Spring 2006 Artificial Intelligence COSC40503 WEEK 3

1
Spring 2006 Artificial Intelligence
COSC 40503WEEK 3
To understand is to perceive patterns
Sir Isaiah Berlin (1909-1997)

• Antonio Sanchez
• Texas Christian University

2
Introduction

• Artificial Intelligence deals with knowledge and
  learning
• Artificial learning is obtained by
• Traversing knowledge bases (rule based and
  logical programming)
• Artificial selection (genetic and evolutive
  algorithms)
• Adaptive methods (conectionism and feedback)
• Adaptive behavior studies have their roots in
  Pavlovs studies of animal conditioning

3
Two recurring concepts
Conectionism
Feedback

• Perceptron (Rosenblatt, Selfridge)
• Learning Automata (Tsetlin, Narendra,
  Barto)
• Neural Networks
  (Rummelhart, McCleland)
• Collective Learning (Samuel, Michie, Bock)

Cybernetic Loop (Wiener, Rosenblueth,
Ashby) RP Policies (Thathachar,Viswanath
an,Fu) Backpropagation (Skjenoswsky,
Huberman) Algedonic Loop (Beer, Bock,
Sanchez)
Learning with the Environment
Credit Assignment
4
Living with the environment
5
One apliccation image recognition with
classification and segmentation
Images
Deterministic Filters
Earth Blue
Sun Yellow
i gt j STM
Results
6
Theoretical framework

• CLS (Collective Learning Systems)
• Algedonic Compensation (Algedolor, Edos placer)
• Transition Matrix
• (STM ? Accumulative histogram for
  Adaptive Learning)
• Accumulative Histogram Analisis
• Color Texture Filters (color pixels texels)
• Fast Fourier Transform (FFT spectograms)

7
Three Examples of the same aplication

• Classification and segmentation of celestial
  bodies
• Segmentationof skin texture
• Recognition of animal sounds

8
First ExampleClassification and Segmentation of
celestial bodies
9
Further Segmentation
10
A hierarchical approach towards recognition and
segmentation
By using a hierarchy of ALISA layers, it is
possible to recognize different images classes
according to previous training. Segmentation of
the distinct elements is done by a texel and
color analysis using a cumulative histogram of
the various features of the image. The hierarchy
may be organized as follows Level 0 Milky Way
? Solar System ? Earth Level 1 Solar
System ? Earth Level 2 Earth ? Oceans ? Depth
Levels Level 2 Earth ? Clouds ? Cloud
Formations Level 2 Earth ? Continent ? Land
Segments Level 3 Land
Segments ? Urban Level 4 Urban ? Buildings ?
Building Type Level 5 Building Type ?
Houses
11
Using the Image Histogram
In our system, image size and resolution bears
no little relevance the only preprocessing done
to the images is the reduction of the pixel size
to values less than 500x500 pixels to speed up
the processing time. As it may be obvious,
texture varies considerably as a function of the
distance from the source. Therefore in order to
maintain texture consistency, all the images used
for each celestial body must come from pictures
taken within the same distance range. The
histograms of all the images are important in
ALISA, for example in this case, no texture
features or color components are needed since the
histograms of the are nicely spread out.
12
Histograms for Complex Images
Analysis of image histograms allow the system to
separate their components. however in complex
images the use of texture features as well color
separation is required for pattern
disambiguation.
13
Accumulated Histograms
Using texture features and examining the joint
and marginal histograms obtained from their
application, the classification process is highly
enhanced, since an accumulated histogram will
provide the spread histogram required to obtain
image segmentation, even though they still use
the full spectrum of the histogram. It is only
necessary to use few features with little
precision, in other cases if a texture feature is
eliminated or if the precision is reduced,
performance is degraded.
14
Training

• Select characteristic images for training
• Careful analysis of texture feature histograms
• Trial-and-error pilots to determine the optimal
  texture feature and color filters
• Anomaly detection and classification for the
  processing modules
• Define a distinctive color per pattern
• Two sets of different sets of images
• For training (small and representive of the
  application)
• For testing (where to prove the system)

