Statistical Evaluation of High-FRR Fingerprints 

Dr. Manfred Bromba

http://www.bromba.com/contacte.htm
Permanent address for citation: urn:nbn:de:0125-20080325104
First version: 2000-03-28 - Corrected: 2008-04-23

 
False acceptance rates (FAR) and false rejection rates (FRR) of fingerprint systems depend on a large number of influences. As a result, personal FRRs show strong variations with respect to person and "time", where time is an indirect variable pointing to much more complex influences.

There are several reasons to focus on persons which deliver too large error rates, especially FRRs:

  • Depending on the FRR measurement definition, small groups with high personal FRRs have great influence on total FRR
  • The failure-to-enroll / failure-to-acquire rates are important quantities for the application environment and the application feasibility, depending on application
  • Large error rates have to be analyzed to improve fingerprint sensor and recognition software
  • The investigation of problem cases and their statistics within a population is inevitable for the determination of a representative test population and test data basis.

High-FRR cases
Determination of critical personal FRRs and FTA frequency
What can be done to improve the FRRs?

High-FRR cases

In some cases and for some users the rejection rate is unacceptably high so that multiple recognition trials become inevitable.

What are possible reasons for high-FRR cases?

Classifications

  1. Physiological/operational/hardware/software problems
  2. temporal/permanent problems

Physiological problems

Physiological problems cannot be solved by the user alone and thus have mainly to be addressed by the system developer.
Wet fingers
Especially capacitive sensors may have problems with wet fingers. Wet fingers often occur with young users, in warm environment or with excited users. The situation with wet fingers may partly be defused by the user (drying fingers).
Dry fingers
Especially capacitive sensors may have problems with dry fingers. Dry fingers often happen with elder users. The situation with dry fingers may partly be defused by the user (increase pressure and wait longer).
Minutia scarcity
Some users may have too few minutia to be detected without problems, depending on sensor area.
Skin disease
Some kind of skin disease may destroy or disturb the natural finger structure.
Skin abrasion
Depending on sensor type, many handicraft activities may decrease the ridge heights such that many sensors deliver only small-contrast pictures. This effect is reversible.
Other problems
  • ...

Operational problems

These problems arise from wrong usage and can be solved by system design and user behavior.
Wrong finger pressure
Is the pressure on the sensor too high, the image quality may degrade (same effect as wet fingers). If the pressure is non-uniform and non-vertical, warping may prevent proper recognition.
Wrong finger positioning
Usually rotation and translation is limited (this limitations may be adjusted by software) because of limited sensor area and because of fake protection. Sometimes a finger guide is not accepted by a user because it may be insufficient or too difficult to be understood.
Finger contamination
Contaminated or even dirty fingers degrade image quality of fingerprints. For different sensor types the type of harmful contamination may be different. Frequent contaminations come from skin care substances such as cream.
Sensor contamination
Some types of fingerprint sensors are sensitive to sensor surface contaminations, e.g. skin cream. Surface contaminations superimpose nonlinearly with actual fingerprints and may prevent recognition. Moving the finger slightly on the sensor surface or cleaning the sensor may solve the problem.
Other problems
  • Some users like to wear stoves and do not like to undress them for identification

Hardware Problems

Hardware problems have to be solved by the fingerprint system supplier.
Finger guidance
The effect of an appropriate mechanical finger guide is better positioning and the minimization of finger warping
Sensor problems
  • Spatial resolution too small
  • Gray level resolution too small
  • Further ADC problems
  • Pixel defects (point defects, line defects, area defects)
  • Resistance to contaminations
  • Noise (correlated and statistical)

Software problems

Feature extraction
  • Finds too many minutia
  • Finds too few minutia
  • Inaccurate minutia position
  • Inaccurate minutia type
  • Inaccurate minutia angle
Matcher
  • Insufficient processing or suboptimal weighing of features (position, type, angle, ...)
  • Warping processing insufficient
  • Suboptimal adjustment of global rejection criteria (fingerprint translation, rotation, minutia quantity, image quality, ...)

Failure-to-acquire (FTA)

Failure-to-acquire is the most unpleasant form of high FRRs: the user delivers such a bad fingerprint image quality that no reasonable detection or enrollment is possible (FRR=100%). This effect may be either temporal or even permanent.

