A Systematic Approach For Feature Extraction in Fingerprint Images

1
A Systematic Approach For Feature Extraction in
Fingerprint Images

• Sharat Chikkerur, Chaohang Wu, Venu Govindaraju
• ssc5,cwu3,govind_at_buffalo.edu

2
Abstract

• A new enhancement algorithm based on Fourier
  domain analysis is proposed.
• Fourier analysis is used extract orientation,
  frequency and quality map in addition to doing
  enhancement.
• The enhancement algorithm uses full contextual
  information and adapts radial and angular extents
  based on block properties.
• A new feature extraction algorithm based on
  chain code analysis is presented.
• An objective metric is used to evaluate the
  efficiency of the feature extraction.

3
Outline

• Related Previous Work
• Overview of the proposed method
• Fourier Analysis
• Fingerprint Image Enhancement
• Feature Extraction
• Performance Evaluation
• Conclusion

4
Motivation Enhancement

• Anisotropic filter (Greenberg et.al, Yang et.al)
• Very fast but cannot handle creases, wide breaks
  and poor quality images
• Pseudo Matched filtering (Wilson, Grother Candela
  et. al)
• Increases SNR but can lead to artefacts due to
  isotropic filtering.
• Directional Filtering (Sherlock,Monro et. al.)
• Very robust even near regions of high curvature
  but marked by large storage requirements.
  Frequency of ridges is assumed to be constant.
• Gabor filter bank(Hong et. al)
• Filter has optimal joint directional and
  frequency resolution but does not handle high
  curvature regions well due to block wise
  approach. Angular and radial bandwidths are
  constant.
• Proposed approach
• A single algorithm is used for contextual
  analysis and enhancement.
• Utilized full contextual information. Adapts both
  frequency and angular bandwidth based on block
  properties.
• Adapts to high curvature regions reducing
  blocking artifacts.
• However, using full contextual information leads
  to processing complexity.

5
Qualitative Comparison Feature Extraction

• MINDTCT,NIST NFIS, (Garris et. al)
• The algorithm is extremely fast.
• Greedy approach to minutia detection leads to
  false positives.Extensive post processing is
  required to eliminate false positives
• Adaptive Flow Orientation technique (Ratha et.
  al.)
• Is capable of correcting breaks in the rides and
  is robust to noise.
• Peak detection leads to false positivies in
  regions of poor ridge constrast.Also, thinning
  and morphological post processing shift minutia
  location.
• Direct Gray Scale Ridge Following (Maio and
  Maltoni)
• Does not have errors introduced due to
  binarization and has low computational
  complexity.
• Cannot handle poor contrast prints and images
  with poor ridge structure.
• Proposed method
• Enhancement reduces spurious and missing
  minutiae. The locations of the minutiae are
  preserved during detection.
• Contour based extraction is sensitive to
  binarization and enhancement errors.

6
Outline

• Related Previous Work
• Overview of the proposed method
• Fourier Analysis
• Fingerprint Image Enhancement
• Feature Extraction
• Performance Evaluation
• Conclusion

7
Overview of the proposed method
Enhancement
Fourier Analysis
Contextual Filtering
Preprocessing
Gray Level Image
Feature Extraction
Feature Extraction
Binarization
Contour Extraction
Minutiae Detection
8
Overview of the proposed method
Enhancement
Fourier Analysis
Contextual Filtering
Preprocessing
Gray Level Image
SNR is increased using Pseudo Matched
filtering Wilson et. Al, 1994, k 0.15 is used
to reduce artifacts
Feature Extraction
Feature Extraction
Binarization
Contour Extraction
Minutiae Detection
9
Overview of the proposed method
Enhancement
Fourier Analysis
Contextual Filtering
Preprocessing
Gray Level Image
The image is divided into blocks and Fourier
analysis is done on each of them. The analysis
produces orientation, frequency, angular
bandwidth and quality maps proposed
Binarization
Contour Extraction
Minutiae Detection
10
Overview of the proposed method
Enhancement
Fourier Analysis
Contextual Filtering
Preprocessing
Gray Level Image
Each block is filtered using a orientation and
frequency selective filter Sherlock and Monro,
1994 with the given bandwidth
Feature Extraction
Feature Extraction
Binarization
Contour Extraction
Minutiae Detection
11
Overview of the proposed method
Enhancement
Fourier Analysis
Contextual Filtering
Preprocessing
Gray Level Image
Feature Extraction
Feature Extraction
Binarization
Contour Extraction
Minutiae Detection
The enhanced image is binarized using an
locally adaptive algorithm
12
Overview of the proposed method
Enhancement
Fourier Analysis
Contextual Filtering
Preprocessing
Gray Level Image
Feature Extraction
Feature Extraction
Binarization
Contour Extraction
Minutiae Detection
Contours of the ridges are extracted and traced
consistently in a counter clockwise
directionGovindaraju et. al, 2003
13
Overview of the proposed method
Enhancement
Fourier Analysis
Contextual Filtering
Preprocessing
Gray Level Image
Feature Extraction
Feature Extraction
Binarization
Contour Extraction
Minutiae Detection
Minutiae are detected as points with 'signficant'
turns in the contour. Vector products are used to
quanity the turns
14
Outline

• Related Previous Work
• Overview of the proposed method
• Fourier Analysis
• Fingerprint Image Enhancement
• Feature Extraction
• Performance Evaluation
• Conclusion

15
Surface Wave Model
16
Validity of the model

• With the exception of singularities such as core
  and delta, any local region of the fingerprint
  has consistent ridge orientation and frequency.
• The ridge flow may be coarsely approximated using
  an oriented surface wave that can be identified
  using a single frequency f and orientation ?.
• However, a real fingerprint is marked by a
  distribution of multiple frequencies and
  orientation.