15
Pixels and Texels Features

• Features used
• Texture Filters (3x3 texels context window)
• Media Range 0 to 20, Precision 30
• Media Range 75 to 100, Precision 30
• Activity de texeles Range 0 to 100,
  Precision 30
• Standard Deviation Range 0 to 100,
  Precision 30
• Color Filters (pixels)
• Blue Range 20 to 75, Precision 30
• Luminosity Range 20 to 75, Precision 30
• Green Range 20 to 75,
  Precision 30

16
Some results The Sun
Solar System Layer
Original Image
Mission Mariner10 Distance Date
Sun 71.4 Class1 11 Class2 8.9
Earth
Sun
Jupiter
Neptune
Rejected
Mercury
Moon
Saturn
Pluto
Tied
Space Ether
Venus
Mars
Uranus
17
Planet Jupiter
Solar System Layer
Original Image
Mission Voyager1 Distance ? Date ?
Jupiter 37.8 Class1 18.4 Class2 13.1
Earth
Sun
Jupiter
Neptune
Rejected
Mercury
Moon
Saturn
Pluto
Tied
Space Ether
Venus
Mars
Uranus
18
Planet Saturn
Solar System Layer
Original Image
Mission Cassini Distance 38.5 Million Kms Date
2004
Saturn 40.5 Class1 22.9 Class2 11.8
Earth
Sun
Jupiter
Neptune
Rejected
Mercury
Moon
Saturn
Pluto
Tied
Space Ether
Venus
Mars
Uranus
19
Reject Examples Note Here we desure the
rejection of non-solar celestial bodies
Rejected 26.6 Class1 21.1 ? Class2 17.5

The Milky Way
Rejected 53.5 Class1 41.4 Class2 1.3
Venus and Pleiades
Solar System Layer
Original Image
Earth
Sun
Jupiter
Neptune
Rejected
Mercury
Moon
Saturn
Pluto
Tied
Space Ether
Venus
Mars
Uranus
20
Planet Mercury
Solar System Layer
Original Image
Mission Mariner10 Distance Date
Mercury 60.8 Class1 17.9 Class2 9.3
Earth
Sun
Jupiter
Neptune
Rejected
Mercury
Moon
Saturn
Pluto
Tied
Space Ether
Venus
Mars
Uranus
21
Urban Distribution Layer

• Classes
• Street type 5 training images
• Highway type 1 training images
• Rural type 3 training images
• Building type 7 training images
• Tree type 2 training images
• Rejects 2 training images
• Frames 2 training images

• Features
• Mean
• D. R. 0 to 65
• Precision 30
• Gradient. Magnitude.
• D. R. 0 to 100,
• Precision 30
• Standard Deviation
• D. R. 0 to 100,
• Precision 30

22
Level 1. Image Recognition
Earth Original Image
Celestial Bodies Layer
Earth Components Layer
Earth Class
Continents
Space Ether Class
Oceans
Clouds
23
Earth Components Layer
Levels 2 3 Image Segmentation
Clouds Components Layer
Continent Components Layer
Ocean Components Layer
24
Earth Components
Clouds Components Layer
Continent Components Layer
Ocean Components Layer
25
South of Italy
(Meris Artifact photo)
Clouds Components Layer
Continent Components Layer
Ocean Components Layer
26
Strait of Gibraltar
(Meris Artifact photo)
Ocean Components Layer
Clouds Components Layer
Continent Components Layer
27
Strait of Gibraltar (Continent Components Layer)
Continent Components Layer
Mountain
Forest
Desert
Rejected
Plain Terrain
28
Continent Components Layer for urban
classification
Sidney, Australia
Continent Components Layer
Urban Class
Rejected
29
Continent Components Layer Urban Classification
(Airplane photo)
Puebla City, México
Continent Components Layer
Urban Class
Rejected
30
Level 4 Image Segmentation Urban Distribution
Layer
Puebla City, México
Urban distribution Layer
Build
Freeway
Trees
Rural
Streets
Rejected
31
Cholula Urban Distribution Layer
Cholula, México
Urban distribution Layer
Build
Freeway
Trees
Rural
Streets
Rejected
32
Puebla City Urban Distribution Layer
Puebla City, México
Urban distribution layer
Build
Freeway
Trees
Rural
Streets
Rejected
33
North America Urban Distribution Layer
Build
Freeway
Trees
Urban distribution layer
Rural
Streets
Rejected
34

