The Course

 Image representation  Books

 Image statistics Computer Vision –
Adrian Lowe
 Histograms (frequency)
Digital Image Processing –
 Entropy (information) Gonzalez, Woods
 Filters (low, high, edge, smooth) Image Processing, Analysis
and Machine Vision – Milan
Sonka, Roger Boyle
Digital Image Processing

 Human vision - perceive and understand world

 Computer vision, Image Understanding / Interpretation,
Image processing.
3D world -> sensors (TV cameras) -> 2D images
Dimension reduction -> loss of information
 low level image processing
transform of one image to another
 high level image understanding
knowledge based - imitate human cognition
make decisions according to information in image
 Introduction to Digital Image
Processing

Classification / decision

 Acquisition,
preprocessing
HIGH
Algorithm Amount of  no intelligence
Complexity Data
 Extraction, edge
Increases MEDIUM Decreases
joining
LOW  Recognition,
interpretation
 intelligent

Raw data
Low level digital image
processing
 Low level computer vision ~ digital image processing
 Image Acquisition
 image captured by a sensor (TV camera) and digitized
 Preprocessing
 suppresses noise (image pre-processing)
 enhances some object features - relevant to understanding the image
 edge extraction, smoothing, thresholding etc.
 Image segmentation
 separate objects from the image background
 colour segmentation, region growing, edge linking etc
 Object description and classification
 after segmentation
Signals and Functions
 What is an image
 Signal = function (variable with physical meaning)
one-dimensional (e.g. dependent on time)
two-dimensional (e.g. images dependent on two co-ordinates in a
plane)
three-dimensional (e.g. describing an object in space)
higher-dimensional
 Scalar functions
sufficient to describe a monochromatic image - intensity images
 Vector functions
represent color images - three component colors
Image Functions
 Image - continuous function of a number of variables
 Co-ordinates x, y in a spatial plane
 for image sequences - variable (time) t
 Image function value = brightness at image points
 other physical quantities
 temperature, pressure distribution, distance from the observer
 Image on the human eye retina / TV camera sensor - intrinsically 2D
 2D image using brightness points = intensity image
 Mapping 3D real world -> 2D image
 2D intensity image = perspective projection of the 3D scene
 information lost - transformation is not one-to-one
 geometric problem - information recovery
 understanding brightness info
Image Acquisition &
Manipulation
 Analogue camera
 frame grabber
 video capture card
 Digital camera / video recorder
 Capture rate  30 frames / second
 HVS persistence of vision
 Computer, digitised image, software (usually c)
 f(x,y)  #define M 128
#define N 128
unsigned char f[N][M]
 2D array of size N*M
 Each element contains an intensity value
Image definition

 Image definition:
A 2D function obtained by sensing a scene N
F(x,y), F(x1,x2), F(x)
F - intensity, grey level f(o,o)
x,y - spatial co-ordinates
M
 No. of grey levels, L = 2B
 B = no. of bits f(N-1,M-1)

B L Description
1 2 Binary Image (black and white)
6 54 64 levels, limit of human visual system
8 256 Typical grey level resolution
Brightness and 2D images
 Brightness dependent several factors
 object surface reflectance properties
surface material, microstructure and marking
 illumination properties
 object surface orientation with respect to a viewer and light source
 Some Scientific / technical disciplines work with 2D images directly
 image of flat specimen viewed by a microscope with transparent
illumination
 character drawn on a sheet of paper
 image of a fingerprint
Monochromatic images
 Image processing - static images - time t is constant
 Monochromatic static image - continuous image
function f(x,y)
arguments - two co-ordinates (x,y)
 Digital image functions - represented by matrices
co-ordinates = integer numbers
Cartesian (horizontal x axis, vertical y axis)
OR (row, column) matrices
 Monochromatic image function range
lowest value - black
highest value - white
 Limited brightness values = gray levels
Chromatic images

Colour
Represented by vector not scalar
Red, Green, Blue (RGB)
Hue, Saturation, Value (HSV)
luminance, chrominance (Yuv , Luv)
S=0
Green Hue degrees:
Red Red, 0 deg
Green 120 deg
Green Blue 240 deg
V=0
Use of colour space
Image quality

Quality of digital image proportional to:

spatial resolution
proximity of image samples in image plane
spectral resolution
bandwidth of light frequencies captured by sensor
radiometric resolution
number of distinguishable gray levels
time resolution
interval between time samples at which images captured
Image summary
N
 F(xi,yj)
 i = 0 --> N-1 f(o,o)
 j = 0 --> M-1
 N*M = spatial resolution, size of image M
 L = intensity levels, grey levels
f(N-1,M-1)
 B = no. of bits
Digital Image Storage

