Pattern Recognition without Features Fred Stentiford University - - PowerPoint PPT Presentation

▶

Nov 20, 2023 17 likes •189 views

Pattern Recognition without Features Fred Stentiford University College London http://www.ucl.ac.uk/~fstentif/ Issues Selection of Features Training Process Commonality Reference Candidate Graphical Similarity Representation

SLIDE 1

Pattern Recognition without Features

Fred Stentiford

University College London http://www.ucl.ac.uk/~fstentif/

SLIDE 2

Issues

Selection of Features
Training Process

SLIDE 3

Commonality

Reference Candidate

SLIDE 4

Graphical Similarity Representation

Pixels are represented as nodes
Nodes possess the properties:
Location of pixel
Brightness gradient orientation
A relationship exists between a pair of nodes if

their properties and relative orientation match that

f a pair in a second image.
A maximal clique is the largest subset of nodes

that all match and possess an orientation relationship with each other.

SLIDE 5

Cliques of Matching Pixels

x1 x2 x3 x4 x5

φ34 φ12 φ23 φ45 φ51 φ13 φ14 φ31 φ35 φ42 φ41 φ52 φ53

y1 y2 y3 y4 y5

Φ’34 Φ’12 Φ’23 Φ’45 Φ’51 Φ’13 Φ’14 Φ’31 Φ’35 Φ’42 Φ’41 Φ’52 Φ’53

Similarity = size of maximal clique

SLIDE 6

Yale Face Database A

SLIDE 7

Yale Face Database A

SLIDE 8

Face Matching

1764 points

Relative orientation ε1<19°; Gradient direction ε2<55°;

reference

Gradient Directions

SLIDE 9

Matching

Close-up showing matching gradient orientations

SLIDE 10

Results - Expressions

1000 1500 2000 2500 3000 3500 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Subject Maximal clique size Happy Surprised Wink Sleepy Glasses Sad No Glasses

No errors

SLIDE 11

Results - Illumination

Gradient direction not affected in many areas

SLIDE 12

Results - Illumination

800 1000 1200 1400 1600 1800 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Subject Maximal clique size Rightlight Centrelight Leftlight

No errors

SLIDE 13

Results - Occlusions

200 400 600 800 1000 1200 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Subject Maximal clique size Top Surprised Left Surprised Right Surprised Bottom Surprised

Errors - bottom:6/15

SLIDE 14

Results - Occlusions

SLIDE 15

Results - Occlusions

SLIDE 16

Neuron Sensitivity

Certain neurons are sensitive to brightness gradient vector orientation (Podvigin 2001)

SLIDE 17

Benefits of the Approach

Avoids the need for feature selection
No training data required
Automatic location of match
Displays immunity to distortions
Potential for parallel implementation