Vectors and more on masks

1
Vectors and more on masks

• Vector space theory applies directly to several
  image processing/representation problems

2
Image as a sum of basic images
What if every persons portrait photo could be
expressed as a sum of 20 special images? ? We
would only need 20 numbers to model any photo ?
sparse rep on our Smart card.
3
Efaces
100 x 100 images of faces are approximated by a
subspace of only 4 100 x 100 images, the mean
image plus a linear combination of the 3 most
important eigenimages
4
The image as an expansion
5
Different bases, different properties revealed
6
Fundamental expansion
7
Basis gives structural parts
8
Vector space review, part 1
9
Vector space review, Part 2
2
10
A space of images in a vector space

• M x N image of real intensity values has
  dimension D M x N
• Can concatenate all M rows to interpret an image
  as a D dimensional 1D vector
• The vector space properties apply
• The 2D structure of the image is NOT lost

11
Orthonormal basis vectors help
12
Represent S 10, 15, 20
13
Projection of vector U onto V
14
Normalized dot product
Can now think about the angle between two
signals, two faces, two text documents,
15
Every 2x2 neighborhood has some constant, some
edge, and some line component
Confirm that basis vectors are orthonormal
16
Roberts basis cont.
If a neighborhood N has large dot product with a
basis vector (image), then N is similar to that
basis image.
17
Standard 3x3 image basis
Structureless and relatively useless!
18
Frie-Chen basis
Confirm that bases vectors are orthonormal
19
Structure from Frie-Chen expansion
Expand N using Frie-Chen basis
20
Sinusoids provide a good basis
21
Sinusoids also model well in images
22
Operations using the Fourier basis
23
A few properties of 1D sinusoids
They are orthogonal
Are they orthonormal?
24
F(x,y) as a sum of sinusoids
25
Spatial direction and frequency in 2D
26
Continuous 2D Fourier Transform
To compute F(u,v) we do a dot product of our
image f(x,y) with a specific sinusoid with
frequencies u and v
27
Power spectrum from FT
28
Examples from images
Done with HIPS in 1997
29
Descriptions of former spectra
30
Discrete Fourier Transform
Do N x N dot products and determine where the
energy is. High energy in parameters u and v
means original image has similarity to those
sinusoids.
31
Bandpass filtering
32
Convolution of two functions in the spatial
domain is equivalent to pointwise multiplication
in the frequency domain
33
LOG or DOG filter

• Laplacian of Gaussian
• Approx
• Difference of Gaussians

34
LOG filter properties
35
Mathematical model
36
1D model rotate to create 2D model
37
1D Gaussian and 1st derivative
38
2nd derivative then all 3 curves
39
Laplacian of Gaussian as 3x3
40
G(x,y) Mexican hat filter
41
Convolving LOG with region boundary creates a
zero-crossing
Mask h(x,y)
Input f(x,y)
Output f(x,y) h(x,y)
42
(No Transcript)
43
LOG relates to animal vision
44
1D EX.
Artificial Neural Network (ANN) for computing
g(x) f(x) h(x) level 1 cells feed 3 level 2
cells level 2 cells integrate 3 level 1 input
cells using weights -1,2,-1
45
Experience the Mach band effect
46
Simple model of a neuron
47
Output conditioning threshold versus smoother
output signal
48
3D situation in the eye
Neuron c has input to neuron A but - input to
neuron B. Neuron d has input to neuron B but
input to neuron A. Neuron b gives no input to
neuron B it is not in the receptive field of B.
49
Receptive fields
50
Experiments with cats/monkeys

• Stabilize/drug animal to stare
• Place delicate probe in visual network
• Move step edge across FOV
• Probe shows response function when the edge
  images to receptive field
• Slightly moving the probe produces similar signal
  when edge is nearby

51
Canny edge detector uses LOG filter
52
Summary of LOG filter

• Convenient filter shape
• Boundaries detected as 0-crossings
• Psychophysical evidence that animal visual
  systems might work this way (your testimony)
• Physiological evidence that real NNs work as the
  ANNs