What are the reasons for FTA?

In general, the reasons for rejection are the same as for high-FRR cases, except for operational problems which can be avoided by the user. The remaining problems are physiological problems, hardware problems, and software problems where the order indicates the priority for toady's advanced systems.

Determination of critical personal FRRs and FTA frequency

Search for critical users

Critical users can be identified by at least two ways:
  1. field trial
  2. purposeful search for critical groups by trial-and-error
The first method is simple and gives some good safety to catch all significant problem users. On the other hand, the second method requires additional statistics. The first method is to be preferred if the result shows a relatively large number of critical cases, the second one for a small number. Since that number is dedicated to be small in practice, the following ideas are based on the second method:
  • find a correspondence between error causes and user groups with known statistics (example: bricklayers (n% of the total population) should have flattened ridges)
  • run a trial with such a  group to confirm the hypothesis and to get the personal FRRs (example: prove that bricklayers have really flat ridges)
  • calculate the total FRR using known statistics of the personal FRRs
However, first of all, it has to be tested that the critical cases are really small in number. This suggests to start with method 1, eventually improve system performance and measurement methodology, and then, if all error rates are small enough, switch over to method 2.

It must be emphasized that method 2 has not yet been evaluated and that there remains some risk that it will not work at all. If that should happen, it will become extremely difficult to prove error rate below 1% at reasonable cost.

Available statistics

  • age
  • sex
  • temperature
  • humidity
  • kind of sports
  • profession
  • disease
  • addiction
  • hobbies
  • ...

Statistics requirements

  • should cover the whole population
  • should be independent
  • should be disjunct

Examples

  • If skin moisture depends on age, with a tendency from wet to dry with increasing age, critical users should be found especially for very young (too wet) and very old people (too dry).
  • The dependency between temperature and error rates has to be tested systematically. From temperature statistics the effect on total FRRs may be determined.
  • Bricklayers stress their fingers' skin more than most other professions. Hence, a field trial with a small (depending on failure rates) number of bricklayers is necessary.

Critical user groups

critical user group available statistics temporal permanent
bricklayer profession
X
 
mountaineer hobby
X
  
young girls age+sex
 
X
old men age+sex
 
X
tiler profession
X
 
neurodermitis disease
X
X
smoker (?) addiction  
X
stove wearer unknown
X
  
swimmer (?) sports
X
 

Test organization

The test starts with a small field trial (100 participants) using the ID Mouse. If the intermediate results show an FTA and an ERR of smaller than 1 % for normal users, small field trials are extended to possibly critical user groups. The critical user group is arranged, e.g., by an institute which organizes focus groups. During a 20-person/2-hour session these users have to leave about 500 fingerprints/person using the ID Mouse equipment. Cost: about 5000€. Tests with medical background should be run, e.g., as medical dissertation.

What can be done to improve the FRRs?

Physiological problems

Physiological problems cannot be solved by the user alone and thus have mainly to be addressed by the system developer.
Wet fingers
For capacitive sensors the amplitude resolution in the region of water seems to show the highest potential for improvements. The situation with wet fingers may partly be defused by the user (drying fingers).
Dry fingers
For capacitive sensors low noise and high resolution in the region of air is important. The situation with wet fingers may partly be defused by the user (increase pressure and wait longer).
Minutia scarcity
Increase sensor area.
Skin disease
No remedy known.
Skin abrasion
Sensors with high amplitude resolution may have advantages.

Hardware Problems

Hardware problems have to be solved by the fingerprint system supplier.
Finger guidance
The finger guide has to be optimized for different finger sizes and types. Finger warping and random positioning has to be avoided. No rules available today.
Sensor problems
  • Spatial resolution > 500 dpi
  • Gray level resolution at least 200 steps
  • ADC must be monotonous
  • Pixel defects (point defects, line defects, area defects) have to be minimized
  • Surface resistant against contaminations
  • Noise (correlated and statistical) < 1 LSB
  • Sensor area > 150 mm²

Software problems

Feature extraction
  • Robust algorithms
Matcher
  • Robust against warping, minutia positioning, minutia rotation, ...