Consortium of Forensic Science Organizations (CFSO) Technology Fair

we hear about last week's NIST Latent Testing Workshop.
_________________________________________

2006 NIST Latent Testing Workshop
by Kasey Wertheim

Last week, April 5-6, 2006, NIST held a Latent Testing Workshop at their campus in Gaithersburg, Maryland. Approximately 50 attendees were present for 2 full days consisting mostly of a series of 20-minute lectures on various aspects of latent prints. The goal of the workshop was to bring together practitioners and academics so NIST could map a pathway toward creating latent sample sets sufficient for automated latent testing. The end goal is a “Latent Grand Challenge” where NIST can process standardized test sets through various combinations of latent print extraction and search algorithms provided by vendors.

The workshop presentations began with briefs by Steve Meagher (FBI/Lab), Kasey Wertheim (DoD) and Danny Greathouse (DHS) on federal/military aspects of latent print needs. This was complimented by the next hour, which included presentations from Thomas Smith (LAPD), Frank Sinese (Illinois) with perspectives on live-scan.  Scott Swann (FBI/CJIS) defined NGI initiatives and specifically discussed the development of a National Marketing Plan for latent prints along with increases in latent print search reliability as a part of algorithm improvements.

Debbie Leben (USSS) spoke on needed improvements for the Universal Latent Workstation:
*Maintain feature sets during cropping, image reversal, and changes in orientation,
*Add filters for use in combination, not just individually: Unsharp mask, pattern removal, fade filter
*Ridge counting, including additional clarification on what to count
*More efficient comparison capabilities, including auto-sizing the latent and tenprint
*Ability to view features on the latent and the tenprint that produced the score
*Ability to know the image being viewed was the image that produced the score
*Case archive/saving option
*Reporting and statistics, which will involve recording hit/no-hit decisions and user profiles
*Queued submission and timed release capability
*Latent sizing ability for non 1:1 latent prints
*Ability to format and submit multi-finger latents, or simultaneous pattern information
*Lights-out auto-search by clusters
*Ability to encode and search Complete Friction Ridge Exemplars
*Ability to search by group (crime type) or region (geographical search)
*Refined ULM receipt process: Single candidates instead of multiple candidates per submission, and real-time ULM reporting (as opposed to batch processing in IAFIS)

Gordon Lowe (CalDOJ), Wade Petroka and Jeri Eaton (King County Washington) spoke on local AFIS issues and needs.

Ambika Suman spoke on the PITO latent test design for the Ident1 benchmark in the UK.

Mark Branchflower discussed the role of latent prints at Interpol as increasing, and that two positions would soon be approved for full-time latent print examiners.

Several AFIS vendors (Sagem Morpho, Cogent, NEC, Motorola) spoke on new and improved functions and capabilities leading to better performance of latent algorithms.

Philip Wasserman from NIST presented "Level 3 Fingerprint Feature Recognition Using Support Vector Machines"  In cases where level 1 and 2 detail (patterns and minutia) are not sufficient to permit identification, trained human examiners rely on level 3 features such as the position and size of sweat pores, incipient ridges, edge formations and other fine detail. Achieving this level of AFIS processing requires machine intelligence that performs on par with the accuracy of the Latent Print Examiner. Sample collection involved an experienced examiner “marking” (selecting) an example of a level 3 feature by clicking on its location in a computerized image. A set of features was extracted from the region around the selected point and the set was used to train an SVM. The trained machine was used to classify samples not included in its training set. Vapnik (1995) provides a highly effective means for pattern classification and regression. SVM provides a more global solution in contrast to a neural network where many non-optimal solutions result. The preliminary results, while encouraging, are not conclusive. Additional examples must be collected to produce statistically significant accuracy estimates.

