1/41

Computer Assisted Image Analysis

Lecture 2 – Point Processing

Anders Brun (anders@cb.uu.se)

Centre for Image Analysis
Swedish University of Agricultural Sciences
Uppsala University
Reading Instructions 2/41
Chapters for this lecture

• Chapter 2.6 2.6.4 and 3.1 3.3 in

Gonzales-Woods.
Previous Lecture 3/41
Digitization

Digital images
• Images denoted by functions, e.g. f (x, y) or g(x, y)
• Sampling in space, e.g. (x, y) 2 I and ||I|| = N , where I is a
discrete set of pixel positions.
• Quantization in amplitude (intensity),
f (x, y) 2 {0, 1, . . . (L 1)}
Image Processing 4/41
In image processing, the operator T transforms the input image
into an output image, g(x, y) = T (f (x, y)).

Typical examples of image processing

• Image restoration: reduce noise and imaging artefacts
• Image enhancement: enhance edges, lines and subtle features
for easier visual inspection
• Feature extraction, as input to subsequent image analysis

Image processing does NOT increase image information!

Image Processing 5/41

• Spatial domain (lectures 2, and 3)

• Brightness transforms, works per pixel ! point processing,
• Spatial filters, local transforms, works on small neighborhoods,
• Geometric transforms, interpolation,
• Frequency domain (lecture 4 and 5).
• The Fast Fourier Transform (FFT)
• Lowpass-, bandpass- and highpass filters,
• The Wiener filter
Image Processing 6/41
In spatial domain processing, the operator T is applied to each
position (x, y) in the input image f , defined over some
neighborhood of (x, y), yielding a value s = g(x, y) as output.

g(x, y) = T [f (x, y)]

In point processing, the operator neighborhood is the pixel itself.

s = T (r), where r = f (x, y), s = g(x, y).

In spatial filtering, larger neighborhoods are used. They are

referred to as masks, filters, kernel windows or templates.

Point processing Spatial filters

Gray Level Transform 7/41

Pixel-wise transform
• Change the gray level for each individual pixel.
• Compare to television: Brightness and contrast
• brightness: addition
• contrast: multiplication

> 45 ! increased contrast

< 45 ! decreased contrast
up ! increased brightness
down ! decreased brightness
Image Histograms 8/41
A gray level histogram shows how many pixels there are at each
intensity level. The bars either sum up to the total number of
pixels, or to 1 (normalized) in a histogram.
Image Processing 9/41
Brightness

Subtract. Add.
Image Processing 10/41
Contrast

Multiply
Gray Level Transformations 11/41
Some basic gray level transformation functions used for image
enhancement.
Gray Level Transformations 12/41
Gray Level Transformations 13/41
Negative or positive

• Original digital mammogram (left).

• Image negative to enhance white or gray details embedded in
dark regions (right).
Gray Level Transformations 14/41
Log transformations

Visualize patterns in the dark region of an image

• Fourier spectrum (left).
• Result of applying the log transform (right).
Histogram Equalization 15/41

Idea: Create an image with evenly distributed gray levels, for

visual contrast enhancement
• The normalized gray level histogram gives the probability for a
pixel to have a certain gray level, pk = nk /N
• Transform the image using the cumulative density function,
Pk R k
cdf(k) = i=0 pi (or = i=0 p(i)di in the continuous case)
Histogram Equalization 16/41

• Continuous formula, where p(i) is the probability measure of

the i grayvalue in the image if s, r 2 [0, L]
Rr
s = T (r) = L i=0 p(i) = L cdf(r)

• Discrete formula, where ni is the number of pixels with

intensity i and N is the total number of pixels and sk and
rk 2 {0, 1, . . . , (L 1)} :
Pk
j=0 nj
sk = T (rk ) = (L 1) N

• Both formulas try to stretch rmin to 0 and rmax to either L

or (L 1) (But do they succeed?)
• The histogram for the output image is uniform (theoretically
in the continuous case), why not in our digital images?
Histogram Equalization 17/41
Why does this work?
• Let pr be the normalized histogram (probability function) for
the input image f (x, y)
Rr
• Transform f (x, y) using s = T (r) = L 0 pr (w)dw
⇥R r ⇤
ds
• Leibniz´s Rule dr = dTdr(r) = L dr
d
0 pr (w)dw = Lpr (r).
• Then from probability theory we have a formula for the
probability density function (histogram) of the transformed
variable (image), ps

dr
ps = pr (r)| |
ds
1
= pr (r)
Lpr (r)
= 1/L
Histogram Equalization 18/41

Original image.

Result of histogram equalization.

