19
1 Fingerprint Classification by SOM Wuzhili Vincent 99050056

1 Fingerprint Classification by SOM Wuzhili Vincent 99050056

  • View
    219

  • Download
    1

Embed Size (px)

Citation preview

1

Fingerprint Classification by SOM

Wuzhili Vincent 99050056

2

Introduction

Detailed Algorithms

Result Analysis

Demo

Fingerprint Classification by SOM

3

IntroductionWhy goes to this topic: 1. Explore the Industrial Usage of SOM ”Fingerprint Classification Through Self Organizing

Feature Maps Modified to Treat Uncertainties” Proceedings of the IEEE, Vol 84, No 10, pp 1497-1512, October 1996

2. Easily start by extending from my Honors Project

“Fingerprint Matching”

Fingerprint Classification by SOM

4

Introduction – Topics about Fingerprint

Fingerprint Classification by SOM

Fingerprint Sensor

Fingerprint Image Preprocessing

Fingerprint Matching

Fingerprint Image

Compression

Fingerprint Classification

5

Fingerprint Manually Classifiedby Experts

Left Loop Right Loop

Whorl Arch Tented Arch

DeltaPore

6

Introduction - Automated Classification

1. Might not according to the traditional 5-class scheme

2. Any uniformly distributed categories3. Consistently and correctly hash new

fingerprints into the categories

Fingerprint Classification by SOM

Class1

Class2

Class3

Class4

Class n

7

How to Classify Fingerprints by SOM

Introduction - SOM

1

4

23

X1

X2

X3

An Input vector X = {x1,x2,x3}

w11w13

w12w14

2x2 SOM

8

How to Classify Fingerprints by SOM

For a well-trained SOM:

1

4

23

X1

X2

X3

An Input vector X = {x1,x2,x3}

w11w13

w12w14

2x2 SOM

Winning Node

So the Input vector X is class 3 !

9

How to Classify Fingerprints by SOM

The Feature Vector of a Fingerprint X:1. X has dimension 1 x 256: {x1,x2,….x256}2. It is the directional Map

10

Extract the fingerprint region(right)

11

Extract the Effective Region

12

Locating Fingerprint Core

90% fingerprints centersare located

13

Uncertainty Value: [0, 1]

1.Directions in the good-quality region has good certainty;

2.In the Left figure:Larger certainty ->longer amplitude

14

Training Algorithm1: Original SOM

1. Contruct a MxM SOM, initialize all the weights2. Input a fingerprint vector: {x1,x2,….x256}3. Find the winning node dmin where: Dmin = min{||x-w||}4. Update the weight vectors:

W(new) =W(old) + Alpha*N*[x-w]Where N is the neighborhood function corresponding to the SOMnode topology

5. Repeat 2-4 till Update is not significant

15

Training Algorithm2: Modified SOM

Note: Each fingerprint is associated with a certainty vector C

1. Contruct a MxM SOM, initialize all weights2. Input a fingerprint vector:

X{x1,x2,….x256} = C*X + (1-C)*Xavg;3. Find the winning node dmin where: Dmin = min{||x-w||}4. Update the weight vectors:

W(new) =W(old) + Alpha*N*[x-w] * CWhere N is the neighborhood function corresponding to the SOMnode topology

5. Repeat 2-4 till Update is not significant

16

Experiment

Fa Fb Fc….. FA FB FC

1. Fa and FA are from the same finger Fb and FB … Fc and FC …2. Each fingerprint in DataA belongs to a class Class(Fa) = k , k within [1 ~ mxm]

Training Set (DataA) Testing Set (DataB)

17

Experiment

Fa Fb Fc….. FA FB FC

If all fingerprints are uniformly classified,Less accumulated worst search price-> Less DataA fingerprints are searched when indexing DataB

Training Set (DataA) Testing Set (DataB)

Class(FA) = ClassX

ClassXThe worst search price to find Fa is Size(ClassX)

Fa

18

  SEARCH%   RECOGNIO ON%    

    3x3 4x4 5x5 8x8 10x10M 10 12.1 18.8 28.8 91.8 100S 20 26.0 38.7 54.0 100  O 30 40.1 61.3 80.5    M 40 55.9 86.6 100      50 71.9 100        60 88.5          70 100          80          90            100        

  10 19.6 16.4 26.4 40.0 62.7 S 20 39.6 38.1 53.1 100 100O 30 57.3 60.9 82.5    M 40 72.9 84.8 100      50 86.3 100        60 97.5          70 100          80            90            100                       

ResultsSearch% column : percentage searched in DataARecognition% Column: percentage found for DataB

19

Advanced work can be done:

1.Increase the layer of SOM to solve the crowded class with many fingerprints2. Principle Component Analysis to reduce the feature vectorfrom 256 to small dimensions. [40 dimensions are feasible]