Yi Chen (co-researcher, Anil Jain, Michigan State University) presented "Pores and Ridges: High Resolution Fingerprint Matching Using Level 3 Features"  Although experienced latent print examiners often rely on level 3 detail to assist in identification, AFIS mostly rely on Level 1 and Level 2 features. With advances in sensing technology, many sensors are now equipped with 1000ppi scanning capability and systematic studies of performance gain need to be conducted in this area. Level 3 features can be automatically extracted by combining images subsequent to wavelet transform and Gabor filtering. Their experiments show that level 3 features carry significant discriminatory information and can be locally matched using “the ICP algorithm”. They showed a relative reduction of 20% in the equal error rate of the matching system when level 3 features are employed in combination with levels 1 and 2, across a variety of fingerprint image quality. Images were shown of extracted images displaying pore location and ridge edge tracings.

Nigel Allinson from the University of Sheffield presented work that has been done in England to support the Lincolnshire Police.  They have equipped their crime scene officers with laptops and scanners capable of scanning 1000ppi images.  Nigel and his team wrote software that allows them to save JPEG 2000 (JP2) images at 16:1 compression, and securely transmit them with PGP Private/Public key incription back to the AFIS section where they can immediately been encoded and searched in AFIS.  On average, they have taken backlog times from 14 days to 1 day, and in their fastest response time provided police a lead less than 3 hours after they had arrived on scene.  When they arrived at the suspect's home, he was found asleep and all the stolen goods were still in his custody.

Tom Hopper spoke on ULW, Elham Tabassi spoke on the NIST Quality Workshop, and George Kiebuzinski spoke on lessons learned during IAFIS source selection.

Kasey Wertheim spoke on 7 Latent Print Quality Measures:
Quantity: square area
Clarity: quality, focus, resolution, crispness
Contrast: ridge to furrow difference in grayscale intensity
Pressure: ridge to furrow difference in width/thickness
Slippage: measure of lateral pressure, smearing (difficult to quantify)
Background: interference of the background – texture/pattern prominence
Focal Points: present/absence of delta, core, scars, creases, or other major features.

Some of these “human” measures of quality (or latent print “difficulty” to compare) may be applied locally in an automated way to boost or reduce scoring (as appropriate) during the matching process to increase reliability and selectivity.

Patrick Grother of NIST spoke on offline biometric testing and Stephen Wood discussed latent test sets (including the NIST Special Database (SD) 27 and a Secret Service (non-public) database of 1000 latent prints).   There is a need for additional databases containing a statistically significant number of cases representative of all latent prints, not just AFIS suitable or AFIS matchable latents. Vladmir Dvornychencko proposed outreach to smaller agencies for 50 samples each.  Mike Garris closed out the conference with a proposal for a latent testing methodology.  In general, it was proposed that an upcoming test evaluate latent fingerprint algorithms from multiple vendors using the following factors:
LFIS versus LFFS
Human versus Machine Encoding
Latent against Tenprint versus Tenprint against Latent
SDK testing model using Subroutine and API for 1) Encoder 2) Matcher 3) Score Normalization
Back End score-based metrics (ROC, DET, Imposter/Genuine disbributions per vendor)
1-1 match scores proposed, but arguments ensued for 1:N
“open” forum prior to a “Latent Search Grand Challenge”

The workshop closed with a discussion of Extended Feature Sets to include potentially major improvements for latent print standards.

______________________________________________________________________
Feel free to pass The Detail along to other examiners.  This is a free newsletter FOR latent print examiners, BY latent print examiners. There are no copyrights on The Detail, and the website is open for all to visit.

If you have not yet signed up to receive the Weekly Detail in YOUR e-mail inbox, go ahead and join the list now so you don't miss out!  (To join this free e-mail newsletter, enter your name and e-mail address on the following page: http://www.clpex.com/Subscribe.htm  You will be sent a Confirmation e-mail... just click on the link in that e-mail, or paste it into an Internet Explorer address bar, and you are signed up!)  If you have problems receiving the Detail from a work e-mail address, there have been past issues with department e-mail filters considering the Detail as potential unsolicited e-mail.  Try subscribing from a home e-mail address or contact your IT department to allow e-mails from Topica.  Members may unsubscribe at any time.  If you have difficulties with the sign-up process or have been inadvertently removed from the list, e-mail me personally at kaseywertheim@aol.com and I will try to work things out.

Until next Monday morning, don't work too hard or too little.

Have a GREAT week!