• Second Example
• Segmentation of skin texturre

35
Configuration of the Skin Layer

• Texture Filters (3x3 texel window)
• Activity in x and y direction
• Gradient Magnitude
• Standard Deviation
• Media
• Color Features (pixeles)
• Blue
• Red
• Greed
• Saturation

• Trained classes and
• number of trainings
• Skin 24
• Almost skin tonality 9
• Almost white skin 15
• Almost brown skin 12
• Almost various skin 10
• Almost skin color 7
• Almost dark skin 12

36
Photo with flash
Image in family photo
segmentation
human skin
37
Photo from a Poster
segmentation
Image from a poster
human skin
38
Photo with textures close to skin texture
Segmentation
Family photo
human skin
39
Photo with flash and a back mirror
Reconocimiento de textura y color de la piel
Foto de familia
Segmentation
40
From Cumulative Histograms to STM entries for
classification
The CLS uses the cumulative histogram as the
entry to recognize a pattern. To classify it
uses an STM per class and looks for the one with
highest of acceptance after training.
41
Third example Recognition of animal sounds
42
Fast Fourier Transform (FFT)
The analysis of complex waves into their
sine/cosine wave components was proposed by Jean
Baptiste Joseph Fourier (1768-1830). Fourier
also proved that the components of periodic
waveforms are all integer multiples of the
fundamental frequency. A computationally fast
Fourier analysis was discovered by the Bell Labs
scientists Cooley and Tukey (1965). N -1 X(n)
1/N ? x(k) e2jk2pn/N for n 0 N-1
K0

• Time Domain
  Frequency Domain

43
Frecuency Spectrum
Rather than long lists of numbers specifying the
frequency and amplitude of each component, this
spectrum may be represented graphically. This 3D
function may be represented simply as a 2D
Spectrum image of vs frequency, where the
intensity of the pixel represents the amplitude
of the signal. In this Gamma Standard
display, white and yellow represent high
amplitude values, while red and blue medium and
small.
44
A hierarchical approach towards recognition
By using a hierarchy of ALISA layers, it is
possible to recognize different classes as
defined by supervised training. In the case of
audio signals the hierarchy may be organized as
follows Level 1 Frequency Layer (khz) ?
distinctive signal frequency Level 2
Time Layer (sec.) ? typical duration of sound
Level 3 Amplitute Layer (rms ) ?
standard power Level 4
Recognition of the Animal Sound Blue
Whale, Elephants, Capped Vireo, Parrot, . . .
45
0-2 KHz
1-4 KHz
3-9 KHz
Rej. / Tied
Original Sound Spectogram
Frequency Layer
Owl
Audio Signal
Bearded Seal
4 Sec
Bowhead Whale
Horse
0-2 Sec
Owl
Rej. / Tied
Rej. / Tied
Time Layer
Animal Layer
46
0-2 KHz
1-4 KHz
3-9 KHz
Rej. / Tied
Original Sound Spectogram
Frequency Layer
Bowhead Whale

Bearded Seal
4 Sec
Bowhead Whale
Horse
0-2 Sec
Owl
Rej. / Tied
Rej. / Tied
Time Layer
Animal Layer
47
0-2 KHz
1-4 KHz
3-9 KHz
Rej. / Tied
Original Sound Spectogram
Frequency Layer
Dog
4 Sec
Dog
0-2 Sec
Chimp
Cow
Rej. / Tied
Rej. / Tied
Time Layer
Animal Layer
48
0-2 KHz
1-4 KHz
3-9 KHz
Rej. / Tied
Original Sound Spectogram
Frequency Layer
Sheep
Reconocimiento de la señal de audio
Turkey
4 Sec
Wolf
0-2 Sec
Sheep
Rej. / Tied
Bees
Time Layer
Animal Layer
Rej. / Tied
49
0-2 KHz
1-4 KHz
3-9 KHz
Rej. / Tied
Original Sound Spectogram
Frequency Layer
Black Vireo
audio signal
Canyon Wren
0.5 Sec
Chesnut Sided Warbler
0.1 Sec
Black Capped Vireo
Rej. / Tied
Rej. / Tied
Time Layer
Animal Layer