The 95 Percent Threshold
According to a study published in PLOS One, these forensic identification specialists require only one or two specific facial features—such as eyes or ears—to achieve accuracy that averaged out to more than 90% correct when matching faces.
They are often used during criminal investigations. Carina Hahn said: Forensic face specialists help investigations and provide expert testimony at court trials.
Hahn also noted: It is remarkable how much identity information they can extract from so little.
Researchers Test How Forensic Experts Process Facial Boundaries
Led by Carina Hahn from the National Institute of Standards and Technology in Maryland, researchers examined how forensic experts process facial boundaries. The investigation tested participants by presenting two side-by-side target images alongside a completely occluded probe image. Analysts had to request facial features to be revealed one by one under a points-based scoring system designed to discourage random guessing.
On each trial, participants viewed two faces on the computer screen as target images, which were always of different people, while a third face began completely occluded as the probe image. To make decisions, participants revealed features on the probe. The predetermined condition was identical except for the available feature options, and participants in that condition had only one button labeled.
Ears emerged as the most self-selected feature during the trials. However, the data revealed a distinct divergence between user preference and diagnostic utility. While specialists leaned toward ears, the highest accuracies in identification were associated with seeing the ears and forehead. Necks and chins proved to be the least helpful.
Cognitive Flexibility in Sequential Matching
When experts bypassed their default preference for ears, they demonstrated a high degree of cognitive flexibility. Specialists dynamically adjusted their feature requests based on how different that feature was between the two faces. Hahn emphasized that this combination of relying on features known to be useful, like ears, with adjusting when needed, had not been documented before.
Hahn noted: When not choosing the ears, experts adapted as hypothesized and generally requested features based on how different that feature was between the two faces.
Findings Could Improve Machine Learning Facial Recognition
The findings extend beyond human cognitive profiling. Researchers say these findings can be used to improve training of machine-learning models for facial identification.
Hahn stated that Understanding the cognitive processes that they use when identifying faces will be useful for improving how future experts are trained, as well as for improving machine-learning algorithms that are used for facial recognition.
