Current and completed projects

Current project: Fingerprint Image processing, Matching and Indexing (independent development)

May, 2009 update:

Fast ID search ("Indexing"): for all interested parties who keep contacting me on this item -- it will remain at the "Proof of concept" stage for at least few months more :(.

04/2002-4/2007: Fingerprint recognition technology (for BIO-key International, Inc.)

The algorithm developed for FVC 2002 became the core of BIO-key (R) International, Inc. VST (TM) engine. Despite my focus shifted away from the accuracy, the algorithm provided highest benchmarks for BIO-key during these 5 years. In 2006, BIO-key was the best performer in North America in the NIST MINEX test. The last submission before I left was BIO-key's entry in the NIST SDK test. In June 2007, at the time of publication, it was uniformly the second best in the EER table, just slightly behind Cogent's I-version; in all other tables the algorithm was also among the top performers.

February, 2009 update: my one-year non-compete agreement with BIO-key long expired on April 30, 2008. This Web page is personal and is created for professionals and for those who seeks for professional algorithm designers. As such, it describes my personal achievements and expertise. However, those visiting this page to collect information about BIO-key -- rest assured that company's engineers and scientists are indeed capable of supporting and improving my input, whatever of it remains in use.
May, 2010 update: This point was just well confirmed. Without my participation for 3 years, the company demonstrated very reasonable accuracy gains for most of the indices in its recent NIST PFT submission.

10/2001-2/2002: Fingerprint recognition algorithm (for Fingerprint Verification Competition 2002)

The international Fingerprint Verification Competition takes place every other year. This is a major event in biometrics. The organizers benchmark fingerprint recognition technologies. 48 teams signed-in initially for FVC 2002.  The final competition was between 31 participants (mostly industrial, with 6 academic teams, another independent developer and me).

My ranking is here: the official table of average results, ordered by perfomance in high security domain.

I got into the medal table, achieved the sixth place by average EER and the third place in the world in the major biometrical domain, performance at low false accept levels: in average on all databases in both False reject rate at Zero false accept and False reject at 0.1% false accept I was behind only Bioscrypt and an anonymous (Bioscrypt's clone?). 

The algorithm outperformed industrial monsters like SAGEM - the winner of FVC 2000 - and many other formal "industry leaders", despite a bug in the executables inflated all the average error indices on four databases by ~0.5% (after all, since the FVC 2002 announcement shortly after 9/11 events, it was just five months to improve the algorithm abandoned in 1998 and I had to replace or completely rewrite about 80% of the code!).

The algorithm is now the core of the BIO-key technology. It completely replaced the previous core code of the company. After bug fixes and few parameter adjustments, on high quality FVC 2002 databases it operated at the same error levels as Bioscrypt (the winner of FVC 2002); on more realistic databases it was better. For instance, in 2002 Bioscrypt published its EERs for FVC 2000 databases. They were significantly better than the winning numbers of Sagem, and best published so far. As tested in July 2002, the error rate of my algorithm on these databases was ~50% less in average than Bioscrypt's, with uniformly better performance on all four databases.

05/2001-10/2001: Text detection from a video stream (for BBN Technologies, with Prem Natarajan).

04/2001-05/2001: Language identification from [spoken] speech utterances (for BBN Technologies, with Prem Natarajan).

8/1999-1/2001: Forecasting of electric power demand (medium-term spatial forecasting of electric loads on Long Island for LIPA/KeySpan, with Prof. E.Feinberg)

6/2000-11/2000: Technology for comparison of speech recognition engines in noisy environments (for Symbol Technologies).

• Download the software manual
  

11/1998-4/1999: Tracking multiple distributed targets from a series of noisy images (with Prof. M.Malioutov)

• Download project presentation (pdf).