Intro1

1
Introduction to Biomedical EngineeringDigital
Image Processing

• Topics
• Fundamentals of DIP Scenes Digital Images
  Image Resolution Criteria
• Mathematical Preliminaries
• Visual Perception
• Image Digitization
• Image Enhancement
• Medical Imaging Facilities and Methodologies

2
Fundamentals of DIP

• Ultimate Goal To help better
• understand, interpret the
• content of the image

3
Scenes and Images

• Optical image of the scene produced by a lens
• a scene -gt an image (via a lens)
• an image illumination pattern (recorded by
  sensors)
• sensor response varies with color

light rays
Object
lens
image
u
v
1/u 1/v 1/f (f focal length)
4
Digital Images and Image Resolution Criteria

• An Image ??A Picture
• Any 2D function that bears information
• Denoted f(x,y) or f(x), where f() is the
  brightness or gray values in BW image, or RGB
  values in color image
• f(x,y) y
• 
  
• 
  x

5

• Properties of Brightness
• Real
• Non-negative
• Bounded (due to finite field of view)
• Image Examples
• x-ray absorption, proton density distribution of
  MR images, radar images, temperature profile,
  luminance of a scene, drawings

A tumor at choroid plexus region (2124s4.11)
6

• Digital Image Processing
• To process a picture using digital techniques,
  i.e., computers
• A typical image processing system
• Digitizer To sample and quantize an image into
  digital form of discrete picture elements (pixel
  , pel)

Image Maker

Scene
Image
Digitizer
Users
Display
CPUs
Storage
7

• Sampling and quantization
• Light dots are not integer samples
• round-off or truncation is needed gt error
• larger dynamic range may alleviate this problem

dynamic range
gray levels
quantization domain

samples
sampling domain
8

• Image Resolution Criteria minimal separation
  that intensity variation can still be discerned
• Resolution ...
• as d gets smaller and smaller, two peaks merge
  into one, i.e., d metric resolution is lost

ideal case
real case
d
9

• Human eye under visible light and viewing
  distance can resolve separation resolution is
  tied to grain size of the silver halides
  crystals.
• Examples
• A film in electron microscopy is 520 ?m
• 300 dpi (dot per inch) corresponds to 85 ?m
• 10 ?m resolution of a Xenopus laevis embryo
  (microscopic MRI)

10

• Typical human visual system can resolve
  approximately 0.003 times of the viewing distance
  to the object

11
Mathematical Preliminaries

• f(x,y) y
• Digital Picture Function
• analytically well-behaved, i.e., bounded,
  integrable, have Fourier transform pairs, etc.

x
12

• Operations on pixels
• point by point operations, e.g., difference image
• local operations, e.g., edge detection
• geometric operation, e.g., image translation
• Problems with pixels on the border
• Assume equal to zero
• Define a sub-picture excluding those
  borderingosed, repetition, (problem oriented)

13

• Arithmetic Operations
• Addition p q, Subtraction p - q
• Multiplication p q, Division p q

q (2124s4.3)
p (2124s4.2)
(p-q) 0.5 128
p - q
14

• Logic Operations
• AND pANDq (also, p q)
• OR pORq (also, p q)
• COMPLEMENT NOTq (also, )
• EXCLUSIVE OR pXORq
• Example of COMPLEMENT
• before after
• Examples of logic operations on binary images
  (from RCG fig. 2.14)

15

• Logic Operations
• p q
• pANDq
• pORq
• pXORq

16
Visual Perception

• Reasons to studying human vision
• 1. Interpretation (detection, recognition,
  classification, ...)
• 2. Information storage and display
• 3. Feature enhancement, correction, image
  processing, algorithm development and design ...

17

• Gestalt laws of organization
• A visual field is usually seen as consisting of a
  small number of regions (objects on a background)
  that abide the following rules
• law of similarity
• law of proximity closely clustered entities tend
  to group together.
• law of good continuation smooth curve effect
• law of closure closed figures tend to be seen as
  a unit

18

• law of simplicity (least information). Example
  Kopfermann cubes, the left one is easily seen as
  3D cube because its simpler and more familiar to
  us
• law of common fate (time or spatial varying
  images). Example uniform motion of a collection
  of objects is easily seen as a single unit

19

• Webers law
• Human visual perception is sensitive to luminance
  contrast rather than the absolute luminance
  values
• f0 luminance of the object
• fs luminance of its surrounds
• f0 - fs ?f just noticeable difference in
  luminance

20

• Simultaneous apparent brightness depends
  strongly on the local background intensity

21

• Mach bands The response of the visual system to
  abrupt changes in luminance (edges or contours)
  display overshoots (mach bands), which have the
  effect of enhancing or deblurring the edges

22

Ebbinghaus illusion
23
Zollner illusion
24
The Benussi ring
25
The Benary cross (although the lower triangle has
more black in its vicinity, it looks
darker because it is seen as outside the black
cross)
26
Filling in from edges having gaps
27
Image Digitization

• Digitization Sampling and Quantization
• f(x,y) an image f(i,j) an image element, pixel,
  or pel
• Sampling partitioning the image as an ordered
  pairs of elements (a,b), with a and b being
  integers. a 0 .. N-1, b 0 .. M-1
• Quantization Assigning a real value to the
  sampled image elements (pixels). In black and
  white images, this value is called the gray level
• For an NxM digital image with 2L gray levels, it
  requires NxMxL bits to represent it

28

• Example of a digitized image

• A tumor at choroid plexus region (2124s4.11)

29

• To better represent the picture, M, N, and L
  should be large. Nothing is gained, however, by
  increasing M, N, and L beyond the resolution
  capabilities of the receiver.
• Question How to choose M, N and L for a fixed
  data size (bytes)?