Stored in two parts

header
width, height … cookie.
• Cookie is an indicator of what type of image file
data
uncompressed, compressed, ascii, binary.
File types
JPEG, BMP, PPM.
PPM, Portable Pixel Map

Cookie
Px
Where x is:
1 - (ascii) binary image (black & white, 0 & 1)
2 - (ascii) grey-scale image (monochromic)
3 - (ascii) colour (RGB)
4 - (binary) binary image
5 - (binary) grey-scale image (monochromatic)
6 - (binary) colour (RGB)
PPM example

 PPM colour file RGB

P3
# feep.ppm
44
15
0 0 0 0 0 0 0 0 0 15 0 15
0 0 0 0 15 7 0 0 0 0 0 0
0 0 0 0 0 0 0 15 7 0 0 0
15 0 15 0 0 0 0 0 0 0 0 0
Image statistics
M 1 N 1

 f ( x, y)
y 0 x 0
 MEAN  =
N *M
M 1 N 1

 ( f ( x , y )   ) 2

 VARIANCE 2 = y 0 x 0

N *M

 STANDARDEVIATION  = var iance

Histograms, h(l)
 Counts the number of occurrences of each grey level in
an image
 l = 0,1,2,… L-1
 l = grey level, intensity level
 L = maximum grey level, typically 256

MAX
Area under histogram


 h (l ) 
l 0
Total number of pixels N*M
unimodal, bimodal, multi-modal, dark, light, low contrast, high
contrast
Probability Density
Functions, p(l)

 Limits 0 < p(l) < 1

 p(l) = h(l) / n
 n = N*M (total number of pixels)
MAX

 p(l ) 1
l 0
Histogram Equalisation, E(l)

Increases dynamic range of an image

Enhances contrast of image to cover all possible
grey levels
Ideal histogram = flat
 same no. of pixels at each grey level
N *M
 Ideal no. of pixels at each grey level = i 
L
Histogram equalisation

Typical histogram Ideal histogram

E(l) Algorithm
 Allocate pixel with lowest grey level in old image to 0 in new image
 If new grey level 0 has less than ideal no. of pixels, allocate pixels
at next lowest grey level in old image also to grey level 0 in new
image
 When grey level 0 in new image has > ideal no. of pixels move up
to next grey level and use same algorithm
 Start with any unallocated pixels that have the lowest grey level in
the old image
 If earlier allocation of pixels already gives grey level 0 in new image
TWICE its fair share of pixels, it means it has also used up its
quota for grey level 1 in new image
 Therefore, ignore new grey level one and start at grey level 2 …..
Simplified Formula

L
E (l )  max( o, round (( ) * t (l ))  1)
N *M
 E(l)  equalised function
 max  maximum dynamic range
 round  round to the nearest integer (up or down)
L  no. of grey levels
 N*M  size of image
 t(l)  accumulated frequencies
Histogram equalisation
examples

Typical histogram After histogram equalisation

Histogram Equalisation e.g.
L
E (l )  max( o, round (( ) * t (l ))  1)
N *M
10 Before HE After HE
9
8
7
6
5
4 Ideal=3
3
2
1
0
1 2 3 4 5 6 7 8 9 10
 2
1
0
g h(g) t(g) e(g) New hist
1 1 1 1 0
2 9 10 3 0
3 8 18 6 9
4 6 24 8 0
5 1 25 8 0
6 1 26 9 8
7 1 27 9 0
8 1 28 9 7
9 2 30 10 3
10 0 30 10 2
Noise in images
 Images often degraded by random noise
 image capture, transmission, processing
 dependent or independent of image content
 White noise - constant power spectrum
 intensity does not decrease with increasing frequency
 very crude approximation of image noise
 Gaussian noise
 good approximation of practical noise
 Gaussian curve = probability density of random variable
 1D Gaussian noise - µ is the mean
  is the standard deviation
Gaussian noise e.g.

