● Fourier Domain useful for analysing LTI systems
● Convolution in Spatial Domain is multiplication in
Fourier Domain

f  x ⊗h x  ℱ F H 

⇒

We will built on this for today's lecture

 Filtering
● Smoothing:
– Linear (low-pass filtering) & non-linear
– Reduce noise
– Down-sample (zoom out)
– Sharpen
– Abstract (simplify)
●
1st & 2nd order derivatives:
– Linear (non-linear exist but we do not discuss)
– Detect edges
– Detect lines
● Normalised convolution
● Next lecture: detection & analysis
 Filtering Properties
● Shift invariant:
– The result of the filter is independent of location within the
image
● Rotation invariant:
– The result of the filter should be independent of the
orientation of the image w.r.t. the axes
● These two rules make the result dependent only on the
object being imaged, not on the exact positioning of the
imaging system
 Linear Smoothing in 1-D
● Low-pass filtering: removing high-frequency
components
smoothing kernel frequency response

ideal (sinc)

uniform (box)

triangle
(=box⊗box)

Gaussian

(note: book pg. 139, eq. 5.47 needs normalization: 1/2πσ2 in 2D)
Phase Reversal of Uniform Filter
uniform Gaussian

reversed!
 Other Linear Low-Pass Filters
● Square neighbourhood, significance of centre pixel
increased:

[ ] [ ]
1 1 1 1 2 1
1 1
1 2 1 2 4 2
10 16
1 1 1 1 2 1
– Book says: “it better approximates the properties of noise
with a Gaussian probability distribution.” (pg. 125) Wrong!
– You now now why these filters are used!
● The Butterworth Filter
– Compare: Chebyshev, Elliptic, etc.
– Designed for electric circuitry
– But: electric circuitry has different constraints!
– Has no purpose in digital signal/image processing
 Linear Smoothing in n-D
● Separable filters vs. non-separable h=h x ⊗h y

● Convolution is associative f ⊗ { h x ⊗h y }= { f ⊗h x }⊗h y

● Separable filters are faster & easier to implement

the square uniform kernel is separable

G x ⊗G  y =G  x , y  the circular uniform kernel

the Gaussian kernel is separable is not separable
Linear Smoothing in n-D
square uniform
(separable) circular uniform
(not separable)

Gaussian
(separable)
Impulse
response
in Fourier
domain
Linear Smoothing in n-D
square uniform
(separable) circular uniform
(not separable)

Gaussian
(separable)
Impulse
response
in Fourier
domain
 Median Filter
● Takes the median of the values in some neighbourhood
● Not separable
● Median is an estimator for the mean
– Better than mean (linear filters) for some noise models
– Conserves edges slightly better than linear filters
● Generalizes to percentile filter
– Median is 50%
– 0% is min filter ( = erosion, Robin's lecture on
mathematical morphology )
– 100% is max filter ( = dilation, Robin's lecture )
– Other useful filters: 5%, 95% (like min and max, but less
sensitive to noise)
 Median Statistics
one input image mean of 5 images median of 5 images

Normally
distributed
noise

Salt &
pepper
noise
 Median Statistics
input image 3x3 uniform filter 3x3 median filter

Normally
distributed
noise

Salt &
pepper
noise
 Max-Min and Min-Max Filter

max-min min-max
(=closing) (=opening)
 Sequence of Max and Min Filters
● max min min max f ≈ min max max min f
● Removes local maxima and minima
● Apply first with very small neighbourhood
● Apply repeatedly with increasing size
 Application: Unsharp Masking
● Subtract smoothed image from original image
● Darkroom technique: implemented by projecting out-of-
focus image onto a negative, then using the two
negatives together
spatial domain frequency domain

Identity (δ)

Unsharp (G)

2δ–G
Unsharp Masking Example
 Application: Shading Correction

background estimated
with Gaussian filter

Subtract the background

or divide by the
background, depending
on the imaging model.

background estimated
with max-min filter
 Normalised Convolution

Input data, with We also know which

missing samples samples are missing

Simply filtering the input image with the

missing samples produces a bad output
 Normalised Convolution

{ f  x m x  }⊗h x 
m x ⊗h  x 

{f  x  m x  }⊗h x  m x ⊗h  x 

By normalising with the mask, the filter

“skips” missing input samples
 First Order Derivatives
∂ f  x =lim f  x −f  x 
∂x 0 
ℱ { ∂ f  x  =i  ℱ { f  x  }
∂x }
● In a discrete grid, the smallest δ is 1
● Convolve with [ 1 -1 ] filter
– Asymmetric
● Convolve with [ 1 0 -1 ] / 2 filter
– Larger δ = worse approximation to derivative
 Gaussian Derivatives
● Both filters have problems:
– High response to noise
– Poor approximation of gradient vector in n-D images
● Solution: use Gaussian derivatives
∂ { f  x ⊗G x  }=f  x ⊗ ∂ G x 
∂x ∂x
– Reduced response to noise
– Computes exact derivative of smoothed function
(meaning gradient vector has correct direction)
– “Band-limited”, so discretisable
– Separable
 Other Derivatives
● In the first course you learned about some other
derivative operators:
– Prewitt

