Fingerprint matching from minutiae texture maps -com

Fingerprint matching from minutiae texture maps

Contents

• The Original Approach--Fingercode– Minutiae, Tessellation, Gabor Filter, Fingercode

• First Improvement– Minutiae as reference points, local orientation

• Second Improvement– Section weight, variations

• Performance and Evaluation

Background

• Fingerprint matching– Widely used

– Give an yes-or-no answer

– fast and convenient

• Shortage for previous approaches– Be of variable length

• Minutiae– Ridge Endings

– Bifurcations

Fingercode

• a 640-byte feature vector

• The matching process

– calculate the Euclidean distance between the sensored fingercode and the template fingercode

Fingercode—1

• Reference Frame Determination

– reference point: Maximum curvature

– reference axis: local symmetry of ref point

Fingercode—2

• Tessellation

– A plate of 120 pixels is selected and divided into 80 section

• Nomalization

Fingercode—3

• 3. Gabor Filter

– 8 directions Gabor filters

Fingercode—3

normalized

135

Fingercode—4

• Feature Vector Extraction

– standard deviation in 8*80 sections

– 640 bytesFinger 1

Finger 2

Why does it work...

• Each Garbor filter contains information about ridges and furrows both globally and locally.

– Band width is 20 pixels.

– Inter-ridge width is 10 pixels.

– So large variation means ridges and furrows along the direction.

Why does it work...

• Detected ref point can be 12 pixels away.

• Orientation can be 20 away

– Since fingerprint is "smooth", and we use statistical (rather than is-or-no) data.

First Adjustment

• Minutiae as reference points

– Fingercode at every minutiae

– Ref axis along the minutiae direction

1.Avoid pre-alignment risks 2.More robust

Second Adjustmdent

• Weighting: ADD

Problem?

• Wrong detection of ref point can make things disastrous

• Errors

– Location Errors

– Orientation Errors

Error Correction

• Orientation variation

• Localization Variation

7 Pixels

1

2

3

4

Matching

• Input fingerprint’s Minutiae list

• Template fingerprint’s Minutiae list

• Compute the Distances for each pair

• Find the minimum one

8 Directions

80 sectors

Test Data Base

• Exploit databases from Fingerprint Verifictation Competition

– FVC2000

– FVC2002

• 100 distinct fingers for each base

• 8 impressions for each finger

Experimental Results—EER-curve

• DB1 FVC2000

• d

• varied the minuti-ae orientations

• Compute FRR

Experimental Results—EERs

• DB1 FVC2000

• EERs estimated for different orientation variations

• ERR < 6%

Experimental Results

Experimental Results

• Matching performance

– Location errors: [-7,+7]

– Orientation variations:

• FVC2002

Experimental Results

• Rank according to EER

• FVC2002

• compared with other TOP 31 participants

Experimental Results

• Ref Points selection compared to fingerCode

• Reasons

– Noise

– The ref point is close to border

– Scars near the ref points(DB3-a)

DB1-a DB2-a DB3-a DB4-a

Cons and Pros

• Pros

– Avoid pre-alignment risk

– Able to get correct result from poor picture.

• Cons

– Computaional expensiveness

Assume 30 minutiae every picture, we have to compute 30*30=900 pairs.

END

Thank you!