Facial Recognition and COVID19
One of the key technologies that will help making the post-COVID19 world a safer place is face recognition. People are usually afraid of being exposed to face recognition systems, thinking they are some kind of malicious Big Brother who watches them. Nothing could be further from the truth. Instead, they are a highly accurate biometric technology that allows -among other applications- touchless and distant access control to workplaces, critical infrastructures, transports or events. With face recognition, there is no need to manipulate access security cards, to approach or put the finger in any device shared by hundreds of persons, thus helping to control pandemic spread. However, there is a main challenge for the face recognition technology in this new normality: the common use of medical masks that occlude half of the face. Ideally, individuals should not have to expose themselves and others to the virus by removing their masks in access controls, but the vast majority of current face recognition algorithms are not yet sufficiently robust to deal with such large facial occlusions.
Indeed, we humans are facing the same challenge. Eye-tracking and psychological works have widely studied the human visual attention mechanisms that take place when confronted to the task of identifying people1. They demonstrate that we innately and systematically fix the triangular face region formed by the two eyes and the mouth. Therefore, if a vertex of this triangle is occluded by a mask, our long-term acquired face recognition mechanisms will encounter difficulties, lose robustness, and take time to adapt to the new facial configuration. Like humans, face recognition algorithms will also need to adapt themselves to the structure of masked faces.
Deep Learning into action
Existing face recognition algorithms are grounded on Artificial Intelligence (AI), particularly on Machine Learning and Deep Learning techniques. This means that they automatically learn an identification strategy from a dataset of millions of facial images used for their training. Until now, facial training datasets have varied in terms of illumination conditions, head poses or backgrounds; they have also presented certain facial occlusions in the form of eye glasses, caps, scarves or beards. But they hardly ever have contained faces with medical masks!
A possible solution to make face recognition models able to identify persons with masks consists on collecting new training images for that purpose. But there are other, more algorithmic-oriented, alternative approaches. Recently, state-of-the-art academic papers have introduced visual attention mechanisms for Deep Neural Networks, so that they can be taught to focus their attention on specific regions of the image during training2. These mechanisms could be applied to teach face recognition models to shift their attention towards the unmasked upper-face region. In any case, face recognition systems must be able to recognize both mask wearers and unmasked people. Therefore, another interesting way to approach the problem could be building a mask detector and then, depending on whether the mask is present or not, and applying different adapted face recognition strategies.
Herta has been working on strong occlusions for the last few months and is currently bringing all these solutions to its products to lead the “new normality” era of face recognition.
1 We refer the reader to this interesting scientific paper by Blais et al. www.ncbi.nlm.nih.gov
2 An introductory read about this topic can be found in the paper “Learn to pay attention” by Jetley et al. arxiv.org/abs/1804.02391
Written by: Isabelle Hupont