50% Gaussian noise

Types of noise

 Image transmission
noise usually independent image signal
 additive, noise v and image signal g are independent

 multiplicative, noise is a function of signal magnitude

 impulse noise (saturated = salt and pepper noise)

Data Information
 Different quantities of data used to represent same
information
people who babble, succinct
 Redundancy
if a representation contains data that is not necessary

Same information Amounts of data

Representation 1 N1
Representation 2 N2
N1
 Compression ratio CR =
 N2
1
1
 Relative data redundancy RD = CR
Types of redundancy
Coding
if grey levels of image are coded in such away that
uses more symbols than is necessary
Inter-pixel
can guess the value of any pixel from its neighbours
Psyco-visual
some information is less important than other info in
normal visual processing
Data compression
when one / all forms of redundancy are reduced / removed
data is the means by which information is conveyed
Coding redundancy
 Can use histograms to construct codes
 Variable length coding reduces bits and gets rid of redundancy
 Less bits to represent level with high probability
 More bits to represent level with low probability
 Takes advantage of probability of events
 Images made of regular shaped objects / predictable shape
 Objects larger than pixel elements
 Therefore certain grey levels are more probable than others
 i.e. histograms are NON-UNIFORM
 Natural binary coding assigns same bits to all grey levels
 Coding redundancy not minimised
Run length coding (RLC)

 Represents strings of symbols in an image matrix

 FAX machines
 records only areas that belong to the object in the image
 area represented as a list of lists
 Image row described by a sublist
 first element = row number
 subsequent terms are co-ordinate pairs
 first element of a pair is the beginning of a run
 second is the end
 can have several sequences in each row
 Also used in multiple brightness images
 in sublist, sequence brightness also recorded
Example of RLC
Inter-pixel redundancy, IPR

 Correlation between pixels is not used in coding

 Correlation due to geometry and structure
 Value of any pixel can be predicted from the value of the neighbours
 Information carried by one pixel is small
 Take 2D visual information
 transformed  NONVISUAL format
 This is called a MAPPING
 A REVERSIBLE MAPPING allows original to be reconstructed after
MAPPING
 Use run-length coding
Psyco-visual redundancy, PVR

 Due to properties of human eye

 Eye does not respond with equal sensitivity to all visual
information (e.g. RGB)
 Certain information has less relative importance
 If eliminated, quality of image is relatively unaffected
 This is because HVS only sensitive to 64 levels

 Use fidelity criteria to assess loss of information

Fidelity Criteria
Info Encoder Channel Decoder Info User
Source Sink

NOISE
 In a noiseless channel, the
encoder is used to remove any
redundancy
PVR removed, image quality is
 2 types of encoding
reduced
 LOSSLESS
2 classes of criteria
 LOSSY
OBJECTIVE fidelity criteria
SUBJECTIVE fidelity criteria
 Design concerns
OBJECTIVE: if loss is expressed
 Compression ratio, CR achieved
as a function of IP / OP
 Quality achieved
 Trade off between CR and quality
Fidelity Criteria

 Input  f(x,y) M 1 N 1

 e( x, y) 2

 compressed output  f(x,y) erms 

y 0 x 0

 error  e(x,y) = f(x,y) -f(x,y) N *M

M 1 N 1

 f ( x, y)
y 0 x 0
2

SNRms  M 1 N 1

 erms = root mean squared error  e( x, y)

y 0 x 0
2

 SNR = signal to noise ratio

 PSNR = peak signal to noise ratio N * M * ( L  1) 2
PSNR  M 1 N 1

 e( x, y)
y 0 x 0
2
Information Theory
How few data are needed to represent an
image without loss of info?
Measuring information
random event, E
probability, p(E)
units of information, I(E)
1
I ( E )  log   log p( E )
p( E )
I(E) = self information of E
amount of info is inversely proportional to the probability
base of log is the unit of info
log2 = binary or bits
e.g. p(E) = ½ => 1 bit of information (black and white)
Infromation channel

Info Encoder Channel Decoder Info User

Source Sink

NOISE

 Connects source and user

physical medium
 Source generates random symbols from a closed set
 Each source symbol has a probability of occurrence
 Source output is a discrete random variable
 Set of source symbols is the source alphabet
Entropy
 Entropy is the uncertainty of the source
 Probability of source emitting a symbol, S = p(S)
 Self information I(S) = -log p(S) L 1
 For many Si , i = 0, 1, 2, … L-1 
H   Pi log 2 ( Pi )
i 0

 Defines the average amount of info obtained by

observing a single source output
 OR average information per source output (bits)
alphabet = 26 letters  4.7 bits/letter
typical grey scale = 256 levels  8 bits/pixel
Filters
 Need templates and convolution  Convolution of Images
 essential for image processing
 Elementary image filters are used
 template is an array of values
enhance certain features
 placed step by step over image
de-enhance others  each element placement of
edge detect template is associated with a
smooth out noise pixel in the image
discover shapes in images  can be centre OR top left of
template
Template Convolution

 Each element is multiplied with its corresponding

grey level pixel in the image
 The sum of the results across the whole template is
regarded as a pixel grey level in the new image
 CONVOLUTION --> shift add and multiply
 Computationally expensive
big templates, big images, big time!
 M*M image, N*N template = M2N2
Convolution
 Let T(x,y) = (n*m) template
 Let I(X,,Y) = (N*M) image
 Convolving T and I gives:
n 1 m 1
T  I ( X , Y )   T (i, j )  I ( X  i, Y  j )
i 0 j 0

