Académique Documents
Professionnel Documents
Culture Documents
Features Editor:
Rebecca L. Deuel
rdeuel@computer.org
Figure 1. Three stages of facial recognition: (a) the range image and texture; (b) the preprocessed surface; (c) the canonical form.
2002, which tested commercial face recog- vide accuracy comparable with fingerprints. examined the use of morphable models—a
nition accuracy and identified numerous For facial recognition, the best packages technique of taking a facial image from any
characteristics for optimizing face recogni- available provide a 90 percent probability of angle and projecting what the subject might
tion technologies’ performance, as well as true verification with a 1 percent probability look like facing forward. There was a dra-
areas for future research. of false verification. This is a helpful find- matic improvement in performance using the
Looking at 10 mature face-recognition ing: The November 2002 report “Summary morphable models. One of the top three sys-
systems, FRVT 2002 tested them on three of NIST Standards for Biometric Accuracy, tems increased its performance from 26 per-
basic tasks: identification, verification, and Tamper Resistance, and Interoperability” cent on nonprocessed, nonfrontal images to
watch-list screening. Identification involves notes that, “within the intelligence commu- 84 percent on morphed images.
matching a biometric record from a single nity, facial data is often the only biometric The size of the database used for iden-
subject probe against an entire database of data that has been and is currently being tification or watch-list screening signifi-
similar biometric records to determine the captured. Face data is one key source for cantly impacted results. The experiments
record owner’s identity—a one-to-many watch lists, and in many situations finger- showed that identification performance
comparison. The verification process con- print data cannot even be captured to use in decreases linearly with respect to the log-
firms that a person is who he or she claims constructing a watch list.” arithm of the database size. NIST reports
to be by matching the biometric record FRVT 2002 showed that facial recogni- a similar effect for the watch-list task—
against that of his or her claimed identity—a tion accuracy varies according to different as the watch list size increases, perfor-
one-to-one comparison. Verification rates factors, which might help in planning better mance decreases. The FRVT 2002 over-
are offset with false accept rates, and verifi- applications and designing future research. view states that, “In general, a watch list
cation performance is described by the two Several image characteristics affected re- with 25 to 50 people will perform better
statistics. Watch-list screening is typically sults. First, as you’d expect, accuracy drops than a larger size watch list.”
the most demanding task, involving two as time increases between the acquisition of For the first time, the test covered the
steps. First, a system must detect if an indi- the database image and the presentation of effects of demographics. The results re-
vidual is even on the watch list, then, if so, the newest image because people age and ported that males are easier to recognize
correctly identify the individual. change in appearance over time. NIST than females and that older people are
Compared with similar tests performed reported that performance degraded at ap- easier to identify than younger people.
two years earlier in FRVT 2000, the FRVT proximately 5 percentage points per year. In For the top systems, identification rates
2002 results show a significant improvement addition, the study found that indoor light- for males were 6 percent to 9 percent
in face recognition systems’ verification ca- ing changes didn’t make an appreciable higher than for females. For every 10
pabilities, indicated by a 50 percent reduc- difference to the top systems’ accuracy, al- years increase in age, on average identifi-
tion in error rates. For the best systems tested though face recognition from outdoor im- cation performance increases approxi-
in FRVT 2002, the top-rank identification agery showed a considerable drop in perfor- mately 5 percentage points.
rate was 85 percent on a database of 800 mance, with the best performing systems
people. With a false accept rate of 10 per- turning in a recognition rate of 50 percent. Reading faces
cent, the top two systems turned in a verifica- The FRVT 2002 also compared the rates Although recognizing faces is important
tion rate of 96 percent. For the best system for still and video images and found, contrary to such security applications as financial
using a watch list of 25 people, the detection to expectations, that recognition performance verification (ATM and credit cards), biomet-
and identification rate was 77 percent. using video sequences was similar to the ric locks, and passport or visa control, read-
FRVT 2002 found that facial recognition performance using still images. For images ing faces has important uses in medicine and
systems attempting verification tasks pro- with nonfrontal presentation, FRVT 2002 security as well. A team of scientists at the