Histogram Equalization Example 19/41

Intensity 0 1 2 3 4 5 6 7
Number of pixels 10 20 12 8 0 0 0 0

p(0) = 10/50 = 0.2

p(1) = 20/50 = 0.4
p(2) = 12/50 = 0.24
p(3) = 8/50 = 0.16
p(r) = 0/50 = 0, r = 4, 5, 6, 7
Histogram Equalization Example (cont.) 20/41
Pk
j=0 nj
Pk
sk = T (rk ) = (L 1) N = (L 1) j=0 p(j)

T (0) = 7 ⇤ (p(0)) ⇡ 1
T (1) = 7 ⇤ (p(0) + p(1)) ⇡ 4
T (2) = 7 ⇤ (p(0) + p(1) + p(2)) ⇡ 6
T (3) = 7 ⇤ (p(0) + p(1) + p(2) + p(3)) = 7
T (r) = 7, r = 4, 5, 6, 7
Intensity 0 1 2 3 4 5 6 7
Number of pixels 0 10 0 0 20 0 12 8
Histogram Equalization 21/41
Example: Original image f (x, y)
Histogram Equalization 22/41
Example: Histogram
Histogram Equalization 23/41
Example: Normalized histogram
Histogram Equalization 24/41
Example: Cumulative histogram
Histogram Equalization 25/41
Example: Normalized cumulative histogram
Histogram Equalization 26/41
Example: Histogram equalization transform
Histogram Equalization 27/41
Example: Histogram equalization
Histogram Equalization 28/41
Example: Equalized histograms
Histogram Equalization 29/41

Transformations for image 1 4.

Note that the transform for figure
4 (dashed line) is close to the
neutral transform (dotted line).
Histogram Equalization 30/41
Not always “optimal” for visual quality

Original. Equalized. Manual choice.

Histogram Equalization 31/41

Transformations for image 1 4.

Note that the transform for figure
4 (dashed line) is close to the
neutral transform (dotted line).
Histogram Matching 32/41

Transform image f (x, y) to match the histogram of image g(x, y)

• If s = T (r) maps f (x, y) to a uniform histogram
• and u = G(t) maps g(x, y) to a uniform histogram
• Then s = G 1 (T (r)) maps f (x, y) to have a histogram
similar to g(x, y)
Arithmetic/Logical Operations 33/41

• Information from two di↵erent images with the same size can
be combined by adding, subtracting, multiplying or comparing
the pixel values, pixel by pixel. Rounding to fit [0, L 1].
• For enhancement, segmentation, change detection.
Arithmetic/Logical Operations 34/41

Image 1. Image 2.
Arithmetic/Logical Operations 35/41
Enhancement by image subtraction

(a) Mask image.

(b) Image (after injection of dye into the bloodstream) with mask
subtracted out.
Arithmetic/Logical Operations 36/41
Images as Vectors

• We may regard images as vectors, i.e. a ordered set of scalars

• All pointwise arithmetic works for both images and vectors
• In fact, sometimes even the geometrical interpretation of
vectors is natural for images, e.g. orthogonality
• However, by subtracting two images we may end up with
negative pixel values. What is that?! Negative coefficients are
natural for vectors, but not for e.g. light intensities or
densities.
• Solution: Let´s not care too much about that ... Deal with
negative, very large and floating-point values by rounding to
the closest integer in [0, L 1] before saving the resulting
image.
Arithmetic/Logical Operations 37/41
Reduction of noise by averaging

Noise can be reduced by

observing the same scene
over a long period of time,
and averaging the images.
Top: original and a noisy
image. Then noisy images
averaged 8, 16, 64 and
128 times.
Arithmetic/Logical Operations 38/41
Reduction of noise by averaging

• Averaging yields a normally distributed resulting image

(Central Limit Theorem)
• Averaging approaches the expected value of the noisy images
(Law of large numbers)
• The standard deviation, after averaging M noisy uncorrelated
images with standard deviation , is p1 .
M
• (However, this only works for noise or image artefacts with
expectation value zero, i.e. it is fine for Gaussian distributed
noise but not for Poisson distributed noise.)
Linear vs Non-Linear Operations 39/41

• An operator H is linear if
• H[ai f i(x, y) + aj fj (x, y)] = ai H[fi (x, y)] + aj H[fj (x, y)]]
• Linear operators have properties that make them useful in
image analysis, in particular for image filtering
• The class of non-linear operators is huge
• Example: sin is non-linear (“The Freshman´s dream”)
• sin(fi (x, y)) + sin(fj (x, y)) 6= sin(fi (x, y) + fj (x, y))
Review Questions 40/41

• Problems 2.22, 2.18, 2.9, 3.1, 3.5 and 3.6 in

Gonzales-Woods.
• Download answers from
http://www.imageprocessingplace.com.
Until Next Lecture 41/41

• Read, read, read

• Experiment in Matlab
• Do the review questions