 CROSS-CORRELATION not CONVOLUTION

 Real convolution is:
n 1 m 1
T  I ( X , Y )   T (i, j )  I ( X  i, Y  j )
i 0 j 0

 convolution often used to mean cross-correlation

Templates
Template Image Result
1 0 1 1 3 3 4 2 5 7 6 *
0 1 1 1 4 4 3 2 4 7 7 *
2 1 3 3 3 3 2 7 7 *
1 1 1 4 4 * * * * *

 Template is not allowed to shift  Periodic Convolution

off end of image  wrap image around a ball
 Result is therefore smaller than  template shifts off left, use
image right pixels
 2 possibilities
 pixel placed in top left position of
 Aperiodic Convolution
new image  pad result with zeros
 pixel placed in centre of template
(if there is one)
 Result
 top left is easier to program  same size as original
 easier to program
Filters
 Need templates and convolution  Convolution of Images
 essential for image processing
 Elementary image filters are used
 template is an array of values
enhance certain features
 placed step by step over image
de-enhance others  each element placement of
edge detect template is associated with a
smooth out noise pixel in the image
discover shapes in images  can be centre OR top left of
template
Template Convolution

 Each element is multiplied with its corresponding

grey level pixel in the image
 The sum of the results across the whole template is
regarded as a pixel grey level in the new image
 CONVOLUTION --> shift add and multiply
 Computationally expensive
big templates, big images, big time!
 M*M image, N*N template = M2N2
Templates
Template Image Result
1 0 1 1 3 3 4 2 5 7 6 *
0 1 1 1 4 4 3 2 4 7 7 *
2 1 3 3 3 3 2 7 7 *
1 1 1 4 4 * * * * *

 Template is not allowed to shift  Periodic Convolution

off end of image  wrap image around a ball
 Result is therefore smaller than  template shifts off left, use
image right pixels
 2 possibilities
 pixel placed in top left position of
 Aperiodic Convolution
new image  pad result with zeros
 pixel placed in centre of template
(if there is one)
 Result
 top left is easier to program  same size as original
 easier to program
Low pass filters

 Moving average of time series Removes high frequency

components
smoothes
Better filter, weights
 Average (up/down, left/right) centre pixel more
smoothes out sudden changes in
pixel values 1 1 1 1 3 1
removes noise 1 1 1 3 16 3
introduces blurring 1 1 1 1 3 1
 Classical 3x3 template
Example of Low Pass

Original Gaussian, sigma=3.0

High pass filters
 Removes gradual changes  Roberts Operators
between pixels  oldest operator
 enhances sudden changes  easy to compute only 2x2
 i.e. edges neighbourhood
 high sensitivity to noise
 few pixels used to calculate
gradient

0 1 1 0
-1 0 0 -1
High pass filters
 Laplacian Operator
 known as 
2
0 1 0 1 1 1
 template sums to zero 1 -4 1 1 -8 1
 image is constant (no sudden 0 1 0 1 1 1
changes), output is zero
 popular for computing second
derivative
2 -1 2 -1 2 -1
 gives gradient magnitude only
-1 -4 -1 2 -4 2
 usually a 3x3 matrix
2 -1 2 -1 2 -1
 stress centre pixel more
 can respond doubly to some
edges
Cont.
 Prewitt Operator
 similar to Sobel, Kirsch, Robinson 1 1 1
 approximates the first derivative
0 0 0
 gradient is estimated in eight
possible directions -1 -1 -1
 result with greatest magnitude is the
gradient direction 0 1 1
 operators that calculate 1st derivative -1 0 1
of image are known as COMPASS -1 -1 0
OPERATORS
 they determine gradient direction
 1st 3 masks are shown below
-1 0 1
(calculate others by rotation …) -1 0 1
 direction of gradient given by mask -1 0 1
with max response
Cont.

