SPIE - Horizonimaging.com

clangedbivalveΤεχνίτη Νοημοσύνη και Ρομποτική

19 Οκτ 2013 (πριν από 4 χρόνια και 20 μέρες)

82 εμφανίσεις

IMAGE PRE
-
PROCESSING FOR CLASSIFICATION


(BIOMETRIC IDENTIFICATION)

BY A NEURAL NETWORK




Anthony Vannelli, Steve Wagner, and Ken McGarvey

Horizon Imaging, LLC


email: info@horizonimaging.com


Horizon Imaging, LLC

Innovative Solutions in Image Processing



Raw
512x480
Image


Neural
Preprocessor


Neural
Network
Classifier

Reduced
data set

Classification
Output

Neural Network Preprocessor and Classifier



Wavelets



PCA



Image “Zones”



Combining Networks



Feed
-
forward Network



Back
-
propagation Training



Single Hidden Layer


Horizon Imaging, LLC

Innovative Solutions in Image Processing





512x480 raw image or 245,760 inputs to network



Large neural network



Poor classification performance



Slow convergence

Curse of dimensionality

Horizon Imaging, LLC

Innovative Solutions in Image Processing



Biometric Identification

Region of Interest

320x160 = 51,200 pixels

Horizon Imaging, LLC

Innovative Solutions in Image Processing



Preprocessing Techniques


Non
-
parametric


“Holistic”


Data
-
driven


No Hand Geometry


No Fidiucial Points


Horizon Imaging, LLC

Innovative Solutions in Image Processing



Preprocessing Techniques


Principal components


Large eigen
-
values help to classify


Reduces dimensionality


Image Processing Zones


Divide and conquer


2x2 zones (160x80 pixels)


4x4 zones (80x40 pixels)


Ensemble of neural networks

Horizon Imaging, LLC

Innovative Solutions in Image Processing



Preprocessing Techniques


Combining Neural Networks


Pick the network with the “best fit”


Average the network outputs


Voting Scheme



Horizon Imaging, LLC

Innovative Solutions in Image Processing



Voting Scheme to Combine Networks



Neural Net #1

Neural Net #2

Neural Net #N




1


2


N

Figure 3.
Voting scheme to combine Neural Networks



Input Vector

















































y
N

y
2

y
1

=



i

y
N

> T

y
2

> T

y
1

> T


i

=

0 for y
i

T

1 for y
i

> T



Combined

Output

Horizon Imaging, LLC

Innovative Solutions in Image Processing



Preprocessing Technique using Wavelets




Coiflet wavelet


Daubechies wavelet


Haar wavelet (averages adjacent pixels)

Second
-
level wavelet approximation

Horizon Imaging, LLC

Innovative Solutions in Image Processing



Image
f(x,y)

Low

High

Low

High

Low

High

LL

LH

HL

HH

Horizontal filter

Vertical filter

2

2

2

2

2

2

One
-
Level of a Wavelet Transform

Horizon Imaging, LLC

Innovative Solutions in Image Processing



Third
-
level Wavelet Decomposition

HH

LH

LL

HL

Horizon Imaging, LLC

Innovative Solutions in Image Processing



Test Case with Single Classifier

Output

Figure

7.

Test case with single classifier

320x160 pixels

Wavelet
Transform


PCA

Neural
Classifier

512 x 480
Image

Image
Preparation

Horizon Imaging, LLC

Innovative Solutions in Image Processing



Test Case with Multiple Classifiers

Image 1

Neural
Classifier

Image N

Neural
Classifier

Combine
Networks

Wavelet
Transform

Output

Image
Preparation

320 x 160
pixels

512 x 480
Image

Figure

8.

Test case with multiple classifiers

Horizon Imaging, LLC

Innovative Solutions in Image Processing



Test Cases

A.
Coiflet 6
-
coefficient wavelet to 3 levels; 3
rd

level
approximation image (40x20 pixels) and 3
sidebands form input to 4 neural networks with
800 inputs each.

B.
Daubechies 6
-
coefficient wavelet to 3 levels; 3
rd

level approximation image (40x20) and 3
sidebands form input to 4 neural networks with
800 inputs each.

C.
Coiflet 6
-
coefficient wavelet to 2 levels (80x40
pixels); 4 image zones fed to 4 separate neural
networks with 800 inputs each.

Horizon Imaging, LLC

Innovative Solutions in Image Processing



Test Cases

D.
Daubechies 6
-
coefficient wavelet to 2 levels
(80x40 pixels); 4 image zones fed to 4 separate
neural networks with 800 inputs each.

E.
Harr wavelet to 2 levels (80x40 pixels); 4 image
zones fed to 4 separate neural networks with 800
inputs each.

F.
Harr wavelet to 2 levels (80x40 pixels) and then
PCA transform fed to a neural network with 512
inputs.

Horizon Imaging, LLC

Innovative Solutions in Image Processing



Test Cases

G.
Harr wavelet to 3 levels (40x20 pixels) fed to a
neural network with 800 inputs.

H.
Coiflet 6
-
coefficient wavelet to 1 level (160X80 =
12800 pixels). The first level approximation image
is divided into 16 image zones (40x20 pixels per
zone). The zones are fed into separate neural
networks with 800 inputs each.

Horizon Imaging, LLC

Innovative Solutions in Image Processing



Summary of Performance

Horizon Imaging, LLC

Innovative Solutions in Image Processing