International. As we continue to conform to social separation mandates, more experts are now recommending wearing face masks as another way to limit the spread of COVID-19. In situations and environments where SAFR facial recognition is vital for safety and safe access, wearing face masks does not compromise their benefits.
It is true that the more facial features that are available as data points, the greater the chances of an accurate match. However, a resilient algorithm can adapt when it sees a darkened or partially covered face and still delivers excellent results.
Occlusion detection has been a feature of the SAFR platform for some time; we've worked hard to make it work well. As preventive masks become more common around the world, the accuracy of partially occluded faces is more important than ever. To meet this need, we are further improving our occlusion logic to ensure we can maintain the highest accuracy and best performance in difficult conditions.
A mask is just a way to occlude a face. When any part of the face is not visible, a facial recognition algorithm must focus on any reference point on the face that it can see to determine an accurate match. It's a process that requires training and adjustments to achieve this, and when implemented commercially, the benefits of accurate recognition under occluded conditions are valuable and, in some cases, save lives:
- Health professionals can walk through safe areas without removing personal protective equipment;
- Providers of essential services can continue to provide products and services without delay;
- Security threats will be recognized and responses can be immediate.
The SAFR evaluates occlusions at various levels, and at some levels, it is not possible to make an exact match. In such cases, SAFR can automate subsequent procedures that ensure the protection of security protocols. For example, different matching thresholds may be required when occlusions are detected, a multi-factor authentication flow may be initiated, or the security team may be notified when a person is not recognized due to a face mask or other occlusion.
Right now, many marketers enthusiastically share the news that they are "mask immune," able to recognize faces with the same accuracy as when people don't wear masks. This may be true when the sample size is small. But what we do know about the human face is that they are all different, and as the sample size increases, so does the variety of features and the likelihood that someone with similar features will appear in the database. Some faces have unique features, especially around the eyes or nose, while others may have these features in the lines of the mouth or jaw. Recognizing this, it becomes obvious that some people will be more easily recognized when the upper half of the face is visible; others when the lower half is available for analysis. For this reason, some people may not have the same level of accuracy as others when they are occluded by similar objects.
We are currently emphasizing training our algorithms to reduce the negative impacts that occlusions, such as face masks, have on accuracy. But you can take advantage of a SAFR feature right now to improve accuracy when a person's external appearance changes significantly: "Grouping in the Person Directory." Refer to the SAFR documentation in "Manage People in the Person Directory" for instructions.
Yes, we can also combine faces when people wear masks, but we feel a responsibility to be honest about the limitations and get even better performance when the SAFR understands what a person can look like when wearing a mask.
Accuracy is always a complex issue, as several variables are involved: the quality of the recorded face image, the degree to which a subject is aware of and cooperates with capturing a good quality image, and the impact of lighting and other environmental variables According to the University of Massachusetts, current SAFR resources for the Faces in the Wild (LFW) tag have an actual ID rate of 99.87%, with only 1:1,000,000 false IDs (false positives). Internal benchmarking shows that when people present themselves in cooperative identification scenarios, such as an Access Control use case, and both the subject and one of their facial images in the database wear a mask, the TRUE SAFR identification rate is 93.5%, with less than 1:3,760 false IDs. Although this is an excellent level of accuracy, considering that 50% to 60% of the face is covered, at SAFR we plan to further improve this capability.
Our goal now, as it always has been, is to equip our end users with high-performance AI-powered software and systems that quickly adapt to changes in environments that increase their need for use. It is a reality that we are becoming familiar with in this time of crisis and we hope that SAFR can help.
* Article written by Eric Ress, Senior Director of Product Management for Facial Recognition and Security Solutions at SAFR.