[ ] []
1 0 −1 1
1 0 −1 /6 = [ 1 0 −1 ] /2 ⊗ 1 /3
1 0 −1 1
– Sobel

[ ] []
1 0 −1 1
2 0 −2 /8 = [ 1 0 −1 ] /2 ⊗ 2 / 4
1 0 −1 1
– Etc.
● You should understand now why Gaussian is better!
 Uses of The Gradient
● The gradient is a vector perpendicular to the edge
● Gradient magnitude is a measure for edge strength

∣∇ f ∣ =
 ∂ f 2 ∂ f 2...
∂x   
∂y
● Gradient direction is a measure for local orientation


∢  ∇ f  = atan2 ∂ f , ∂ f
∂y ∂x 
● Of course, you need a rotation invariant filter to
accurately measure orientation. The Gaussian gradient
operator is rotation invariant
 Gradient Direction

Difference between
computed and true
gradient direction,
for various gradient
filters.
Input Gauss
radian
5 10-5

-5 10-5
Prewitt Sobel
Gradient Direction
Isotropy of Gradient Magnitude
impulse response
(in Fourier Domain)
using Gaussian gradients

impulse response
(in Fourier Domain)
using Sobel derivatives
 Second Order Derivatives
∂2 f  x =lim f  x −2 f  x f  x −
∂ x2 0 2

{ }
2
ℱ ∂ f  x  =− 2
ℱ {f  x  }
2
∂x

● Finite difference approximation

– Convolve with [ 1 -2 1 ]
– Note that [ 1 -2 1 ] = [ 1 -1 ] ⊗ [ 1 -1 ]
● Gaussian approximation
– Convolve with second derivative of Gaussian
 Laplace Operator
● The Laplace operator is everywhere in physics
e.g. remember heat equation:
∂f 2
– =D ∇ f
∂t

 
2 2
2
∇ f =∇⋅∇ f = ∂ ∂
 2 ... f
2
∂x ∂y
● The Laplace operator is:
– isotropic
– generalized 2nd derivative
● It detects lines, and responds strongly to edges
● It does not measure edge magnitude, as reported in
book (pg. 133)
 Laplace Operator

 
2 2
2
∇ f =∇⋅∇ f = ∂ ∂
 2 ... f
2
∂x ∂y

[ ]
1 1 1
● Finite difference approximation: 1 −8 1
1 1 1
● Gaussian approximation:
– Less sensitive to noise, more isotropic
– But: not separable!
● Another approximation: difference of Gaussians (DoG)
– Advantage: separable
 Laplace Operator
H(0) = 0

Laplace of
Gaussian

finite difference Laplace

difference of
Gaussians
 Unsharp Masking Revisited

spatial domain frequency domain

2 −G =1 + =

−∇ 2 G =1 + =

−∇ 2  + =
 Kuwahara / Nagao Filter
● Mean in neighbourhood
● The neighbourhood shifts for each pixel
– Neighbourhood with minimum variance chosen
● Does not average across edges

In book: “averaging using rotating mask”

 Adaptive Filters
● Generalization of Kuwahara / Nagao
● We can turn or grow and shrink our neighbourhood
● Many, many ways of directing the filter
Gaussian (3x0.8) rotated to
align to local contours
 Sigma Filter
● Computes mean in neighbourhood, like uniform filter
● But: excludes pixels that differ more than s from the
central value
7-pixel circular uniform
sigma filter, with s=20

7-pixel circular
uniform filter
 Generalizing the Sigma Filter

position position
intensity

intensity
 Bilateral Filter
● Kernel is Gaussian in distance, like linear Gaussian
● Kernel is also Gaussian in intensity difference
– Edges attenuate kernel significantly
hx  x  = G  x − x 0  G  f  x −f  x 0  
0 x f
 Anisotropic Diffusion
∂f
= ∇⋅ D f , x  ∇ f  , f  x , t 
∂t
● Linear diffusion = heat equation = Gaussian filter
∂f 2
=D ∇ f
∂t
● Choose diffusion coefficient
to be low at edges:
Df , x =g ∣∇ f∣
2
− u/ K 
g u=e
1
g u= 2
1  u / K 
 Summary of Today’s Lecture
● Gaussian filters for:
– Smoothing
– Derivatives
– Laplace operator
● Non-linear filters for edge-preserving smoothing
● Smoothing filters for:
– Noise reduction
– Image simplification
– Shading correction
– Edge sharpening
● First order derivatives used for edge detection