17
Obtaining block parameters

• To obtain the dominant ridge orientation and
  frequency a probabilistic approximation is used
• We can represent the Fourier spectrum in polar
  form as F(r, ?) The power spectrum is reduced to
  a joint probability density function using
• The angular and frequency densities are given by
  marginal density functions

,
18
Obtaining block parameters (contd.)

• The dominant ridge orientation is obtained using
• The dominant frequency can be estimated using the
  expected value of the frequency density function,
• The quality is assumed to be proportional to the
  strength of the ridge flow and is estimated using

19
Fourier Analysis Energy Map
Original Image
Energy Map
20
Fourier Analysis Frequency Map
Original Image
Local Ridge Frequency Map
21
Fourier Analysis-Orientation Map
Original Image
Local Ridge Orientation Map
22
Fourier Analysis Angular Bandwidth
23
Outline

• Related Previous Work
• Overview of the proposed method
• Fourier Analysis
• Fingerprint Image Enhancement
• Feature Extraction
• Performance Evaluation
• Conclusion

24
Fourier Domain Based Enhancement
Enhanced Image
Contextual Filter
Original Image
25
Additional Enhancement Results
26
Outline

• Related Previous Work
• Overview of the proposed method
• Fourier Analysis
• Fingerprint Image Enhancement
• Feature Extraction
• Performance Evaluation
• Conclusion

27
Determination of Turn Points

• When the ridge contours are traced in a counter
  clockwise direction, minutiae are encountered as
  points with significant turn.
• Types of turn points left(ridge),right(bifurcatio
  n)
• S(Pin, Pout) S( )S(x1y2 x2y1)
• Pin Vector leading into the candidate point
• Pout Vector leading out of the point of interest
• S(Pin, Pout) gt0 indicates left turn, S(Pin, Pout)
  lt0 indicates right turn
• Significant turn can be determined by
• ( )x1y1 x2y2 lt T

28
Turn points
(a) Potential minutia location (b) Determination
of turn points
29
Post processing

• Feature Extraction errors
• Missing minutiae
• Spurious minutiae
• Spurious minutia can be removed using post
  processing
• Heuristic rules
• Merge minutiae that are a certain distance of
  each other and have similar angles
• Discard minutiae whose angles are inconsistent
  with ridge direction
• Discard all border minutia
• Discard opposing minutiae within certain distance
  of each other

30
Example Result
31
Outline

• Related Previous Work
• Overview of the proposed method
• Fourier Analysis
• Fingerprint Image Enhancement
• Feature Extraction
• Performance Evaluation
• Conclusion

32
Quantitative Analysis

• Test Data
• 150 prints from FVC2002(DB1) were randomly
  selected for evaluation.
• Ground truth was established using a semi
  automated truthing tool.
• Results compared using NIST NFIS open source
  software.
• Metrics
• We use feature extraction metrics proposed by
  Sherlock et. Al
• Sensitivity Ability of the algorithm to detect
  true minutiae
• Specificity Ability of the algorithm to avoid
  false positives
• Additional Metrics
• Flipped Minutiae whose type has been exchanged

33
Quantitative Analysis Results

• Examples

34
Quantitative Analysis Results

• Summary results
• Count TP(ANSI) gt proposed 40 of 150
• Count E(ANSI) lt proposed 40 of 150

Sensitivity distribution
Overall statistics
35
Conclusion

• A new effective enhancement algorithm based on
  Fourier domain analysis is proposed
• A single algorithm is used to derive orientation,
  frequency, angular bandwidth and quality maps
• A new feature extraction algorithm based on chain
  code contour analysis is presented
• Heuristic rules specific to the feature
  extraction algorithm has been derived
• The algorithm is evaluated using an objective
  metric

36
Thank You

• http//www.cubs.buffalo.edu

37
Related Previous Work Enhancement

• Spatial Domain
• Anisotropic filter (Greenberg et.al, Yang et.al)
• Uses a locally adaptive kernel
• Blurs along the ridge direction. Increases the
  discrimination between ridges and valleys along
  the perpendicular direction.
• Frequency Domain
• Pseudo Matched filtering (Wilson, Grother Candela
  et. al)
• The Fourier transform of the block is multiplied
  by its power spectrum raised to a power of k
• Directional Filtering (Sherlock,Monro et. al.)
• The image is decomposed into a set of eight
  directional responses using a bank of
  directionally selective filters. The frequency is
  assumed constant.
• The enhanced image is obtained by composing the
  filter responses using the local orientations.
• Gabor filter bank(Hong et. al)
• The image is enhanced by using a Gabor filter
  bank
• Gabor fillters have the optimum orientation and
  frequency resolution.

38
Related Previous Work Feature Extraction

• Binarized Images
• MINDTCT, NIST NFIS, (Garris et. al)
• An oriented grid is placed at each pixel and the
  projection sums are taken at each row. The pixel
  is assigned 0 if the projections sum at the
  center row is less than average, otherwise the
  pixel is assigned 1
• The minutiae are detected using structural rules.
• Adaptive Flow Orientation technique (Ratha et.
  al.)
• Orientation of each 16x16 block is determined by
  computing the gray level projections at various
  angles. The projection along a scan line
  perpendicular the ridge direction has maximum
  variance.
• The image is binarized by detecting the peaks
  along this scan line.
• The minutiae are detected using the thinned image
• Gray Scale Image
• Direct Gray Scale Ridge Following (Maio and
  Maltoni)
• A set of starting points are chosen by
  superimposing a grid on the image
• The ridge is traced from each starting point
  until a bifurcation or ridge ending is found.
• A labelling strategy is used to preven
  traversing the same ridge twice.