Incorrect User Identification with Facial Recognition

Facial recognition technology is incredibly good at recognising individual faces from hundreds and even thousands of known faces. However, it is possible that facial recognition (and even humans!) can incorrectly identify an individual. In facial recognition terminology, an incorrect identification is known as a "false positive".

What Causes Incorrect Matches?

There are several possible causes of incorrect matches as follows:

• Incorrect Registration. If an individual manually registers as the wrong user (eg: by using their passcode), then that individual will be recognised as that user on an ongoing basis.
• Poor Quality Images. If images are of poor quality, for example, because of strong back light or bad camera settings, then facial recognition will be less reliable.
• Facial Coverings. If individuals wear dark or reflective glasses, wear masks or scarves, or wear hats that cast substantial shadows, their faces are less likely to be identified reliably.
• Very Similar Faces. If two individuals have very similar facial measurements, then it is possible their faces can be mistaken. This can be the case even if the individuals have different hair styles or different skin colors, because facial recognition only compares facial measurements.

The sections below explain how to respond to incorrect matches and more importantly how to avoid them.

Responding to Incorrect Matches

If an individual is incorrectly identified when they present to NoahFace, they simply need to press the "Not Me" button on the Facial Recognition screen:

NoahFace will automatically unlearn the association between their face and the identified user, and the next time they present they will either be correctly identified or asked to re-register.

Alternatively, if an individual reports to an Administrator that they were incorrectly matched (and they did not know to press the "Not Me" button), the Administrator can:

1. Login to the NoahFace Dashboard.
2. Navigate to the User Details page of the identified user (eg: if Mandy was incorrectly identified as Sammy, then navigate to Sammy's User Details page).
3. Press the Remove button next to their Consent date/time:

Removing a user's consent will delete:

• Any biometrics stored in the Cloud (immediately).
• Any biometrics stored on attached Access Points (on the next synchronisation).
• The Profile Picture.

Once this has been done, the identified user (eg: Sammy in this example), will need to re-register.

Avoiding Incorrect Matches

There are many steps you can take to minimize the possibility of incorrect matches as follows:

Mounting Position

You should mount your iPad so that individuals do not have significant back light when they present to NoahFace. For example, make sure the iPad is not facing a window. If this is not possible, you might consider installing artificial shading or front lighting.

Camera Brightness

You should set your camera brightness (under Settings/Camera in the NoahFace App) so that the images produced are slightly brighter (or over-exposed) than normal.

Employee Training

You should train your employees to:

• Remove their facial coverings (dark glasses, hats, masks, scarves, etc) when they present to NoahFace.
• Press the "Not Me" button if they are ever mis-recognised, rather than clocking in or out as someone else.

Limiting Users

Facial recognition will produce far more accurate results if it is matching an individual's face against 100 known faces compared with matching it against 10,000 known faces. So, if you have multiple sites and employees do not travel between those sites, you should either:

• Configure your access Rules so that employees can only use iPads at their sites (see: Multi-Site Deployments).
• Configure NoahFace so that biometrics are not shared between iPads.

Facial Recognition Tuning

You can adjust the following facial recognition settings (under Settings/Facial Recognition in the NoahFace App):

Detection Distance indicates how close an individual needs to be to the screen in order to be recognized. Increasing this to "Very Near" will mean individuals need to stand very close to the iPad to be recognized. This will improve the quality of captured images and therefore reduce the possibility of incorrect matches.

Detection Threshold indicates the level of certainty that the image is a face before it is analysed. Increasing this to "High" or "Very High" will the possibility of incorrect matches, however, it will also make it more likely that individuals need to remove any facial coverings (eg: dark glasses, masks, scarves, etc) to be recognized.

Matching Threshold indicates the level of certainty that the detected face belongs to a known user before declaring a match. Increasing this to "Very High" or "Maximum" will reduce the possibility of incorrect matches.

‍

‍

‍