COMPUTERIZED FEATURE RETRIEVAL OF IMAGES - SUSPECT IDENTIFICATION

In suspect identification, witnesses examine photographs of known offe nders in mug-shot albums. The probability of correct identification de teriorates rapidly, however, as the number of mug-shots examined incre ases. Feature approaches, where mug-shots are displayed in order of si milarity to a witness's description of a suspect, attempt to increase identification success by reducing this number. A methodology is propo sed for the design, development, and evaluation of these systems based on computer simulations, experiments, and four classes of system perf ormance measures: identification performance, retrieval rank, toleranc e performance, and feature quality. This approach was used to develop a suspect identification system for 640 mug-shots of known offenders. System performance was good. In three experimental tests, > 90% of all witness searches resulted in target suspects retrieved in the first e ight mug-shots. The system is highly tolerant of witness errors and om issions.