How does NEC NeoFace Watch real-time facial recognition work?

NeoFace Watch real-time face recognition is designed to find a needle in a haystack, reducing a very large problem down to a much smaller review process. It is used to determine the degree of similarity of facial images of people captured by cameras to watchlists of facial images. Human users are involved throughout the entire process, from managing the watchlist data, setting review thresholds, monitoring alerts and checking these and taking appropriate action where needed.

For more detailed information on features, capabilities and other specifications, please contact ABT Security Systems at 9878 7111.
Email enquiries can be sent to: info [at] abtgroup.com.au

Create Watchlist
Customer chooses which facial images to add to a watchlist. The system creates a unique signature representing the features of each face.

Set Review Threshold
A review threshold is set by users to determine the degree of similarity for alerting and review by users. Adjustments are made for different operational needs.

Observe Crowd
Video feeds from cameras are analysed in real-time. Faces found in the crowd are compared for similarity against those in the watchlist.

Raise Alerts
Similar facial images above the review threshold are raised as alerts to users. All other facial images are automatically discarded immediately.

Review Alerts
Users review the alerts, comparing facial images from cameras against the corresponding watchlist images to determine whether further action is needed.

Delete Data
Images in alerts are only retained for a limited, user-defined time period, in accordance with customer’s policy.

What is accuracy? – a worked example

It is incorrect to state accuracy based simply on true positives vs false positives as these will vary with different image qualities, user-selected thresholds and other factors.

In the example given here, review thresholds have been optimised to maximise identifications. It would be wrong and misleading to conclude an accuracy rate of 33% based on a simplistic ratio of true positives to total positives.

The reality is that the identification success rate is 100% in this example. The false positive rate is very low but could be reduced or eliminated if required by some further fine-tuning.

The quantity of faces presented for review by users is determined by setting and fine-tuning review thresholds and are affected by many environmental variables. To know more about how NeoFace Watch Facial Recognition identifies similar image, click to download the full version of the report.

Download

While NEC has consistently achieved the highest accuracy rankings from many tests by its customers and other independent bodies, it continues to enhance its face recognition system, embracing the latest AI and deep learning technologies; and makes this available to customers regularly. NEC does not have access to customer data, neither the database of images in the watchlists, nor the images of people captured by the cameras, and NEC does not use any of this specific data to improve the NeoFace Watch system.