Sobel Robinson
good horizontal / vertical 1 1 1
edge detector 1 -2 1
1 2 1 -1 -1 -1
0 0 0
-1 -2 -1
Kirsch
-1 0 1
-2 0 2 3 3 3
3 0 3
-1 0 1
-5 -5 -5
0 1 2
-1 0 1
-2 -1 0
Example of High Pass

Laplacian Filter - 2nd derivative

More e.g.’s

Horizontal Sobel Vertical Sobel

1st derivative
Morphology

 The science of form and structure

 the science of form, that of the outer form, inner structure,
and development of living organisms and their parts
about changing/counting regions/shapes
 Used to pre- or post-process images
via filtering, thinning and pruning

 Smooth region edges  Count regions (granules)

create line drawing of face number of black regions
 Force shapes onto region  Estimate size of regions
edges area calculations
curve into a square
Morphological Principles
 Easily visulaised on binary image
 Template created with known origin
1 *
1 1
 Template stepped over entire image
similar to correlation
 Dilation
if origin == 1 -> template unioned
resultant image is large than original
 Erosion
only if whole template matches image
origin = 1, result is smaller than original
Dilation

Dilation (Minkowski addition)

fills in valleys between spiky regions
increases geometrical area of object
objects are light (white in binary)
sets background pixels adjacent to object's
contour to object's value
smoothes small negative grey level regions
Dilation e.g.
Erosion

Erosion (Minkowski subtraction)

removes spiky edges
objects are light (white in binary)
decreases geometrical area of object
sets contour pixels of object to background value
smoothes small positive grey level regions
Erosion e.g.
Hough Transform

Intro
edge linking & edge relaxation join curves
require continuous path of edge pixels
HT doesn’t require connected / nearby points
Parametric representation
Finding straight lines
consider, single point (x,y)
infinite number of lines pass through (x,y)
each line = solution to equation
simplest equation:
y = kx + q
HT - parametric
representation

y = kx + q
(x,y) - co-ordinates
k - gradient
q - y intercept
Any stright line is characterised by k & q
use : ‘slope-intercept’ or (k,q) space not (x,y)
space
(k,q) - parameter space
(x,y) - image space
can use (k,q) co-ordinates to represent a line
Parameter space

q = y - kx
a set of values on a line in the (k,q) space ==
point passing through (x,y) in image space
OR
every point in image space (x,y) ==
line in parameter space
HT properties

 Original HT designed to detect straight lines and

curves
 Advantage - robustness of segmentation results
 segmentation not too sensitive to imperfect data or noise
 better than edge linking
 works through occlussion
 Any part of a straight line can be mapped into
parameter space
Accumulators

Each edge pixel (x,y) votes in (k,q) space for

each possible line through it
i.e. all combinations of k & q
This is called the accumulator
If position (k,q) in accumulator has n votes
n feature points lie on that line in image space
Large n in parameter space, more probable
that line exists in image space
Therefore, find max n in accumulator to find
lines
 HT Algorithm

 Find all desired feature points in

image space
i.e. edge detect (low pass filter)
 Take each feature point
increment appropriate values in
parameter space
i.e. all values of (k,q) for give (x,y)
 Find maxima in accumulator array
 Map parameter space back into
image space to view results
Alternative line representation
 ‘slope-intercept’ space has problem
verticle lines k -> infinity
q -> infinity
 Therefore, use (,) space
 = xcos  + y sin 
 = magnitude
drop a perpendicular from origin to the line
 = angle perpendicular makes with x-axis
, space
In (k,q) space
point in image space == line in (k,q) space
In (,) space
point in image space == sinusoid in (,) space
where sinusoids overlap, accumulator = max
maxima still = lines in image space
 Practically, finding maxima in accumulator is non-
trivial
often smooth the accumulator for better results
HT for Circles

Extend HT to other shapes that can be

expressed parametrically
Circle, fixed radius r, centre (a,b)
(x1-a)2 + (x2-b)2 = r2
accumulator array must be 3D
unless circle radius, r is known
re-arrange equation so x1 is subject and x2 is the
variable
for every point on circle edge (x,y) plot range of
(x1,x2) for a given r
Hough circle example
General Hough Properties

Hough is a powerful tool for curve detection

Exponential growth of accumulator with
parameters
Curve parameters limit its use to few
parameters
Prior info of curves can reduce computation
e.g. use a fixed radius
Without using edge direction, all accumulator
cells A(a) have to be incremented
Optimisation HT
With edge direction
edge directions quantised into 8 possible directions
only 1/8 of circle need take part in accumulator
Using edge directions
a & b can be evaluated from

 = edge direction in pixel x

delta  = max anticipated edge direction error
 Also weight contributions to accumulator A(a) by edge
magnitude
General Hough

Find all desired points in image

For each feature point
for each pixel i on target boundary
get relative position of reference point from i
add this offset to position of i
increment that position in accumulator
Find local maxima in accumulator
Map maxima back to image to view
General Hough example

 explicitly list points on shape

 make table for all edge pixles for target
 for each pixel store its position relative to some
reference point on the shape
 ‘if I’m pixel i on the boundary, the reference point is at
ref[i]’