30

• Image Sampling - 1D case
• One dimensional sampling function
• Let f(t) denote the 1D signal and T be the
  sampling period

t
0
T 2T ... nT ...
-T
0
31

• Two dimensional Sampling
• Ideal sampling function

y
x
32

• Sampling and Reconstruction from sampled data

f(t)
F(w)
-fc
fc
fs(t)
Fs(w)
-fc
fc
-1/T
1/T
f(t)
F(w)
-fc
fc
-1/T
1/T
33

• Two problems associated with the reconstruction
  of the original signal from its samples
• 1. If f(t) is not a bandlimited signal, i.e.,
• F(w) ? 0, -8 w 8 or fc 8
• 2. If 1/T is not sufficiently distanced.
• Both cases result overlapped spectrum, a
  phenomenon called aliasing.
• For bandlimited signal, the sufficient condition
  to reconstruct the original signal back from its
  sampled signal is
• 1/T 2 fc
• where fc is the bandwidth of the original signal.
  The lower bound is called the Nyquist rates or
  Nyquist frequency.

34
Image Enhancement

• Purposes To make an image better appealing and
  easier to deal with than the original image
• Three categories
• 1. Spatial domain methods operate on the images
  itself, examples as
• Point processing, e.g., image averaging logic
  operation contrast stretching ...
• Mask processing, e.g., filtering or mask
  operation, (blurring, median

35

• 2. Frequency domain methods work on the Fourier
  transformed output of the image, examples from
  the convolution theory
• g(x,y) f(x,y) ??h(x,y)
• gt G(u,v) F(u,v) H(u,v)
• gt certain properties of F(u,v) can be
• emphasized into G(u,v)
• gt spatial domain g(x,y) F-1G(u,v)
• 3. Combination of the above two categories

36

• Spatial Domain Methods
• Point processing enhancement
• Image intensity transformation
• Negative image

L-1
T(r)
s
0
L-1
r
37

• Contrast stretching to increase the dynamic
  range of the gray levels in the image

38

• Dynamic range compression
• linear scaling
• input range R, output range L
• output gray level s (r-I0)L/R
• I0 lower bound of the input
• logarithm scaling s c log(1 r)
• useful when the input range is very large, e.g.,
  106, and the brightest parts are dominating (from
  RC Gonzalez)

39

• Histogram Processing
• A histogram is a plot of the number of gray
  level, rk, its occurrencesk, 0kL-1, versus the
  range of gray levels normalized to the total
  number of pixels, n
• Histogram of a dark image
• Histogram Equalization to equalize the histogram
  according to its probability density function

40

• Image subtraction
• g(x,y) f(x,y) - h(x,y)

Image subtraction enhancement (a) mask image (b)
image with mask subtracted out (after injection
of dye into the bloodstream) (from RC Gonzalez)
41

• Image averaging consider a noisy image

M 1
M 2
M 16
M 8
M 32
M 128
(from RC Gonzalez)
42

• Spatial Filtering (Mask processing)
• lowpass filtering eliminating high frequency
  components gt image blurring (from RC Gonzalez)

43

• highpass filtering sharpening edges or fine
  details in an image gt deblurring (from RC
  Gonzalez)

44

• Derivative filters
• averaging ? integration gt blurring
• difference ? differentiation gt sharpening
• Gradient operator

In discrete case
45

• Consider the digital image

At point z5
?f(z5-z4)2(z5-z2) 21/2
z5-z4z5-z2
Laplacian operator
Digital Laplacian has the effect of increasing
the ramp steepness, and of increasing the
contrast at the edges
46

• High-emphasis filtering
• differentiation enhances high spatial frequencies
• integration weakens high frequencies
• the effect of subtracting a Laplacian from an
  image itself

47
Original (2124s4.1)
Laplace(3x3)
Laplace-(3x3)
High Emphasis
48

• Various image processing effects

Original
Smooth
1 1 1 1 1 1 1 1 1
Shadow
Sharpen
-2 -1 0 -1 1 1 0 1 2
-1 -1 -1 -1 -12 -1 -1 -1 -1
49
Original
Trace edge
1 1 1 1 1 1 1 1 1
1 1 1 0 0 0 -1 -1 -1
-1 0 1 -1 0 1 -1 0 1
Dither (Floyd-Steinberg)
Reduce Noise (median filter)
50
Grad-NW (3x3)
Grad-N (3x3)
Laplace (5x5)
Laplace (9x9)
51
Grad-W (7x7)
High-Emphasis (3x3)
Hat (5x5)
Hat (13x13)
52
Mean (5x5)
Mean (15x15)
Gauss (15x15)
Gauss (5x5)
53
Original
Subtract Background
Enhance Contrast
Equalize
54

• Frequency domain processing

aOriginal (2124s4.1)
bFFTa
cHigh_passb
dIFFTc
55

• Two pictures are worth more than ten thousand
  words